{ "schema_version": "1.1", "generated_at": "2026-10-15", "note": "Every license, commercial_use and commercial_cost value was read from the URLs listed in that record's sources / commercial_cost_sources fields on the verified_at date. commercial_use is about RIGHTS (may a company use it), commercial_cost is about PRICE (what commercial use costs). Fields that could not be verified are set to \"unverified\".", "count": 62, "tools": [ { "id": "rfdiffusion", "name": "RFdiffusion", "category": "de novo binder and protein design", "one_line": "Diffusion model that generates protein backbones, including binders to a target.", "what_it_does": "Generates protein backbones. Input: a target structure (PDB) with hotspot residues, or just a length/symmetry spec; output: backbone-only PDBs (no sequence) for de novo binders, scaffolds or motif grafts. You then run inverse folding (ProteinMPNN) and a structure predictor to filter.", "maker": "Institute for Protein Design, University of Washington (RosettaCommons)", "url": "https://github.com/RosettaCommons/RFdiffusion", "paper_url": "https://www.nature.com/articles/s41586-023-06415-8", "license": "BSD License (as titled in the repo LICENSE file)", "commercial_use": "yes", "commercial_use_note": "The LICENSE text states it covers both source code and the model weights referenced in the README; it is a permissive BSD-style license with no field-of-use restriction.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/RosettaCommons/RFdiffusion/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Default starting point for de novo backbone generation and target-conditioned binder scaffolds. Output needs sequence design (ProteinMPNN) and in-silico filtering; raw success rates per design are low, so expect to order many variants.", "sources": [ "https://raw.githubusercontent.com/RosettaCommons/RFdiffusion/main/LICENSE", "https://raw.githubusercontent.com/RosettaCommons/RFdiffusion/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "rfantibody", "name": "RFantibody", "category": "antibody-specific design", "one_line": "Pipeline for structure-based de novo antibody and nanobody design.", "what_it_does": "Designs antibody/nanobody CDR loops onto a real framework. Input: target PDB + epitope hotspots + a VHH or scFv framework; output: designed CDR backbones, then sequences via its ProteinMPNN step and a filter pass with RoseTTAFold2 antibody scoring.", "maker": "Institute for Protein Design, University of Washington (RosettaCommons)", "url": "https://github.com/RosettaCommons/RFantibody", "paper_url": "https://www.biorxiv.org/content/10.1101/2024.03.14.585103v1", "license": "MIT License", "commercial_use": "unverified", "commercial_use_note": "The repo LICENSE is MIT, which covers the code. The README downloads model weights via include/download_weights.sh but states no separate license for those weights, so weight terms are unverified.", "commercial_cost": "free", "commercial_cost_note": "No fee is charged for the code (MIT) or the weight download script, but the weights carry no licence of their own, so the right to use them commercially is a separate, unresolved question from the price.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/RosettaCommons/RFantibody/main/README.md" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Use when you specifically need CDR-loop design on a real framework (VHH or scFv) rather than generic binders. Published experimental hit rates are low and affinities are typically weak starting points needing affinity maturation.", "sources": [ "https://raw.githubusercontent.com/RosettaCommons/RFantibody/main/LICENSE", "https://raw.githubusercontent.com/RosettaCommons/RFantibody/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "bindcraft", "name": "BindCraft", "category": "de novo binder and protein design", "one_line": "AlphaFold2-backpropagation pipeline that designs binders end to end.", "what_it_does": "One-command de novo binder design. Input: target PDB, chain and target residues; output: binder sequences plus predicted complex PDBs and filter statistics (iPTM, pAE interaction, ddG, clashes), using AlphaFold2 backpropagation and MPNN redesign.", "maker": "Martin Pacesa / Correia lab, EPFL", "url": "https://github.com/martinpacesa/BindCraft", "paper_url": "https://www.nature.com/articles/s41586-025-09429-6", "license": "MIT License", "commercial_use": "restricted", "commercial_use_note": "Code is MIT, but the README states the installer installs PyRosetta, 'which requires a license for commercial purposes'. PyRosetta's own page confirms licensing is free for academic/non-profit and fee-based for commercial users. AlphaFold2 parameters it uses are CC BY 4.0, which does allow commercial use.", "commercial_cost": "paid-license", "commercial_cost_note": "The pipeline itself is free, but its installer pulls in PyRosetta, and PyRosetta's licensing page states that commercial users pay a fee (academic and non-profit use is free). Budget for a PyRosetta commercial licence, or replace that dependency.", "commercial_cost_sources": [ "https://www.pyrosetta.org/home/licensing-pyrosetta", "https://raw.githubusercontent.com/martinpacesa/BindCraft/main/README.md" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Good one-command route from target PDB to candidate binder sequences with built-in filters. Depends on PyRosetta (commercial license needed) and is GPU-hungry per trajectory.", "sources": [ "https://raw.githubusercontent.com/martinpacesa/BindCraft/main/LICENSE", "https://raw.githubusercontent.com/martinpacesa/BindCraft/main/README.md", "https://www.pyrosetta.org/home/licensing-pyrosetta", "https://raw.githubusercontent.com/google-deepmind/alphafold/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "boltzgen", "name": "BoltzGen", "category": "de novo binder and protein design", "one_line": "Generative model for binder design across protein and non-protein targets.", "what_it_does": "Generative binder design against protein and non-protein targets. Input: target (protein, nucleic acid or small molecule) plus binding-site spec in YAML; output: designed binder backbones and sequences with self-consistency/refolding metrics.", "maker": "Boltz team (MIT CSAIL and collaborators)", "url": "https://github.com/HannesStark/boltzgen", "paper_url": "https://hannes-stark.com/assets/boltzgen.pdf", "license": "MIT License", "commercial_use": "unverified", "commercial_use_note": "The repo LICENSE is MIT, covering the code. The README documents automatic download of ~6 GB Boltz weights plus ESMFold2 weights but makes no license statement for those weights, so weight terms are unverified.", "commercial_cost": "free", "commercial_cost_note": "No fee: code is MIT and the weights download automatically. The weights' own terms are not published, so 'free' here describes price, not a settled right to use them commercially.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/HannesStark/boltzgen/main/README.md" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Reach for it when you want a newer generative binder designer with built-in refolding metrics. Large download (~40 GB with ESMFold2) and needs a recent CUDA driver; very little independent wet-lab replication published yet.", "sources": [ "https://raw.githubusercontent.com/HannesStark/boltzgen/main/LICENSE", "https://raw.githubusercontent.com/HannesStark/boltzgen/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "germinal", "name": "Germinal", "category": "antibody-specific design", "one_line": "End-to-end pipeline for de novo nanobody and scFv design.", "what_it_does": "End-to-end de novo nanobody and scFv design. Input: target PDB + epitope residues; output: ranked designed antibody sequences with predicted complexes, by combining backbone hallucination, sequence design and in-silico filtering in one pipeline.", "maker": "Santiago Mille-Fragoso et al. (open-source release)", "url": "https://github.com/SantiagoMille/germinal", "paper_url": "https://www.biorxiv.org/content/10.1101/2025.09.19.677538v1", "license": "Apache License 2.0", "commercial_use": "yes", "commercial_use_note": "Apache-2.0 LICENSE file in the repo; permissive, commercial use allowed. Any bundled third-party weights/dependencies should be checked separately.", "commercial_cost": "free", "commercial_cost_note": "Apache-2.0 code, no fee. Check the dependency stack it installs before commercial deployment - the pipeline pulls in several third-party models.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/SantiagoMille/germinal/main/LICENSE" ], "open_weights": "unverified", "compute": "heavy-gpu", "when_to_use": "Consider for antibody/nanobody design when you want an alternative to RFantibody with a reported wet-lab validated pipeline. Young project; dependency stack is heavy and results are not widely reproduced.", "sources": [ "https://raw.githubusercontent.com/SantiagoMille/germinal/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "proteinmpnn", "name": "ProteinMPNN", "category": "sequence design / inverse folding", "one_line": "Designs amino-acid sequences that fold to a given backbone.", "what_it_does": "Inverse folding. Input: a backbone PDB (optionally with fixed positions, chain bias, symmetry); output: amino-acid sequences predicted to fold to that backbone, with per-residue probabilities and scores. CPU-fast; the standard step after any backbone generator.", "maker": "Justas Dauparas, Institute for Protein Design, University of Washington", "url": "https://github.com/dauparas/ProteinMPNN", "paper_url": "https://www.science.org/doi/10.1126/science.add2187", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "MIT LICENSE file in the repo; no field-of-use restriction.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/dauparas/ProteinMPNN/main/LICENSE" ], "open_weights": true, "compute": "cpu", "when_to_use": "The standard inverse-folding step after any backbone generator; runs fine on CPU. It optimises foldability, not binding or developability, so you still need structure prediction and liability filtering afterwards.", "sources": [ "https://raw.githubusercontent.com/dauparas/ProteinMPNN/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "ligandmpnn", "name": "LigandMPNN", "category": "sequence design / inverse folding", "one_line": "Inverse folding that is aware of ligands, nucleotides and metals.", "what_it_does": "Inverse folding that sees the non-protein context. Input: a PDB containing the protein plus ligands, nucleotides, metals or waters; output: sequences designed around that context, with side-chain packing - the right choice for binding pockets and metal sites.", "maker": "Justas Dauparas, Institute for Protein Design, University of Washington", "url": "https://github.com/dauparas/LigandMPNN", "paper_url": "https://www.nature.com/articles/s41592-025-02626-1", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "MIT LICENSE file in the repo; no field-of-use restriction.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/dauparas/LigandMPNN/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Use instead of ProteinMPNN when the binding site contains a small molecule, nucleic acid or metal that must be respected. Overkill for plain protein-protein interfaces.", "sources": [ "https://raw.githubusercontent.com/dauparas/LigandMPNN/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "alphafold3", "name": "AlphaFold 3", "category": "structure and complex prediction", "one_line": "Predicts structures of proteins with ligands, nucleic acids and modifications.", "what_it_does": "Predicts the structure of a complex. Input: a JSON listing protein/RNA/DNA chains, ligands, ions and modifications; output: an all-atom predicted complex with pLDDT, PAE and ranking scores. Code is public; weights must be requested and are non-commercial.", "maker": "Google DeepMind and Isomorphic Labs", "url": "https://github.com/google-deepmind/alphafold3", "paper_url": "https://www.nature.com/articles/s41586-024-07487-w", "license": "Apache License 2.0 (code); 'AlphaFold 3 Model Parameters Terms of Use' (weights)", "commercial_use": "restricted", "commercial_use_note": "Code is Apache-2.0, but the README states you may only use the model parameters if received directly from Google, and the weights Terms of Use make them available free of charge only for certain non-commercial uses, by or on behalf of non-commercial organisations. Commercial organisations cannot use the released weights under those terms.", "commercial_cost": "none", "commercial_cost_note": "There is no price at which a company can use the AlphaFold 3 weights: the Model Parameters Terms of Use restrict them to non-commercial use by non-commercial organisations, and Google does not offer a commercial licence for download. The code is free. For commercial work use Boltz-2 or Chai-1 instead.", "commercial_cost_sources": [ "https://github.com/google-deepmind/alphafold3/blob/main/WEIGHTS_TERMS_OF_USE.md", "https://github.com/google-deepmind/alphafold3/blob/main/WEIGHTS_PROHIBITED_USE_POLICY.md" ], "open_weights": false, "compute": "heavy-gpu", "when_to_use": "Best-in-class for complex/ligand-containing predictions if you qualify for the weights. For a company, the weight terms are a blocker, so use Boltz-2 or Chai-1 instead.", "sources": [ "https://raw.githubusercontent.com/google-deepmind/alphafold3/main/LICENSE", "https://raw.githubusercontent.com/google-deepmind/alphafold3/main/WEIGHTS_TERMS_OF_USE.md", "https://raw.githubusercontent.com/google-deepmind/alphafold3/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "alphafold2", "name": "AlphaFold 2", "category": "structure and complex prediction", "one_line": "Original AlphaFold; monomer and multimer structure prediction from sequence.", "what_it_does": "Predicts protein structure from sequence. Input: FASTA (monomer or multimer); output: 5 ranked all-atom models with pLDDT and PAE, plus MSAs. Still the workhorse for binder-design filtering and the engine inside BindCraft/ColabFold.", "maker": "Google DeepMind", "url": "https://github.com/google-deepmind/alphafold", "paper_url": "https://www.nature.com/articles/s41586-021-03819-2", "license": "Apache License 2.0 (code); CC BY 4.0 (model parameters)", "commercial_use": "yes", "commercial_use_note": "README's 'Model Parameters License' section states parameters are made available under CC BY 4.0, which replaced the earlier CC BY-NC 4.0 terms; attribution required. Some mirrored MSA databases carry their own CC BY-SA 4.0 terms.", "commercial_cost": "free", "commercial_cost_note": "Free: code Apache-2.0, parameters CC BY 4.0 (changed from CC BY-NC in January 2022). Attribution is the only cost.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/google-deepmind/alphafold/main/README.md" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Still the workhorse for monomer and protein-protein complex prediction and the engine inside many binder pipelines. Weaker on ligands and nucleic acids than AF3/Boltz-2, and MSA generation is the slow step.", "sources": [ "https://raw.githubusercontent.com/google-deepmind/alphafold/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "boltz-2", "name": "Boltz-2", "category": "structure and complex prediction", "one_line": "Open structure and binding-affinity prediction model, MIT for code and weights.", "what_it_does": "Open AlphaFold3-class structure prediction plus binding affinity. Input: YAML/FASTA with protein, ligand and nucleic-acid chains, optional templates and affinity query; output: predicted complex with confidence scores and a predicted binding affinity value for the ligand.", "maker": "Wohlwend, Corso, Passaro et al. (MIT CSAIL) / boltz.bio", "url": "https://github.com/jwohlwend/boltz", "paper_url": "https://doi.org/10.1101/2025.06.14.659707", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "LICENSE file is MIT and the README states explicitly that all code and weights are MIT, freely usable for academic and commercial purposes.", "commercial_cost": "free", "commercial_cost_note": "Free, and cleanly so: MIT covers both code and weights, which is why it is the usual commercial replacement for AlphaFold 3.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/jwohlwend/boltz/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "The pragmatic commercial substitute for AlphaFold 3, with an affinity-prediction head. Its affinity outputs are predictions, not measured KD, and should only be used to rank, never to report numbers.", "sources": [ "https://raw.githubusercontent.com/jwohlwend/boltz/main/LICENSE", "https://raw.githubusercontent.com/jwohlwend/boltz/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "chai-1", "name": "Chai-1", "category": "structure and complex prediction", "one_line": "Multi-modal structure prediction for proteins, ligands, nucleic acids and glycans.", "what_it_does": "Structure prediction for proteins with ligands, nucleic acids, glycans and modified residues. Input: FASTA plus optional restraints/MSAs; output: all-atom predicted complex with per-atom confidence - usable as an AlphaFold3 substitute whose weights are usable commercially.", "maker": "Chai Discovery", "url": "https://github.com/chaidiscovery/chai-lab", "paper_url": "https://www.biorxiv.org/content/10.1101/2024.10.10.615955v2", "license": "Apache License 2.0", "commercial_use": "yes", "commercial_use_note": "LICENSE file is Apache-2.0 and the README's Licence section states both code and model weights are Apache-2.0 and usable for academic and commercial purposes including drug discovery.", "commercial_cost": "free", "commercial_cost_note": "Free to download and run commercially under Apache-2.0. Chai's hosted web platform is a separate product with its own terms; the open weights are what carry the free commercial grant.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/chaidiscovery/chai-lab/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Good second opinion alongside Boltz-2 for complexes; easy pip install. Like all co-folding models it is unreliable for de novo binder interfaces it has not seen.", "sources": [ "https://raw.githubusercontent.com/chaidiscovery/chai-lab/main/LICENSE", "https://raw.githubusercontent.com/chaidiscovery/chai-lab/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "esmfold", "name": "ESMFold", "category": "structure and complex prediction", "one_line": "Single-sequence structure prediction with no MSA required.", "what_it_does": "MSA-free structure prediction. Input: a single protein sequence; output: an all-atom PDB with pLDDT in seconds-to-minutes, no MSA search needed. Good for screening thousands of designed sequences; weaker than AlphaFold2 on natural targets.", "maker": "Meta AI (FAIR) Evolutionary Scale Modeling", "url": "https://github.com/facebookresearch/esm", "paper_url": "https://www.science.org/doi/10.1126/science.ade2574", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "MIT LICENSE file in the facebookresearch/esm repo covering the released code and models.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/facebookresearch/esm/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Use when you need fast structures for thousands of sequences without MSAs, e.g. filtering design libraries. Less accurate than AF2/Boltz on hard targets and not a complex predictor.", "sources": [ "https://raw.githubusercontent.com/facebookresearch/esm/main/LICENSE", "https://raw.githubusercontent.com/facebookresearch/esm/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "esm2", "name": "ESM-2", "category": "protein language models", "one_line": "Protein language model for embeddings, fine-tuning and variant scoring.", "what_it_does": "Protein language model. Input: a protein sequence; output: per-residue and per-sequence embeddings, masked-token likelihoods and variant pseudo-likelihood scores - the standard feature extractor for fitness/liability models and design filters.", "maker": "Meta AI (FAIR)", "url": "https://huggingface.co/facebook/esm2_t33_650M_UR50D", "paper_url": "https://www.science.org/doi/10.1126/science.ade2574", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "The Hugging Face model card front-matter declares 'license: mit' for the released weights; the source repo LICENSE is also MIT.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/facebookresearch/esm/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Default embedding model for property predictors and sequence ranking; the 650M checkpoint runs on one GPU. Its likelihoods track evolutionary plausibility, not binding affinity or developability.", "sources": [ "https://huggingface.co/facebook/esm2_t33_650M_UR50D/raw/main/README.md", "https://raw.githubusercontent.com/facebookresearch/esm/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "esm-c", "name": "ESM C / ESMFold2 (esm package)", "category": "protein language models", "one_line": "Newer ESM protein world-model family with folding, now MIT licensed.", "what_it_does": "Newer ESM encoder family (ESM C). Input: a protein sequence; output: embeddings and likelihoods at better quality-per-parameter than ESM-2, as a drop-in replacement for representation and scoring workloads.", "maker": "EvolutionaryScale / Chan Zuckerberg Biohub", "url": "https://github.com/evolutionaryscale/esm", "paper_url": "https://www.biorxiv.org/content/10.64898/2026.06.03.729735", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "LICENSE.md in the repo is MIT (Copyright 2026 Chan Zuckerberg Biohub) and the README's Licenses section states the models are available under the MIT license. Note this differs from the earlier ESM3 non-commercial community license; verify the checkpoint you actually download.", "commercial_cost": "free", "commercial_cost_note": "Free: the esm repository's licence file is MIT (Copyright 2026 Chan Zuckerberg Biohub) and the open checkpoints are downloadable without payment. The maintainers also sell hosted API access to the larger models, which is optional.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/evolutionaryscale/esm/main/LICENSE.md", "https://raw.githubusercontent.com/evolutionaryscale/esm/main/README.md" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Use for the strongest open ESM-family embeddings/folding where license matters. Large weight downloads and newer API; older ESM3 checkpoints may still carry different terms.", "sources": [ "https://raw.githubusercontent.com/evolutionaryscale/esm/main/LICENSE.md", "https://raw.githubusercontent.com/evolutionaryscale/esm/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "ablang2", "name": "AbLang2", "category": "protein language models", "one_line": "Antibody-specific language model for paired heavy/light sequences.", "what_it_does": "Antibody-specific language model. Input: a paired heavy+light (or single-chain) antibody sequence; output: embeddings and residue likelihoods used to score naturalness, restore missing residues and rank humanisation or affinity-maturation edits.", "maker": "OPIG, University of Oxford", "url": "https://github.com/oxpig/AbLang2", "paper_url": "https://academic.oup.com/bioinformatics/article/40/11/btae618/7845256", "license": "BSD 3-Clause License", "commercial_use": "yes", "commercial_use_note": "BSD 3-Clause LICENSE file in the repo; permissive, attribution and no-endorsement conditions only.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/oxpig/AbLang2/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Use for antibody-aware likelihoods, germline-ness scoring and restoring missing residues in repertoire data. Trained on natural repertoires, so it penalises unusual-but-valid designed CDRs.", "sources": [ "https://raw.githubusercontent.com/oxpig/AbLang2/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "igfold", "name": "IgFold", "category": "antibody-specific design", "one_line": "Fast antibody Fv structure prediction from sequence.", "what_it_does": "Antibody Fv structure prediction. Input: heavy and light chain sequences; output: an Fv PDB with per-residue error estimates, in under a minute - fast enough for whole design libraries, but the licence is academic-only.", "maker": "Gray lab, Johns Hopkins University", "url": "https://github.com/Graylab/IgFold", "paper_url": "https://www.nature.com/articles/s41467-023-38063-x", "license": "JHU Academic Software License Agreement", "commercial_use": "no", "commercial_use_note": "LICENSE.md states only a non-exclusive, nontransferable license for non-commercial purposes; 'Commercial use of the Software requires a separately executed written license agreement.' Derivatives are bound by the same non-commercial terms.", "commercial_cost": "none", "commercial_cost_note": "No fee is charged, but no commercial right is granted either: the JHU Academic Software License covers non-profit academic/research institutions only. A company would have to negotiate separately with JHU; no published commercial price exists. ImmuneBuilder or ABodyBuilder3 avoid the problem entirely.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/Graylab/IgFold/main/LICENSE.md" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Only for academic work: fast Fv models in seconds. For any commercial antibody program use ImmuneBuilder or ABodyBuilder3 instead.", "sources": [ "https://raw.githubusercontent.com/Graylab/IgFold/main/LICENSE.md" ], "verified_at": "2026-10-15" }, { "id": "immunebuilder", "name": "ImmuneBuilder (ABodyBuilder2, NanoBodyBuilder2)", "category": "antibody-specific design", "one_line": "Fast structure prediction for antibodies, nanobodies and TCRs.", "what_it_does": "Fast immune-receptor structure prediction. Input: heavy/light (ABodyBuilder2), nanobody (NanoBodyBuilder2) or TCR sequences; output: an all-atom Fv/VHH/TCR PDB in seconds, refined with OpenMM.", "maker": "OPIG, University of Oxford", "url": "https://github.com/oxpig/ImmuneBuilder", "paper_url": "https://www.nature.com/articles/s42003-023-04927-7", "license": "BSD 3-Clause License", "commercial_use": "yes", "commercial_use_note": "BSD 3-Clause LICENSE file in the repo; permissive, commercial use allowed.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/oxpig/ImmuneBuilder/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "The practical, commercially-usable choice for bulk Fv and VHH modelling, seconds per structure. CDR-H3 accuracy remains the weak point, as for all Fv predictors.", "sources": [ "https://raw.githubusercontent.com/oxpig/ImmuneBuilder/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "abodybuilder3", "name": "ABodyBuilder3", "category": "antibody-specific design", "one_line": "Updated antibody structure predictor with per-residue error estimates.", "what_it_does": "Updated antibody Fv predictor. Input: heavy and light chain sequences; output: an Fv structure plus a per-residue predicted error (RMSD) estimate, so you can tell which CDR loops to trust before docking.", "maker": "Exscientia", "url": "https://github.com/Exscientia/abodybuilder3", "paper_url": "https://academic.oup.com/bioinformatics/article/40/10/btae576/7762510", "license": "Apache License 2.0", "commercial_use": "yes", "commercial_use_note": "LICENSE.txt in the repo is Apache-2.0; permissive, commercial use allowed.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/Exscientia/abodybuilder3/main/LICENSE.txt" ], "open_weights": "unverified", "compute": "single-gpu", "when_to_use": "Use when you want Fv models plus a calibrated confidence estimate per residue. Weight-file terms were not separately stated, and like other Fv models CDR-H3 is the limiting region.", "sources": [ "https://raw.githubusercontent.com/Exscientia/abodybuilder3/main/LICENSE.txt", "https://github.com/Exscientia/abodybuilder3/blob/main/LICENSE.txt" ], "verified_at": "2026-10-15" }, { "id": "anarci", "name": "ANARCI", "category": "antibody-specific design", "one_line": "Numbers antibody and TCR sequences (IMGT, Kabat, Chothia).", "what_it_does": "Antibody numbering. Input: antibody or TCR sequences (FASTA); output: the same sequences numbered in IMGT/Kabat/Chothia/Martin with germline V/J assignment and CDR boundaries - the preprocessing step almost every antibody tool expects.", "maker": "OPIG, University of Oxford", "url": "https://github.com/oxpig/ANARCI", "paper_url": "https://academic.oup.com/bioinformatics/article/32/2/298/1743894", "license": "BSD 3-Clause style license (LICENCE file, 2019 Deane et al.)", "commercial_use": "yes", "commercial_use_note": "The LICENCE file contains BSD 3-Clause redistribution terms with no field-of-use restriction; GitHub's license detection also reports BSD-3-Clause.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/oxpig/ANARCI/master/LICENCE" ], "open_weights": true, "compute": "cpu", "when_to_use": "The basic plumbing for any antibody pipeline: numbering, CDR definition, species assignment. Depends on HMMER germline HMMs and can fail on heavily engineered or truncated sequences.", "sources": [ "https://raw.githubusercontent.com/oxpig/ANARCI/master/LICENCE", "https://api.github.com/repos/oxpig/ANARCI/license" ], "verified_at": "2026-10-15" }, { "id": "anarcii", "name": "ANARCII", "category": "antibody-specific design", "one_line": "Language-model successor to ANARCI for immune receptor numbering.", "what_it_does": "Language-model antibody/TCR numbering. Input: immune receptor sequences; output: numbered sequences with CDR annotation, handling unusual, truncated and non-human sequences where HMM-based ANARCI fails.", "maker": "OPIG, University of Oxford", "url": "https://github.com/oxpig/ANARCII", "paper_url": "https://www.biorxiv.org/content/10.1101/2025.04.16.648720v1", "license": "BSD 3-Clause License", "commercial_use": "yes", "commercial_use_note": "LICENCE file reported by the GitHub license API as BSD-3-Clause; permissive.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/oxpig/ANARCII/main/LICENCE" ], "open_weights": true, "compute": "cpu", "when_to_use": "Use when ANARCI's HMM approach fails, e.g. unusual species, nanobodies or noisy sequences. Newer and less battle-tested than ANARCI.", "sources": [ "https://api.github.com/repos/oxpig/ANARCII/license", "https://raw.githubusercontent.com/oxpig/ANARCII/main/LICENCE" ], "verified_at": "2026-10-15" }, { "id": "humatch", "name": "Humatch", "category": "developability and sequence-liability prediction", "one_line": "Humanness scoring and humanisation guidance for antibody sequences.", "what_it_does": "Humanness scoring and humanisation. Input: heavy/light antibody sequences; output: a humanness score per chain plus suggested germline-directed mutations, so you can de-risk a non-human design before ordering it.", "maker": "OPIG, University of Oxford", "url": "https://github.com/oxpig/Humatch", "paper_url": "https://www.tandfonline.com/doi/full/10.1080/19420862.2024.2434121", "license": "BSD 3-Clause License", "commercial_use": "yes", "commercial_use_note": "BSD 3-Clause LICENSE file on the repo's master branch; permissive.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/oxpig/Humatch/master/LICENSE" ], "open_weights": true, "compute": "cpu", "when_to_use": "Use to check how human-like a designed or murine-derived sequence looks before humanisation work. Humanness scores correlate with, but do not measure, clinical immunogenicity.", "sources": [ "https://raw.githubusercontent.com/oxpig/Humatch/master/LICENSE", "https://github.com/oxpig/Humatch/blob/master/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "tap", "name": "TAP (Therapeutic Antibody Profiler)", "category": "developability and sequence-liability prediction", "one_line": "Flags antibody developability risks against clinical-stage antibody ranges.", "what_it_does": "Developability red flags. Input: an antibody Fv structure (or sequence to be modelled); output: five TAP metrics - CDR length, patches of surface hydrophobicity, positive and negative charge, and charge asymmetry - each placed against the distribution of clinical-stage antibodies.", "maker": "OPIG, University of Oxford (SAbPred web server)", "url": "https://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/sabpred/tap", "paper_url": "https://www.pnas.org/doi/10.1073/pnas.1810576116", "license": "unverified", "commercial_use": "unverified", "commercial_use_note": "The SAbPred TAP web page states no license or terms of use, and no public source repository license was found; terms for commercial use are therefore unverified.", "commercial_cost": "unverified", "commercial_cost_note": "The TAP web server is offered for free use through OPIG, but no licence or terms page was found that states whether a company may use it, so neither the right nor any commercial fee is settled. Treat as unknown.", "commercial_cost_sources": [ "https://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/sabpred/tap" ], "open_weights": "unverified", "compute": "hosted-service", "when_to_use": "Quick red-flag check (CDR length, patch hydrophobicity, charge) on a handful of candidates. It is a comparison against clinical-stage distributions, not a measurement, and the web server is rate-limited for bulk work.", "sources": [ "https://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/sabpred/tap", "https://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/sabpred/" ], "verified_at": "2026-10-15" }, { "id": "netsolp", "name": "NetSolP", "category": "developability and sequence-liability prediction", "one_line": "Predicts solubility and usability of proteins from sequence.", "what_it_does": "Solubility and expression risk. Input: protein sequences (FASTA); output: predicted solubility and 'usability' probabilities per sequence, for triaging designs before paying for expression.", "maker": "DTU Health Tech (Technical University of Denmark)", "url": "https://github.com/tvinet/NetSolP-1.0", "paper_url": "https://academic.oup.com/bioinformatics/article/38/4/941/6444984", "license": "BSD 3-Clause License", "commercial_use": "yes", "commercial_use_note": "BSD 3-Clause LICENSE file in the repo (Copyright 2021, Technical University of Denmark); permissive.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/tvinet/NetSolP-1.0/main/LICENSE" ], "open_weights": "unverified", "compute": "single-gpu", "when_to_use": "Cheap pre-expression triage of many constructs for solubility risk. Trained largely on E. coli expression data, so it transfers poorly to mammalian-expressed antibodies.", "sources": [ "https://raw.githubusercontent.com/tvinet/NetSolP-1.0/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "thermompnn", "name": "ThermoMPNN", "category": "developability and sequence-liability prediction", "one_line": "Predicts stability change (ddG) for single point mutations.", "what_it_does": "Stability effect of point mutations. Input: a protein structure (PDB); output: predicted ddG for single-point mutations (one site or the full saturation scan), used to pick stabilising mutations for a designed scaffold.", "maker": "Kuhlman lab, UNC Chapel Hill", "url": "https://github.com/Kuhlman-Lab/ThermoMPNN", "paper_url": "https://www.pnas.org/doi/10.1073/pnas.2314853121", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "MIT LICENSE file in the repo.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/Kuhlman-Lab/ThermoMPNN/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Use to rank stabilising point mutations across a scan quickly. Trained on single mutations in globular proteins; it does not handle multi-mutant epistasis or antibody-specific effects well.", "sources": [ "https://raw.githubusercontent.com/Kuhlman-Lab/ThermoMPNN/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "autodock-vina", "name": "AutoDock Vina", "category": "docking and affinity estimation", "one_line": "Classic small-molecule docking with an empirical scoring function.", "what_it_does": "Small-molecule docking. Input: receptor PDBQT + ligand + search box; output: docked poses with an empirical affinity score - the baseline for pocket-level triage, not for protein-protein work.", "maker": "Scripps Research (CCSB)", "url": "https://github.com/ccsb-scripps/AutoDock-Vina", "paper_url": "https://pubs.acs.org/doi/10.1021/acs.jcim.1c00203", "license": "Apache License 2.0", "commercial_use": "yes", "commercial_use_note": "Apache-2.0 LICENSE file in the repo; permissive.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/ccsb-scripps/AutoDock-Vina/develop/LICENSE" ], "open_weights": true, "compute": "cpu", "when_to_use": "Still the default for cheap small-molecule pose generation on CPU. Its scores rank poses poorly as affinities and it is not meant for protein-protein or antibody work.", "sources": [ "https://raw.githubusercontent.com/ccsb-scripps/AutoDock-Vina/develop/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "gnina", "name": "GNINA", "category": "docking and affinity estimation", "one_line": "Molecular docking with CNN-based pose scoring.", "what_it_does": "Docking with a CNN rescorer. Input: receptor + ligand (and optionally a reference pose); output: poses scored by both the Vina-style function and a convolutional neural network, which usually improves pose ranking.", "maker": "Koes lab, University of Pittsburgh", "url": "https://github.com/gnina/gnina", "paper_url": "https://jcheminf.biomedcentral.com/articles/10.1186/s13321-021-00522-2", "license": "Apache License 2.0 (LICENSE.APACHE)", "commercial_use": "yes", "commercial_use_note": "The repo's license file LICENSE.APACHE is Apache-2.0, reported by the GitHub license API. Bundled dependencies (e.g. OpenBabel) may carry their own copyleft terms and should be checked before redistribution.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/gnina/gnina/master/LICENSE.APACHE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Use when CNN rescoring improves pose selection over Vina. Affinity output is still a score, not a KD, and GPU builds can be awkward to install.", "sources": [ "https://raw.githubusercontent.com/gnina/gnina/master/LICENSE.APACHE", "https://api.github.com/repos/gnina/gnina/license" ], "verified_at": "2026-10-15" }, { "id": "prodigy", "name": "PRODIGY", "category": "docking and affinity estimation", "one_line": "Estimates binding affinity of protein-protein complexes from structure.", "what_it_does": "Affinity estimate for a protein-protein complex. Input: a complex structure (PDB); output: a predicted binding affinity (dG/Kd estimate) from interfacial contact composition - a structure-based ranking signal, not a measurement.", "maker": "Bonvin lab, Utrecht University", "url": "https://github.com/haddocking/prodigy", "paper_url": "https://elifesciences.org/articles/07454", "license": "Apache License 2.0", "commercial_use": "yes", "commercial_use_note": "Apache-2.0 LICENSE file reported by the GitHub license API for haddocking/prodigy.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/haddocking/prodigy/main/LICENSE" ], "open_weights": true, "compute": "cpu", "when_to_use": "Fast contact-based ΔG estimate to rank protein-protein complexes. It was fitted on experimental crystal complexes, so predictions on modelled or designed interfaces are rough at best and never a substitute for SPR/BLI.", "sources": [ "https://api.github.com/repos/haddocking/prodigy/license", "https://raw.githubusercontent.com/haddocking/prodigy/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "openmm", "name": "OpenMM", "category": "MD and refinement", "one_line": "GPU-accelerated molecular dynamics toolkit with a Python API.", "what_it_does": "Molecular dynamics engine with a Python API. Input: a structure plus a force field; output: trajectories, energies and minimised structures. Used inside design pipelines for relaxation and side-chain/loop refinement.", "maker": "Stanford University and the OpenMM community", "url": "https://github.com/openmm/openmm", "paper_url": "https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1005659", "license": "Mixed: MIT for most of OpenMM, with some parts under LGPL (see Licenses.txt)", "commercial_use": "yes", "commercial_use_note": "Licenses.txt states most of OpenMM is under the MIT License while certain portions are under the LGPL (with GPL referenced for those parts). Commercial use is allowed; redistribution of the LGPL portions carries LGPL obligations.", "commercial_cost": "free", "commercial_cost_note": "Free. Most of OpenMM is MIT; some components are LGPL, which costs nothing but obliges you to allow relinking if you distribute a modified build.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/openmm/openmm/master/docs-source/licenses/Licenses.txt" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Use for relaxation, minimisation and MD of designed complexes, including AlphaFold-style structure relaxation. MD stability is not affinity, and useful sampling of an interface is expensive.", "sources": [ "https://raw.githubusercontent.com/openmm/openmm/master/docs-source/licenses/Licenses.txt" ], "verified_at": "2026-10-15" }, { "id": "gromacs", "name": "GROMACS", "category": "MD and refinement", "one_line": "High-performance molecular dynamics engine for biomolecular simulation.", "what_it_does": "Production MD engine. Input: structure + topology + run parameters; output: long trajectories with free-energy and analysis tooling - for stability, aggregation and interface-dynamics studies of a design.", "maker": "GROMACS development team (KTH, Uppsala and others)", "url": "https://www.gromacs.org/", "paper_url": "https://www.sciencedirect.com/science/article/pii/S2352711015000059", "license": "GNU Lesser General Public License v2.1", "commercial_use": "yes", "commercial_use_note": "The COPYING file in the official GitLab repository is LGPL v2.1. Commercial use is permitted; distributing modified GROMACS libraries triggers LGPL source-sharing obligations.", "commercial_cost": "free", "commercial_cost_note": "Free under LGPL-2.1. No fee; if you distribute a modified GROMACS you must release those modifications under the same terms.", "commercial_cost_sources": [ "https://gitlab.com/gromacs/gromacs/-/raw/main/COPYING" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Reach for it for long or large production MD where throughput matters. Steeper learning curve than OpenMM and overkill for simple minimisation of design models.", "sources": [ "https://gitlab.com/gromacs/gromacs/-/raw/main/COPYING" ], "verified_at": "2026-10-15" }, { "id": "colabfold", "name": "ColabFold", "category": "structure and complex prediction", "one_line": "Fast AlphaFold2 with MMseqs2-based MSAs, runnable in Colab.", "what_it_does": "AlphaFold2/multimer with fast MSAs. Input: FASTA (or a Colab notebook cell); output: ranked predicted structures with pLDDT/PAE plots, using MMseqs2 remote MSA search instead of a local database.", "maker": "Mirdita, Ovchinnikov, Steinegger et al.", "url": "https://github.com/sokrypton/ColabFold", "paper_url": "https://www.nature.com/articles/s41592-022-01488-1", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "MIT LICENSE file in the repo. The AlphaFold2 parameters it downloads are CC BY 4.0 per the AlphaFold README. Note the free public MSA server is intended for modest academic-scale use; heavy or commercial batch use should run a local MMseqs2 server.", "commercial_cost": "free", "commercial_cost_note": "The software is free (MIT) and the public MSA server costs nothing, though it is rate-limited and fair-use. Heavy or production use means running your own MSA server or paying for your own compute, not a licence fee.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/sokrypton/ColabFold/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "The cheapest practical way to get AF2 predictions, including on free Colab GPUs. MSA depth from the public server is lower than full AF2 databases, and the server is rate-limited.", "sources": [ "https://raw.githubusercontent.com/sokrypton/ColabFold/main/LICENSE", "https://raw.githubusercontent.com/google-deepmind/alphafold/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "openfold", "name": "OpenFold", "category": "structure and complex prediction", "one_line": "Trainable open reimplementation of AlphaFold 2.", "what_it_does": "Trainable AlphaFold2 reimplementation. Input: FASTA/MSAs; output: predicted structures, with the ability to fine-tune or retrain on your own data - which the original AlphaFold code does not give you.", "maker": "AlQuraishi lab, Columbia University", "url": "https://github.com/aqlaboratory/openfold", "paper_url": "https://www.nature.com/articles/s41592-024-02272-z", "license": "Apache License 2.0", "commercial_use": "yes", "commercial_use_note": "Apache-2.0 LICENSE file in the repo. The README notes DeepMind's pretrained parameters fall under CC BY 4.0, replacing the earlier CC BY-NC 4.0 terms; OpenFold's own trained weights are distributed with the project.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/aqlaboratory/openfold/main/LICENSE" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Use when you need to fine-tune or retrain an AF2-class model on your own data rather than just run inference. Training is genuinely expensive in GPU time.", "sources": [ "https://raw.githubusercontent.com/aqlaboratory/openfold/main/LICENSE", "https://raw.githubusercontent.com/aqlaboratory/openfold/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "pyrosetta", "name": "PyRosetta", "category": "MD and refinement", "one_line": "Python interface to Rosetta for protein modelling and design.", "what_it_does": "Rosetta's Python API. Input: structures and movers you script; output: relaxed/minimised and scored structures, interface metrics (ddG), loop remodelling and packing. Dependency of several design pipelines.", "maker": "RosettaCommons / University of Washington", "url": "https://www.pyrosetta.org/", "paper_url": "https://academic.oup.com/bioinformatics/article/26/5/689/212442", "license": "RosettaCommons license (free academic/non-profit; paid commercial)", "commercial_use": "restricted", "commercial_use_note": "The official licensing page states licensing is free for academic and non-profit institutions and available to commercial users for a fee, handled through Rosetta Commons. A commercial entity needs a paid license.", "commercial_cost": "paid-license", "commercial_cost_note": "Free for academic and non-profit users; commercial users pay a licence fee, quoted by RosettaCommons/UW on request rather than published as a list price. This is the single most common hidden cost in otherwise-free design pipelines (BindCraft, several Rosetta-based filters).", "commercial_cost_sources": [ "https://www.pyrosetta.org/home/licensing-pyrosetta" ], "open_weights": true, "compute": "cpu", "when_to_use": "Still the most complete toolbox for interface scoring, relaxation, loop modelling and FastRelax used by many design pipelines. The commercial license cost is the main blocker, and energies are not affinities.", "sources": [ "https://www.pyrosetta.org/home/licensing-pyrosetta" ], "verified_at": "2026-10-15" }, { "id": "mmseqs2", "name": "MMseqs2", "category": "data and structure resources", "one_line": "Very fast sequence search and clustering, the MSA engine behind ColabFold.", "what_it_does": "Sequence search and clustering at scale. Input: query and target FASTA/databases; output: homology hits, clusters and MSA profiles in minutes where BLAST takes days - the MSA engine behind ColabFold.", "maker": "Steinegger and Söding labs", "url": "https://github.com/soedinglab/MMseqs2", "paper_url": "https://www.nature.com/articles/nbt.3988", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "LICENSE.md in the repo is 'The MIT License (MIT)'.", "commercial_cost": "free", "commercial_cost_note": "Nothing to pay: the licence read at the source grants commercial use with no fee, and the code and weights are public downloads. You still pay for your own compute.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/soedinglab/MMseqs2/master/LICENSE.md" ], "open_weights": true, "compute": "cpu", "when_to_use": "Use to build your own MSAs or cluster large sequence sets locally instead of hitting public servers. Needs lots of RAM and disk for big databases.", "sources": [ "https://raw.githubusercontent.com/soedinglab/MMseqs2/master/LICENSE.md" ], "verified_at": "2026-10-15" }, { "id": "foldseek", "name": "Foldseek", "category": "data and structure resources", "one_line": "Fast structural similarity search across millions of structures.", "what_it_does": "Structure search. Input: a query structure; output: structurally similar entries across PDB/AFDB with alignment and E-values - how you check whether a designed fold or epitope already exists.", "maker": "Steinegger lab, Seoul National University", "url": "https://github.com/steineggerlab/foldseek", "paper_url": "https://www.nature.com/articles/s41587-023-01773-0", "license": "GNU General Public License v3.0", "commercial_use": "restricted", "commercial_use_note": "LICENSE.md in the repo is GPL-3.0. Internal commercial use is allowed, but distributing software that incorporates Foldseek requires releasing the derivative work under GPL-3.0, which matters if you plan to ship a product containing it.", "commercial_cost": "free", "commercial_cost_note": "No fee, but GPL-3.0: if you ship Foldseek inside a product you distribute, you must release that product's source under GPL-3.0. Running it internally as a service carries no such obligation.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/steineggerlab/foldseek/master/LICENSE.md" ], "open_weights": true, "compute": "cpu", "when_to_use": "Use to find structural neighbours of a designed backbone, e.g. novelty checks and self-similarity screening. GPL copyleft makes embedding it in a closed product awkward.", "sources": [ "https://raw.githubusercontent.com/steineggerlab/foldseek/master/LICENSE.md", "https://api.github.com/repos/steineggerlab/foldseek/license" ], "verified_at": "2026-10-15" }, { "id": "pymol-open-source", "name": "Open-Source PyMOL", "category": "data and structure resources", "one_line": "Molecular visualisation for inspecting structures and interfaces.", "what_it_does": "Molecular visualisation and measurement. Input: structures and trajectories; output: rendered images/sessions plus interface inspection - contacts, distances, buried surface - for sanity-checking a design by eye.", "maker": "Schrödinger, LLC (open-source build)", "url": "https://github.com/schrodinger/pymol-open-source", "paper_url": null, "license": "Open-Source PyMOL Copyright Notice (custom permissive, MIT/X11-style)", "commercial_use": "yes", "commercial_use_note": "The LICENSE file grants permission to use, copy, modify and distribute for any purpose without fee, provided copyright notices are kept; the PyMOL name may not be used in advertising without permission, and derivatives must be plainly distinguished from Schrödinger's PyMOL products. The paid Incentive PyMOL product is separate.", "commercial_cost": "free", "commercial_cost_note": "The open-source build is free for any use. The commercial PyMOL sold by Schrodinger is a separate, paid product; you do not need it, and incentive/educational builds have their own restrictions.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/schrodinger/pymol-open-source/master/LICENSE" ], "open_weights": true, "compute": "cpu", "when_to_use": "Use for visual QC of designs, epitopes and docked poses. The open-source build lacks some features and support of the commercial distribution.", "sources": [ "https://raw.githubusercontent.com/schrodinger/pymol-open-source/master/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "pdb", "name": "Protein Data Bank (RCSB PDB)", "category": "data and structure resources", "one_line": "The archive of experimentally determined biomolecular structures.", "what_it_does": "Primary structure archive. Input: a PDB ID or search query; output: experimentally determined coordinates and metadata - the source of your target structure and of nearly every training set in this list.", "maker": "wwPDB / RCSB PDB", "url": "https://www.rcsb.org/", "paper_url": "https://academic.oup.com/nar/article/28/1/235/2384399", "license": "CC0 1.0 Universal (archive data)", "commercial_use": "yes", "commercial_use_note": "The RCSB usage policy page states the wwPDB policy makes PDB archive data files available under CC0 1.0, as is data from RCSB programmatic APIs. RCSB software products are separately under MIT or Apache-2.0.", "commercial_cost": "free", "commercial_cost_note": "Free, CC0: no fee, no attribution requirement (citation is requested as courtesy). Bulk download via FTP/rsync is open.", "commercial_cost_sources": [ "https://www.rcsb.org/pages/usage-policy" ], "open_weights": true, "compute": "hosted-service", "when_to_use": "The ground truth for benchmarks, templates and epitope analysis. Coverage is biased to crystallisable, well-behaved proteins and many therapeutic complexes are simply absent.", "sources": [ "https://www.rcsb.org/pages/usage-policy" ], "verified_at": "2026-10-15" }, { "id": "sabdab", "name": "SAbDab", "category": "data and structure resources", "one_line": "Curated, annotated database of antibody structures from the PDB.", "what_it_does": "Antibody structure database. Input: search by antigen, species, CDR length, affinity or date; output: curated, numbered antibody/nanobody structures with annotations - the benchmark set behind antibody design papers.", "maker": "OPIG, University of Oxford", "url": "https://sabdab.opig.stats.ox.ac.uk/", "paper_url": "https://academic.oup.com/nar/article/42/D1/D1140/1044118", "license": "Creative Commons Attribution 4.0 International (CC BY 4.0)", "commercial_use": "yes", "commercial_use_note": "The SAbDab site states the data is made available under CC BY 4.0 (link to creativecommons.org/licenses/by/4.0/ in the site's about text); attribution required.", "commercial_cost": "free", "commercial_cost_note": "Free to download and use commercially under CC BY 4.0; the cost is attribution.", "commercial_cost_sources": [ "https://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/sabdab/" ], "open_weights": true, "compute": "hosted-service", "when_to_use": "Best starting point for antibody structure datasets with numbering, species and antigen annotation. It inherits PDB's coverage bias and annotation errors do occur.", "sources": [ "https://sabdab.opig.stats.ox.ac.uk/", "https://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/sabdab/about/" ], "verified_at": "2026-10-15" }, { "id": "oas", "name": "Observed Antibody Space (OAS)", "category": "data and structure resources", "one_line": "Over a billion annotated antibody repertoire sequences, paired and unpaired.", "what_it_does": "Antibody repertoire sequences. Input: a query over studies/species/chains; output: over a billion annotated BCR sequences, paired and unpaired, for training language models and checking whether a CDR looks natural.", "maker": "OPIG, University of Oxford", "url": "https://opig.stats.ox.ac.uk/webapps/oas/", "paper_url": "https://onlinelibrary.wiley.com/doi/10.1002/pro.4205", "license": "Creative Commons Attribution 4.0 International (CC BY 4.0)", "commercial_use": "yes", "commercial_use_note": "The OAS homepage states: 'Please note that the data contained within OAS is available under a CC-BY 4.0 license.'", "commercial_cost": "free", "commercial_cost_note": "Free under CC BY 4.0, including for commercial training runs; attribute OAS and the underlying studies.", "commercial_cost_sources": [ "https://opig.stats.ox.ac.uk/webapps/oas/" ], "open_weights": true, "compute": "hosted-service", "when_to_use": "Use for training or benchmarking antibody language models and for germline/naturalness baselines. Study-level batch effects and uneven donor coverage mean naive frequency statistics mislead.", "sources": [ "https://opig.stats.ox.ac.uk/webapps/oas/" ], "verified_at": "2026-10-15" }, { "id": "alphafold-db", "name": "AlphaFold Protein Structure Database", "category": "data and structure resources", "one_line": "Precomputed AlphaFold structure predictions for most known proteins.", "what_it_does": "Precomputed structure predictions. Input: a UniProt accession; output: the AlphaFold model with pLDDT/PAE - the fastest way to get a usable structure for a target with no crystal.", "maker": "Google DeepMind and EMBL-EBI", "url": "https://alphafold.ebi.ac.uk/", "paper_url": "https://academic.oup.com/nar/article/50/D1/D439/6430488", "license": "Creative Commons Attribution 4.0 International (CC BY 4.0)", "commercial_use": "yes", "commercial_use_note": "The AlphaFold DB homepage and download page both state: 'Data is available for academic and commercial use, under a CC-BY-4.0 licence.'", "commercial_cost": "free", "commercial_cost_note": "Free under CC BY 4.0 - note this is the database of predictions, which is licensed far more openly than the AlphaFold 3 weights.", "commercial_cost_sources": [ "https://alphafold.ebi.ac.uk/about" ], "open_weights": true, "compute": "hosted-service", "when_to_use": "Use to grab a target model instantly instead of folding it yourself. These are predictions with pLDDT confidence, not experimental structures, and low-pLDDT regions should not be treated as real conformations.", "sources": [ "https://alphafold.ebi.ac.uk/", "https://alphafold.ebi.ac.uk/download" ], "verified_at": "2026-10-15" }, { "id": "uniprot", "name": "UniProt", "category": "data and structure resources", "one_line": "Reference protein sequence and functional annotation database.", "what_it_does": "Reference sequences and annotation. Input: accession, gene or text query; output: canonical sequences, isoforms, domains, PTMs and functional annotation - the identity layer for any target you design against.", "maker": "UniProt Consortium (EMBL-EBI, SIB, PIR)", "url": "https://www.uniprot.org/", "paper_url": "https://academic.oup.com/nar/article/51/D1/D523/6835362", "license": "Creative Commons Attribution 4.0 International (CC BY 4.0)", "commercial_use": "yes", "commercial_use_note": "The UniProt license help page states CC BY 4.0 applies to all copyrightable parts of their databases, with a disclaimer that some data may be covered by patents or other rights.", "commercial_cost": "free", "commercial_cost_note": "Free under CC BY 4.0, bulk downloads included.", "commercial_cost_sources": [ "https://www.uniprot.org/help/license" ], "open_weights": true, "compute": "hosted-service", "when_to_use": "Canonical source for target sequences, isoforms, domains and PTM annotation before any design run. Annotation quality varies sharply between reviewed (Swiss-Prot) and unreviewed (TrEMBL) entries.", "sources": [ "https://rest.uniprot.org/help/license", "https://www.uniprot.org/help/license" ], "verified_at": "2026-10-15" }, { "id": "rfdiffusion2", "name": "RFdiffusion2", "category": "de novo binder and protein design", "one_line": "Atom-level successor to RFdiffusion, aimed at enzyme and motif scaffolding.", "what_it_does": "Generates backbones around an atomic motif without having to specify its indices. Input: a set of catalytic/functional atoms or a target site; output: backbone PDBs scaffolding that geometry, which you then sequence-design (LigandMPNN) and refold. Inference only in this release.", "maker": "Institute for Protein Design, University of Washington (RosettaCommons)", "url": "https://github.com/RosettaCommons/RFdiffusion2", "paper_url": "https://www.biorxiv.org/content/10.1101/2025.04.09.648075v1", "license": "BSD 3-Clause License (code); no licence stated for the downloaded weights", "commercial_use": "unverified", "commercial_use_note": "The repo LICENSE.md is BSD 3-Clause (2025, Institute for Protein Design) and mentions only source and binary forms; the README's setup step downloads large model weight files and containers with no licence statement of their own, so the weights' commercial terms are unverified. The README also marks the release 'under construction'.", "commercial_cost": "unverified", "commercial_cost_note": "Nothing is charged for the download. Whether a company may use the weights at all is unsettled, so no price can be stated.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/RosettaCommons/RFdiffusion2/main/README.md" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Reach for it over RFdiffusion v1 when the design problem is an atom-level site (enzyme active site, metal or ligand geometry) rather than a plain binder. Early release, inference only, and the weight licence question is unresolved - for commercial work that matters.", "sources": [ "https://raw.githubusercontent.com/RosettaCommons/RFdiffusion2/main/LICENSE.md", "https://raw.githubusercontent.com/RosettaCommons/RFdiffusion2/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "protenix", "name": "Protenix", "category": "structure and complex prediction", "one_line": "Open AlphaFold3-class complex predictor from ByteDance, Apache-2.0 for code and v1 parameters.", "what_it_does": "Predicts all-atom structures of biomolecular complexes. Input: a JSON/YAML of protein, nucleic-acid and ligand entities (optionally MSAs); output: ranked all-atom complexes with confidence scores. Also ships lightweight 'Mini' variants that cut inference cost.", "maker": "ByteDance (AML AI for Science)", "url": "https://github.com/bytedance/Protenix", "paper_url": "https://www.biorxiv.org/content/10.1101/2025.01.08.631967v1", "license": "Apache License 2.0 (code and released model parameters)", "commercial_use": "yes", "commercial_use_note": "The README's License section states the project, code and model parameters, is released under Apache 2.0 and is 'free for both academic research and commercial use'. Important exception read in the same README: the newer Protenix-v2 weights are described as proprietary and confidential, not released under any open-source licence - so the free commercial grant covers the openly released checkpoints, not v2.", "commercial_cost": "free", "commercial_cost_note": "No fee for the open checkpoints - the README states Apache 2.0 for code and parameters and explicitly says commercial use is free. Protenix-v2 is not available at any published price.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/bytedance/Protenix/main/README.md" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "A third commercially usable AlphaFold3 substitute alongside Boltz-2 and Chai-1, with a stated antibody-antigen focus in later versions. Use the openly licensed checkpoints; do not assume v2 weights come with the same rights.", "sources": [ "https://raw.githubusercontent.com/bytedance/Protenix/main/LICENSE", "https://raw.githubusercontent.com/bytedance/Protenix/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "chai-2", "name": "Chai-2", "category": "antibody-specific design", "one_line": "Chai Discovery's zero-shot antibody and miniprotein design model - access by partnership only.", "what_it_does": "Designs de novo antibodies, nanobodies and miniproteins against a target epitope, reported to need only ~20 designs per target. Input (as described by the vendor): a target antigen and epitope; output: scFv/VHH/miniprotein designs. There is no public code, no downloadable weights and no self-serve API.", "maker": "Chai Discovery", "url": "https://www.chaidiscovery.com/news/introducing-chai-2", "paper_url": "https://www.chaidiscovery.com/news/introducing-chai-2", "license": "unverified (no public licence; access granted case-by-case under the vendor's Responsible Deployment policy)", "commercial_use": "unverified", "commercial_use_note": "No licence text exists to read. The announcement says access is being opened to 'a number of academic and industry partners' via an early-access sign-up and granted under a Responsible Deployment policy. Terms for a commercial partner are not published, so rights are unverified. Listed here because it is widely cited and people ask - not because it is usable off the shelf.", "commercial_cost": "unverified", "commercial_cost_note": "No price is published. Access is by application, and nothing in the announcement states whether partners pay, grant rights, or share results.", "commercial_cost_sources": [ "https://www.chaidiscovery.com/news/introducing-chai-2" ], "open_weights": false, "compute": "hosted-service", "when_to_use": "Only relevant if you want to approach Chai as a partner. For in-house work today, the open antibody designers (RFantibody, Germinal, IgGM, mBER) are what you can actually run.", "sources": [ "https://www.chaidiscovery.com/news/introducing-chai-2" ], "verified_at": "2026-10-15" }, { "id": "esm3", "name": "ESM3", "category": "protein language models", "one_line": "Generative multimodal protein model reasoning jointly over sequence, structure and function.", "what_it_does": "Generates and edits proteins across three tracks at once. Input: any partial specification - sequence, backbone coordinates, secondary structure, SASA or function keywords, with the rest masked; output: the completed sequence and/or structure. Used for prompted generation and scaffolding rather than plain embedding extraction.", "maker": "EvolutionaryScale (repository now copyright Chan Zuckerberg Biohub)", "url": "https://github.com/evolutionaryscale/esm", "paper_url": "https://www.science.org/doi/10.1126/science.ads0018", "license": "MIT License (LICENSE.md, Copyright 2026 Chan Zuckerberg Biohub, Inc.)", "commercial_use": "yes", "commercial_use_note": "The repository LICENSE.md is MIT and the README's Licenses section states 'These models are available under the MIT license' - a change from the earlier non-commercial Cambrian open licence. Larger checkpoints remain behind a hosted API; the open ones are the MIT-licensed download.", "commercial_cost": "free", "commercial_cost_note": "Free for the open checkpoints under MIT. The vendor's hosted API for the larger models is a separate paid service you are not required to use.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/evolutionaryscale/esm/main/README.md" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Use when you want prompted generation conditioned on structure or function, not just embeddings - that is what separates it from ESM-2 and ESM C. Open checkpoint is the small one; the headline results used much larger models served over the API.", "sources": [ "https://raw.githubusercontent.com/evolutionaryscale/esm/main/LICENSE.md", "https://raw.githubusercontent.com/evolutionaryscale/esm/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "genie2", "name": "Genie 2", "category": "de novo binder and protein design", "one_line": "Diffusion model for unconditional backbone generation and multi-motif scaffolding.", "what_it_does": "Generates protein backbones, optionally scaffolding several functional motifs at once. Input: a length range, or one or more motif PDB fragments with their desired spacing; output: backbone PDBs, which you then sequence-design and refold to check designability. Checkpoints ship in the v1.0.0 release.", "maker": "AQ Laboratory (Mohammed AlQuraishi lab, Columbia University)", "url": "https://github.com/aqlaboratory/genie2", "paper_url": "https://arxiv.org/abs/2405.15489", "license": "Apache License 2.0", "commercial_use": "yes", "commercial_use_note": "Apache-2.0 LICENSE in the repo. The README states the repository provides training and inference code 'as well as model weights', released from the same repository, and no separate weight terms are given - so the permissive grant is the only licence in play.", "commercial_cost": "free", "commercial_cost_note": "Free: Apache-2.0 code and checkpoints downloaded from the GitHub release, no fee.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/aqlaboratory/genie2/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "The practical option when you need multi-motif scaffolding, which RFdiffusion handles less directly. Research-grade: expect to filter heavily and to pair it with ProteinMPNN plus a refolding check.", "sources": [ "https://raw.githubusercontent.com/aqlaboratory/genie2/main/LICENSE", "https://raw.githubusercontent.com/aqlaboratory/genie2/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "proteina", "name": "Proteina", "category": "de novo binder and protein design", "one_line": "Large flow-matching backbone generator from NVIDIA - research and evaluation only.", "what_it_does": "Generates protein backbones at scale, including long chains and fold-class-conditioned samples. Input: length, fold class label or motif; output: backbone structures for downstream sequence design. Trained on a very large synthetic+PDB corpus.", "maker": "NVIDIA (BioNeMo)", "url": "https://github.com/NVIDIA-BioNeMo/proteina", "paper_url": "https://arxiv.org/abs/2503.00710", "license": "NVIDIA License (non-commercial)", "commercial_use": "no", "commercial_use_note": "The LICENSE file's section 3.3 'Use Limitation' states the work and any derivative works 'only may be used or intended for use non-commercially', defining non-commercially as 'for research or evaluation purposes only', with an explicit carve-out letting NVIDIA itself use it commercially. Unambiguous: a company cannot use this in product work.", "commercial_cost": "none", "commercial_cost_note": "No commercial option exists at any price: the licence limits use to research or evaluation and NVIDIA publishes no commercial licence for it.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/NVIDIA-BioNeMo/proteina/main/LICENSE" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Academic benchmarking of backbone generation, or inspiration for scale. For anything commercial, use RFdiffusion or Genie 2 instead - note that NVIDIA's own La-Proteina ships under different, commercially usable weight terms.", "sources": [ "https://raw.githubusercontent.com/NVIDIA-BioNeMo/proteina/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "la-proteina", "name": "La-Proteina", "category": "de novo binder and protein design", "one_line": "Partially latent flow-matching model generating sequence and all-atom structure jointly.", "what_it_does": "Generates a protein's sequence and full all-atom structure (backbone plus side chains) in one pass, including atomistic motif scaffolding. Input: length, or a motif with its atom coordinates; output: all-atom designs that do not need a separate inverse-folding step to get a sequence.", "maker": "NVIDIA (BioNeMo)", "url": "https://github.com/NVIDIA-BioNeMo/la-proteina", "paper_url": "https://arxiv.org/abs/2507.09466", "license": "Apache License 2.0 (code); NVIDIA Open Model License Agreement (weights); separate dataset licence", "commercial_use": "yes", "commercial_use_note": "This repo splits its licences into a LICENSE/ directory: license_code.txt is Apache-2.0, license_weights.txt is the NVIDIA Open Model License Agreement, and license_datasets.txt covers the data. The Open Model License is the permissive one NVIDIA uses to allow commercial use of model outputs and derivatives, in contrast to the plain 'NVIDIA License' on Proteina which is non-commercial - read the conditions (attribution, trustworthy-AI and no-reverse-engineering terms) before shipping.", "commercial_cost": "free", "commercial_cost_note": "No fee. Code Apache-2.0 and weights under the NVIDIA Open Model License, which grants use without payment subject to its conditions.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/NVIDIA-BioNeMo/la-proteina/main/LICENSE/license_weights.txt" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Attractive when you want sequence and side chains out of the generator itself, especially for atomistic motif scaffolding. Check the Open Model License conditions with counsel; they are permissive but not as plain as MIT.", "sources": [ "https://raw.githubusercontent.com/NVIDIA-BioNeMo/la-proteina/main/LICENSE/license_code.txt", "https://raw.githubusercontent.com/NVIDIA-BioNeMo/la-proteina/main/LICENSE/license_weights.txt", "https://github.com/NVIDIA-BioNeMo/la-proteina" ], "verified_at": "2026-10-15" }, { "id": "chroma", "name": "Chroma", "category": "de novo binder and protein design", "one_line": "Programmable protein generator with composable conditioners - weights are academic-only.", "what_it_does": "Generates all-atom proteins and complexes under programmable constraints. Input: 'conditioners' you compose - symmetry, shape, substructure, natural-language-style class conditioning; output: structures plus sequences, since it models both. Also does sequence design on given backbones and side-chain packing.", "maker": "Generate Biomedicines", "url": "https://github.com/generatebio/chroma", "paper_url": "https://www.nature.com/articles/s41586-023-06728-8", "license": "Apache License 2.0 (code); Chroma Parameters License (weights, academic/non-profit only)", "commercial_use": "restricted", "commercial_use_note": "The README's License section is explicit: code is Apache-2.0, but 'Chroma weights are freely available to academic researchers and non-profit entities' under the Chroma Parameters License, and anyone wanting to use the parameters 'beyond the scope of the rights granted in the License (including for commercial purposes)' is told to contact licensing@generatebiomedicines.com. So commercial use needs a negotiated licence from the company.", "commercial_cost": "paid-license", "commercial_cost_note": "Free for academics and non-profits; a company must contact Generate Biomedicines' licensing address for terms. No price is published, so assume a negotiated fee.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/generatebio/chroma/main/README.md", "https://chroma-weights.generatebiomedicines.com/" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Unmatched for constraint-driven generation (symmetry, shape, substructure) if you are academic. For a company, the conditioner idea is worth studying but the weights need a deal - plan on RFdiffusion/Genie 2 instead.", "sources": [ "https://raw.githubusercontent.com/generatebio/chroma/main/LICENSE.txt", "https://raw.githubusercontent.com/generatebio/chroma/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "mber", "name": "mBER", "category": "antibody-specific design", "one_line": "Open antibody binder design by backprop through AlphaFold-Multimer, format-specific.", "what_it_does": "Designs antibody binders in a chosen format. Input: target structure plus a framework/format template (VHH, Fab, scFv) and epitope; output: designed CDR sequences and predicted complexes, by optimising through AlphaFold-Multimer with sequence conditioning and structure templates.", "maker": "Manifold Bio", "url": "https://github.com/manifoldbio/mber-open", "paper_url": "https://www.biorxiv.org/content/10.1101/2025.02.25.640114v1", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "MIT LICENSE file (Copyright 2025 Manifold Bio), and the README's License section repeats MIT. Caveat to check yourself: its download_weights.sh pulls roughly 9 GB of third-party weights (AlphaFold2, NanoBodyBuilder2, ESM-2) whose own terms apply - AlphaFold2 parameters are CC BY 4.0 and the others are permissive, which is why this still reads as a yes.", "commercial_cost": "free", "commercial_cost_note": "No fee: MIT code and freely downloadable third-party weights. Your cost is GPU time, which for backprop-through-AF2 design is the real expense.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/manifoldbio/mber-open/main/LICENSE", "https://raw.githubusercontent.com/manifoldbio/mber-open/main/README.md" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "A good open alternative to RFantibody when you care about the final antibody format rather than a generic binder, and you already have AlphaFold2 infrastructure. Backprop designs are GPU-expensive per candidate.", "sources": [ "https://raw.githubusercontent.com/manifoldbio/mber-open/main/LICENSE", "https://raw.githubusercontent.com/manifoldbio/mber-open/main/README.md" ], "verified_at": "2026-10-15" }, { "id": "iggm", "name": "IgGM", "category": "antibody-specific design", "one_line": "Generative model for antibody design that states commercial use is allowed.", "what_it_does": "Designs antibody sequences and structures together for a given antigen. Input: antigen structure plus the framework and the CDR positions to design; output: designed CDR sequences with the predicted antibody-antigen complex. Weights download automatically on first run.", "maker": "Tencent AI4S", "url": "https://github.com/TencentAI4S/IgGM", "paper_url": "https://arxiv.org/abs/2504.09248", "license": "MIT License", "commercial_use": "yes", "commercial_use_note": "MIT LICENSE (Copyright 2024, Tencent Inc.) and an unusually direct README statement: 'Our model and code are released under MIT License, and can be freely used for both academic and commercial purposes.' That sentence covers the weights as well as the code, which is rare in this field. Note the same group's tFold is non-commercial - do not assume the licence carries across repos.", "commercial_cost": "free", "commercial_cost_note": "Free: MIT, with the README stating free use for commercial purposes and weights fetched automatically at no charge.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/TencentAI4S/IgGM/master/README.md" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "One of the few antibody design models whose weights are explicitly commercial-use-clear, so it belongs in any company pipeline as a cross-check against RFantibody and Germinal. Published validation is still largely in-silico.", "sources": [ "https://raw.githubusercontent.com/TencentAI4S/IgGM/master/LICENSE", "https://raw.githubusercontent.com/TencentAI4S/IgGM/master/README.md" ], "verified_at": "2026-10-15" }, { "id": "tfold", "name": "tFold", "category": "antibody-specific design", "one_line": "Tencent's antibody structure prediction suite - PolyForm Noncommercial licence.", "what_it_does": "Predicts antibody and nanobody structures, and antibody-antigen complexes, from sequence without MSAs. Input: heavy/light (or VHH) sequences, plus antigen sequence for the complex mode; output: predicted Fv or complex structures with confidence estimates.", "maker": "Tencent AI4S", "url": "https://github.com/TencentAI4S/tFold", "paper_url": "https://www.nature.com/articles/s42256-024-00939-y", "license": "PolyForm Noncommercial License 1.0.0", "commercial_use": "no", "commercial_use_note": "The repo LICENSE is PolyForm Noncommercial 1.0.0, which grants rights only for noncommercial purposes. Contrast with the same lab's IgGM, which is MIT and explicitly commercial-friendly - read each repo's own file.", "commercial_cost": "none", "commercial_cost_note": "No fee and no commercial route: PolyForm Noncommercial grants noncommercial rights only, and no commercial licence is offered on the repo.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/TencentAI4S/tfold/master/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Academic antibody-antigen complex prediction. In a company, use ImmuneBuilder/ABodyBuilder3 for Fv structures and Boltz-2 or Chai-1 for complexes.", "sources": [ "https://raw.githubusercontent.com/TencentAI4S/tfold/master/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "helixfold3", "name": "HelixFold3", "category": "structure and complex prediction", "one_line": "Baidu's AlphaFold3 reproduction - code non-commercial, commercial use only via the paid server.", "what_it_does": "Predicts all-atom biomolecular complexes (proteins, ligands, nucleic acids, ions). Input: a JSON entity list with optional MSAs; output: ranked all-atom complex structures with confidence scores. Built on PaddlePaddle, so it needs that stack rather than PyTorch.", "maker": "Baidu (PaddleHelix team)", "url": "https://github.com/PaddlePaddle/PaddleHelix/tree/dev/apps/protein_folding/helixfold3", "paper_url": "https://arxiv.org/abs/2408.16975", "license": "Custom 'Terms of Use Agreement' - non-commercial use only for code and parameters", "commercial_use": "restricted", "commercial_use_note": "The helixfold3 LICENSE states the open-source code and derivative works are 'only available for non-commercial use by individuals and non-commercial organizations... or for journalism', and the README repeats that code and model parameters are non-commercial only. The same README says the free version of the HelixFold3 server is non-commercial while 'the paid version offers unrestricted usage, enabling commercial applications' - so commercial use exists, but only as a paid hosted service, not as a self-hosted model.", "commercial_cost": "paid-service", "commercial_cost_note": "Self-hosting commercially is not permitted at any price. Commercial use runs through the paid tier of the hosted HelixFold3 server; the README states the free tier is non-commercial and the paid tier is unrestricted, but does not publish the price.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/PaddlePaddle/PaddleHelix/dev/apps/protein_folding/helixfold3/README.md" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Mostly of academic or comparative interest. If you need an AlphaFold3-class predictor you can self-host commercially, use Boltz-2, Chai-1 or Protenix instead; HelixFold3 would mean paying for Baidu's server.", "sources": [ "https://raw.githubusercontent.com/PaddlePaddle/PaddleHelix/dev/apps/protein_folding/helixfold3/LICENSE", "https://raw.githubusercontent.com/PaddlePaddle/PaddleHelix/dev/apps/protein_folding/helixfold3/README.md" ], "verified_at": "2026-10-15" }, { "id": "diffab", "name": "DiffAb", "category": "antibody-specific design", "one_line": "Diffusion model for antigen-specific CDR design and optimisation.", "what_it_does": "Designs or redesigns CDRs against a given antigen. Input: an antibody-antigen complex structure (or a docked framework) and the CDR to vary; output: sampled CDR sequence-structure pairs, supporting de novo design, sequence-structure co-design and optimisation of an existing antibody.", "maker": "Shitong Luo et al. (academic release)", "url": "https://github.com/luost26/diffab", "paper_url": "https://www.biorxiv.org/content/10.1101/2022.07.10.499510v5", "license": "Apache License 2.0", "commercial_use": "yes", "commercial_use_note": "Apache-2.0 LICENSE in the repo; the trained checkpoints are distributed from the same repository with no separate terms. It does depend on antibody structure tooling, which you should licence-check separately.", "commercial_cost": "free", "commercial_cost_note": "Free: Apache-2.0 code and checkpoints, no fee.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/luost26/diffab/main/LICENSE" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "A well-known, easy-to-run baseline for CDR redesign when you already have a complex structure - good for optimisation and for sanity-checking newer tools. It needs a starting complex; it will not find an epitope for you.", "sources": [ "https://raw.githubusercontent.com/luost26/diffab/main/LICENSE" ], "verified_at": "2026-10-15" }, { "id": "iglm", "name": "IgLM", "category": "protein language models", "one_line": "Antibody-specific generative language model - academic licence only.", "what_it_does": "Generates and infills antibody sequences conditioned on species and chain type. Input: a partial antibody sequence with a span to fill, plus species/chain tokens; output: completed sequences with likelihoods, used for library generation and for scoring how natural a designed CDR looks.", "maker": "Gray Lab, Johns Hopkins University", "url": "https://github.com/Graylab/IgLM", "paper_url": "https://www.cell.com/cell-systems/fulltext/S2405-4712(23)00271-5", "license": "JHU Academic Software License Agreement", "commercial_use": "no", "commercial_use_note": "LICENSE.md is the JHU Academic Software License Agreement, effective only between JHU and 'a non-profit academic or non-profit research institution'. Same pattern as IgFold from the same lab. A company needs a separate agreement with JHU.", "commercial_cost": "none", "commercial_cost_note": "No fee for academics; no published commercial price. Commercial use would require negotiating with JHU's technology transfer office.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/Graylab/IgLM/main/LICENSE.md" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Academic antibody library generation and naturalness scoring. For commercial work, AbLang2 (BSD-3) and IgBert/IgT5 (MIT) cover the same ground.", "sources": [ "https://raw.githubusercontent.com/Graylab/IgLM/main/LICENSE.md" ], "verified_at": "2026-10-15" }, { "id": "igbert-igt5", "name": "IgBert / IgT5", "category": "protein language models", "one_line": "Exscientia's paired-antibody BERT and T5 models, MIT licensed on Hugging Face.", "what_it_does": "Encode or generate antibody sequences with paired heavy+light context. Input: heavy and light chain sequences (space-separated residues, paired or unpaired); output: per-residue and sequence embeddings, masked-residue predictions (IgBert) or sequence-to-sequence generation (IgT5), for naturalness scoring, humanness work and property prediction.", "maker": "Exscientia", "url": "https://huggingface.co/Exscientia/IgBert", "paper_url": "https://arxiv.org/abs/2403.17889", "license": "MIT License (as declared on the model cards)", "commercial_use": "yes", "commercial_use_note": "Both model repositories declare the MIT licence in their card metadata on Hugging Face, covering the published weights; the models are ungated and download without an access request. Verified against the Hugging Face model API for Exscientia/IgBert and Exscientia/IgT5.", "commercial_cost": "free", "commercial_cost_note": "Free: MIT-declared weights, public download, no access request and no fee.", "commercial_cost_sources": [ "https://huggingface.co/Exscientia/IgBert", "https://huggingface.co/Exscientia/IgT5" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "A commercially clean antibody language model pair, trained with paired-chain context - useful alongside AbLang2 for scoring designs. IgT5 if you want generation, IgBert if you want embeddings.", "sources": [ "https://huggingface.co/Exscientia/IgBert", "https://huggingface.co/Exscientia/IgT5" ], "verified_at": "2026-10-15" }, { "id": "antifold", "name": "AntiFold", "category": "sequence design / inverse folding", "one_line": "Antibody-specific inverse folding, fine-tuned from ESM-IF1.", "what_it_does": "Designs antibody sequences for a given Fv backbone. Input: an antibody structure (PDB) with the regions to redesign, e.g. specific CDRs; output: per-position amino-acid probabilities and sampled sequences that keep the fold - the antibody-tuned counterpart to ProteinMPNN.", "maker": "OPIG (Oxford Protein Informatics Group) and collaborators", "url": "https://github.com/oxpig/AntiFold", "paper_url": "https://academic.oup.com/bioinformaticsadvances/article/doi/10.1093/bioadv/vbae202/7921183", "license": "BSD 3-Clause License (code and the released model weights)", "commercial_use": "yes", "commercial_use_note": "The repo LICENSE is BSD 3-Clause, and the README's model section links the weight download alongside an explicit 'License: BSD 3-Clause' line for those weights. One caveat read in the same README: an optional --esm_if1_mode runs ESM-IF1 weights instead, which carry Meta's own terms - check those if you use that flag.", "commercial_cost": "free", "commercial_cost_note": "Free: BSD 3-Clause for both code and weights, no fee.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/oxpig/AntiFold/master/README.md" ], "open_weights": true, "compute": "cpu", "when_to_use": "Use instead of plain ProteinMPNN whenever the backbone is an antibody: it respects antibody sequence statistics and is the practical tool for CDR resurfacing and affinity-maturation libraries. Runs on CPU.", "sources": [ "https://raw.githubusercontent.com/oxpig/AntiFold/master/LICENSE", "https://raw.githubusercontent.com/oxpig/AntiFold/master/README.md" ], "verified_at": "2026-10-15" }, { "id": "progen3", "name": "ProGen3", "category": "protein language models", "one_line": "Profluent's sparse protein language model family - code open, weights non-commercial.", "what_it_does": "Generates and scores protein sequences at scale. Input: a prompt (family/function tags or a partial sequence) for generation, or sequences for scoring; output: generated sequences, or log-likelihood fitness scores used to rank variants. Ships as a CLI for scoring and generation.", "maker": "Profluent Bio", "url": "https://github.com/Profluent-AI/progen3", "paper_url": "https://www.biorxiv.org/content/10.1101/2025.04.15.649055v1", "license": "Apache License 2.0 (code, LICENSE-CODE); CC BY-NC-SA 4.0 (model weights)", "commercial_use": "no", "commercial_use_note": "The README states up front that 'the model weights are released under a non-commercial usage license', and its LICENSE section splits them: Apache-2.0 for the code in LICENSE-CODE, CC BY-NC-SA 4.0 for the weights. The Hugging Face model card metadata also declares cc-by-nc-sa-4.0. The code being Apache-2.0 does not help: without the weights there is no model.", "commercial_cost": "none", "commercial_cost_note": "No fee, but also no commercial grant: CC BY-NC-SA 4.0 excludes commercial use and Profluent publishes no commercial licence price for the weights.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/Profluent-AI/progen3/main/README.md" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Academic sequence generation and variant scoring at scale. Commercially, use ESM-2/ESM C or ProtGPT2, all of which are permissive.", "sources": [ "https://raw.githubusercontent.com/Profluent-AI/progen3/main/LICENSE-CODE", "https://raw.githubusercontent.com/Profluent-AI/progen3/main/README.md", "https://huggingface.co/Profluent-Bio/progen3-3b" ], "verified_at": "2026-10-15" }, { "id": "protgpt2", "name": "ProtGPT2", "category": "protein language models", "one_line": "GPT-2-style generative model that writes de novo protein sequences.", "what_it_does": "Generates novel protein sequences unconditionally or from a seed. Input: an optional sequence prefix and sampling parameters; output: new sequences in natural-protein-like space, plus perplexity scores. A sequence-only generator: it says nothing about structure, so fold it afterwards.", "maker": "Noelia Ferruz et al.", "url": "https://huggingface.co/nferruz/ProtGPT2", "paper_url": "https://www.nature.com/articles/s41467-022-32007-7", "license": "Apache License 2.0 (as declared on the model card)", "commercial_use": "yes", "commercial_use_note": "The Hugging Face model card metadata declares apache-2.0 for the published weights, and the model is ungated. Verified against the Hugging Face model API.", "commercial_cost": "free", "commercial_cost_note": "Free: Apache-2.0 weights, public download, no fee.", "commercial_cost_sources": [ "https://huggingface.co/nferruz/ProtGPT2" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Cheap sequence-space exploration and baseline generation, especially for scaffolds or linkers. Outputs need structure prediction and filtering; it has no notion of a target.", "sources": [ "https://huggingface.co/nferruz/ProtGPT2" ], "verified_at": "2026-10-15" }, { "id": "evo2", "name": "Evo 2", "category": "protein language models", "one_line": "Genome-scale DNA/RNA/protein foundation model, Apache-2.0 including weights.", "what_it_does": "Models biological sequence at the nucleotide level across very long contexts. Input: DNA, RNA or protein sequence (up to ~1 Mb context); output: likelihoods for variant-effect prediction, embeddings, or generated sequence. Relevant to design work for variant scoring and regulatory/coding sequence generation rather than structure.", "maker": "Arc Institute, with NVIDIA and Stanford", "url": "https://github.com/ArcInstitute/evo2", "paper_url": "https://www.science.org/doi/10.1126/science.adt2044", "license": "Apache License 2.0", "commercial_use": "yes", "commercial_use_note": "Repo LICENSE is Apache-2.0 and the Hugging Face checkpoints (arcinstitute/evo2_7b and siblings) declare apache-2.0 in their card metadata, so code and weights are both permissive. The published checkpoints are ungated.", "commercial_cost": "free", "commercial_cost_note": "Free: Apache-2.0 code and weights with no fee. NVIDIA also sells hosted inference, which is optional.", "commercial_cost_sources": [ "https://huggingface.co/arcinstitute/evo2_7b" ], "open_weights": true, "compute": "heavy-gpu", "when_to_use": "Useful when the question is nucleotide-level - variant effects, expression constructs, codon and regulatory design - not when you need a 3D structure. Large models need serious GPU memory.", "sources": [ "https://raw.githubusercontent.com/ArcInstitute/evo2/main/LICENSE", "https://huggingface.co/arcinstitute/evo2_7b" ], "verified_at": "2026-10-15" }, { "id": "evobind", "name": "EvoBind", "category": "de novo binder and protein design", "one_line": "In-silico directed evolution of cyclic and linear peptide binders with AlphaFold.", "what_it_does": "Evolves a peptide binder against a target without needing a known binding site. Input: target sequence/structure and optional target residues; output: optimised peptide (including macrocyclic) sequences with AlphaFold-predicted complexes, by iterative mutation under an AlphaFold-based objective.", "maker": "Patrick Bryant (Stockholm University / SciLifeLab)", "url": "https://github.com/patrickbryant1/EvoBind", "paper_url": "https://www.nature.com/articles/s42003-025-08009-8", "license": "unverified - no LICENSE file published in the repository", "commercial_use": "unverified", "commercial_use_note": "Checked via the GitHub licence API and the repository root on the date shown: the repo publishes no LICENSE file and the README states no terms, so by default no rights are granted and the position is unverified. It also runs AlphaFold2 parameters (CC BY 4.0) underneath, which are themselves fine for commercial use.", "commercial_cost": "unverified", "commercial_cost_note": "No fee is charged for the code, but with no licence published there is no stated right to use it commercially at any price.", "commercial_cost_sources": [ "https://github.com/patrickbryant1/EvoBind" ], "open_weights": false, "compute": "heavy-gpu", "when_to_use": "Interesting for peptide and macrocycle binders where you have no structural starting point. Treat the missing licence as a blocker for commercial use until the author states terms - ask him; it is usually a one-line answer.", "sources": [ "https://github.com/patrickbryant1/EvoBind" ], "verified_at": "2026-10-15" }, { "id": "pepmlm", "name": "PepMLM", "category": "de novo binder and protein design", "one_line": "Masked-language-model generation of linear peptide binders from target sequence alone.", "what_it_does": "Generates peptide binders without any structure. Input: the target protein sequence and the desired binder length; output: candidate peptide sequences, produced by masked-token generation conditioned on the target - cheap enough to run on a laptop-class GPU.", "maker": "Chatterjee Lab (Duke / Programmable Biology)", "url": "https://github.com/programmablebio/pepmlm", "paper_url": "https://arxiv.org/abs/2310.03842", "license": "Model card declares MIT, but access is gated behind terms that require non-commercial use", "commercial_use": "no", "commercial_use_note": "A genuine contradiction, read on the date shown: the Hugging Face card for ChatterjeeLab/PepMLM-650M declares license: mit, while the same card's gated-access fields make you tick 'I agree to use this model for non-commercial use ONLY', plus authorship, data-sharing and no-patent conditions. The GitHub repo publishes no LICENSE file at all. Those access terms are what you actually accept to obtain the weights, so treat it as non-commercial until the authors reconcile the two.", "commercial_cost": "none", "commercial_cost_note": "No money changes hands, but the access agreement demands non-commercial use and rights concessions, so there is no commercial path. No commercial licence is offered.", "commercial_cost_sources": [ "https://huggingface.co/ChatterjeeLab/PepMLM-650M" ], "open_weights": true, "compute": "single-gpu", "when_to_use": "Fast sequence-only peptide binder ideas when you have no structure at all. The access conditions - no patents on generated sequences, authorship on resulting papers - are disqualifying for most company work; read them before downloading.", "sources": [ "https://huggingface.co/ChatterjeeLab/PepMLM-650M", "https://github.com/programmablebio/pepmlm" ], "verified_at": "2026-10-15" }, { "id": "alphaproteo", "name": "AlphaProteo", "category": "de novo binder and protein design", "one_line": "DeepMind's binder design system - no code or weights released; trusted-tester access only.", "what_it_does": "Designs de novo protein binders to a specified target site, with reported nanomolar-to-picomolar affinities straight out of the model. Input (per the trusted-tester notebook): a target PDB ID and the target region; output: binder designs returned by DeepMind, not computed locally. There is no model you can run.", "maker": "Google DeepMind", "url": "https://github.com/google-deepmind/ap_trusted_testers", "paper_url": "https://www.nature.com/articles/s41586-025-09721-5", "license": "Apache License 2.0 covers only the input-preparation notebooks in the public repo; the model itself is unreleased", "commercial_use": "unverified", "commercial_use_note": "The only public artefact is the ap_trusted_testers repository, whose README describes a Colab that helps 'users of the AlphaProteo Trusted Tester Program' prototype requests and visualise target regions - an input builder, not the model. No weights, no inference code and no terms for commercial use of the designs are published, so rights are unverified.", "commercial_cost": "unverified", "commercial_cost_note": "No price exists because no product exists. Access is by DeepMind's trusted-tester programme, whose commercial terms are not published.", "commercial_cost_sources": [ "https://raw.githubusercontent.com/google-deepmind/ap_trusted_testers/master/README.md" ], "open_weights": false, "compute": "hosted-service", "when_to_use": "Nothing to adopt today: include it in a landscape review, not a pipeline. BindCraft, RFdiffusion and BoltzGen are what you can actually run for binder design.", "sources": [ "https://raw.githubusercontent.com/google-deepmind/ap_trusted_testers/master/README.md", "https://github.com/google-deepmind/ap_trusted_testers" ], "verified_at": "2026-10-15" } ] }