Best Open-Source Computational Drug Discovery Tools in 2026

The open-source stack for computational drug discovery is no longer a collection of research code. In 2026 you can fold a protein, co-fold a complex, dock a ligand, screen it for pharmacokinetics, generate a better analogue, plan its synthesis and read the single-cell consequences — with tools that each carry a peer-reviewed paper and a permissive licence. This is the 2026 edition of our 2025 survey, re-checked tool by tool: every licence below was read from the project's own repository and every release date from its own release feed, on 2026-10-24.

One thing has not changed. The hard part is not finding the tools, it is chaining them: different input formats, different GPU requirements, different output schemas. The last two sections deal with that.

Best open-source tools for protein structure prediction in 2026#

For a single chain from sequence alone, ESMFold is still the fastest credible answer and needs no multiple sequence alignment; ColabFold is the practical way to run AlphaFold2 when you want MSA-grade accuracy; and for a complex — several chains, a ligand, a nucleic acid — Boltz-2 is the one to reach for first, because it is the strongest co-folding model whose weights you can simply download. AlphaFold 3 is more capable than any of them on interactions, but its weights are not openly licensed.

What the numbers actually say. ESMFold's contribution was scale: its authors used it to predict structures for over 617 million metagenomic sequences, more than 225 million of them at high confidence. ColabFold's was throughput — a 40–60-fold faster homology search, enough for close to 1,000 structures a day on a single GPU. Boltz-2's is affinity: it is, by its authors' account, the first model to approach free-energy perturbation accuracy on small-molecule binding affinity while being at least 1,000× cheaper to run. Chai-1 can be prompted with wet-lab restraints, which its preprint reports lifting performance by double-digit percentages, and it runs in single-sequence mode without MSAs.

Two licence details are worth knowing before you plan around them. The original facebookresearch/esm repository that shipped ESMFold is archived, with its last commit in February 2024; the ESM line continues in a separate repository under an MIT licence, where the current generation of sequence and folding models is documented in a 2026 preprint. And AlphaFold 3's code being Apache-2.0 does not make the model open: the parameters come only from Google and carry their own terms of use.

For the head-to-head on the two single-chain methods, see ESMFold vs AlphaFold2; for co-folding, DiffDock vs Chai-1 vs Boltz-2.

Top open-source molecular docking software in 2026#

Dock in two steps, not one. P2Rank finds and ranks the likely ligand-binding pockets from the structure alone, then DiffDock poses the ligand without needing you to define a box, or AutoDock Vina poses it inside one you do define. DiffDock wins on finding the right pose; Vina wins on returning a pose that is physically valid; using both and comparing is cheaper than arguing about which is better.

DiffDock reports 38% top-1 success at RMSD below 2 Å on PDBBind, against 23% for traditional docking and 20% for the best deep-learning baseline in the same comparison, and it holds up better on computationally folded receptors — 21.7% against a maximum of 10.4% for the methods it compares with. PoseBusters is the necessary counterweight: scored on stereochemistry, bond geometry, ring planarity and protein-ligand clashes, as well as on generalisation to sequences unlike the training data, no deep-learning docking method in that study outperformed the classical tools, and a molecular-mechanics minimisation step closed part of the gap. Treat the two findings as complementary, not contradictory — and run a validity check on any pose you intend to act on.

More detail in DiffDock vs GNINA and P2Rank vs fpocket.

Top ADMET prediction software for drug candidates in 2026#

ADMET-AI is the open-source default: a Python package and web server that predicts a broad panel of absorption, distribution, metabolism, excretion and toxicity endpoints, and reports the highest average rank on the Therapeutics Data Commons ADMET leaderboard in its paper. The free web servers SwissADME and pkCSM are useful orthogonal second opinions, but they are hosted services, not code you can audit or batch.

ADMET-AI's throughput is the reason it suits a screening cascade: its paper reports one million molecules predicted in 3.1 hours locally. Its second major version, released in February 2026, re-trained the models on Chemprop v2 and added PAINS, BRENK and NIH structural alerts; the release notes are explicit that v2 predictions are not identical to v1's, so do not mix outputs from the two inside one ranking.

A caveat that applies to this whole section: every model here is trained on public assay data of uneven provenance, and an endpoint with few measured compounds will be confidently wrong. We wrote up the gap between predicted and measured properties in in-silico ADMET vs in vivo, and compared the three tools directly in ADMET-AI vs SwissADME vs pkCSM.

Best open-source generative chemistry tools in 2026#

Two mature choices with different control surfaces. GenMol generates and edits molecules as fragments using masked discrete diffusion over SAFE strings, which makes fragment-constrained design and scaffold work natural. REINVENT 4 steers a sequence generator with reinforcement learning against a scoring function you define, which makes multi-objective optimisation natural. Pick by whether your constraint is structural (GenMol) or numerical (REINVENT).

GenMol's paper reports state-of-the-art results on goal-directed hit generation and lead optimisation and better performance than the previous GPT-based baseline on de novo and fragment-constrained generation. REINVENT 4 covers more design tasks in one tool — R-group replacement, linker design, scaffold hopping, library design — and is maintained as a reference implementation rather than a single model. Note the licence asymmetry: REINVENT 4 is Apache-2.0 throughout, while GenMol's code is Apache-2.0 but its weights sit under NVIDIA's model licence, which matters if you plan to redistribute.

Side by side: GenMol vs REINVENT 4.

Best machine learning protein design and protein engineering tools in 2026#

For sequence design on a fixed backbone, ProteinMPNN remains the baseline every other method is measured against, and LigandMPNN is its successor for sites where a ligand, metal or nucleic acid is part of the context. For a new backbone, RFdiffusion generates one. For a binder against a named target,

BindCraft is the first open pipeline to report an experimental success rate you can plan a campaign around. For stability triage on point mutations,

ThermoMPNN. For antibodies specifically, ImmuneBuilder for structure and AntiFold for inverse folding.

The two numbers to carry from this section. ProteinMPNN recovers 52.4% of native sequence on native backbones against 32.9% for Rosetta, which is why it displaced physics-based design as the default. BindCraft reports experimental success rates of 10–100% across targets including cell-surface receptors, allergens and multi-domain nucleases, producing nanomolar binders without high-throughput screening and without a known binding site — a different claim in kind from an in-silico recovery rate, and the reason binder design is now a planning option rather than a research project.

Order matters more than tool choice here: generate or fix a backbone, design sequences onto it, then predict the structure of what you designed and check that it folds back to what you asked for. On stability specifically, ThermoMPNN vs FoldX vs ESM zero-shot explains why only two of those three answer in kcal/mol; on sequence design, ProteinMPNN recovery vs experimental success explains why recovery is not a hit rate.

Best open-source retrosynthesis software in 2026#

AiZynthFinder is the most widely used open retrosynthesis planner and the easiest to run against your own stock of starting materials. ASKCOS is the larger suite — route planning plus condition recommendation, impurity and selectivity prediction. Syntheseus is not a planner but a benchmarking library, and it is the one to read first if you intend to compare methods.

AiZynthFinder 4.0 added filter policies, support for arbitrary one-step models, a scoring framework and several search algorithms, and its authors report a large-scale analysis over hundreds of thousands of targets covering route shape, stock usage and reaction-space coverage. The Syntheseus result is the one that should change how you read retrosynthesis claims: re-evaluated under controlled conditions, the ranking of published state-of-the-art methods changed. A route score from any single planner is a hypothesis about synthesisability, not a verdict — and a chemist's read is still the arbiter.

Best open-source single-cell analysis tools in 2026#

Scanpy is the Python standard for the whole scRNA-seq path — quality control, normalisation, dimensionality reduction, clustering, differential expression — and scvi-tools is what you add when you need probabilistic models for batch correction, imputation and multi-modal integration. Seurat is the equally mature R alternative, and your choice between them is mostly your lab's language. CellTypist handles automated cell-type annotation.

All four are under active development — Scanpy 1.12.4, scvi-tools 1.5.1, Seurat 5.2.1 and CellTypist 1.7.1 as of this writing — and the two Python packages share the AnnData object, so moving between them costs nothing. The divergence that bites is in defaults rather than capability, which is the subject of Scanpy vs Seurat.

AI drug discovery tools with no coding required in 2026#

You do not need to write Python to run any of the predictions above. There are three ways in, and they trade off differently.

The AlphaFold Database deserves its own line: if your protein is a known UniProt entry, the structure has probably already been computed, and fetching it takes seconds against the minutes or hours a prediction would. Check there before you fold anything.

How to run these on .bloom*#

On .bloom*, you ask Ivy in plain words and it plans the run, picks the tools, executes them and writes up the result with the parameters it used. A typical small-molecule chain is fold or fetch the structure, find the pocket, dock, then profile the hit:

1. Structure — esmfold_predict folds a sequence with no MSA, or alphafold_fetch pulls the pre-computed model from the AlphaFold Database when a UniProt ID exists. For a multi-chain complex or a ligand co-fold, boltz2_predict takes up to 20 typed chain entries, signed in.
2. Pocket — p2rank_predict ranks candidate sites from the structure alone, no ligand needed.
3. Dock — diffdock_dock docks blind against the receptor and returns ranked poses.
4. Profile — admet_ai_predict screens the ligand for pharmacokinetic and toxicity endpoints, and rdkit_full_analysis gives you descriptors, fingerprints and conformers.
5. Design or generate — genmol_generate proposes fragment-aware analogues; proteinmpnn_design designs sequences onto a backbone; thermompnn_predict scores point mutations for folding stability; esm_scan_predict ranks a whole sequence by language-model likelihood; immunebuilder_predict and antifold_redesign cover antibodies.
6. Synthesis — aizynthfinder_retro plans routes back to purchasable starting materials.

The fold → dock triage recipe chains the first four steps, and hit-to-lead carries ranked poses into property filtering. Which tool? maps a goal to a job when you are not sure what to ask for. ESMFold, DiffDock, P2Rank, ADMET-AI, RDKit, AiZynthFinder, ProteinMPNN and ThermoMPNN run without an account; Boltz-2 and Chai-1 need you signed in. Reach the same tools from the web app, from curl, from the ivybloom terminal client, or from your own agent over MCP — and GNINA, AutoDock Vina, Seurat, RFdiffusion, BindCraft, ASKCOS and the ADMET web servers are not offered on .bloom*, so bring their output in if you want it compared.

Key takeaways#

• Structure prediction is solved for the easy cases and open for the hard ones. ESMFold and ColabFold cover single chains; Boltz-2 and Chai-1 publish complex-prediction weights under MIT and Apache-2.0 while AlphaFold 3's stay behind a request process.
• Dock in two steps and check the pose. P2Rank then DiffDock or Vina, then a PoseBusters-style validity check — because the method that finds the pose best is not the one that returns the most physical one.
• Protein design now has experimental numbers, not just recovery rates. ProteinMPNN's 52.4% sequence recovery made it the default; BindCraft's 10–100% experimental success rate is what moved binder design into planning.
• Retrosynthesis claims need a benchmark you trust. Syntheseus showed the published ranking changes under controlled evaluation.
• Licences are not uniform, so read them before you build. Code and weights can carry different terms — AlphaFold 3 and GenMol both do.
• The bottleneck is integration. Every tool here is free; chaining them, converting formats and keeping a record of what ran is the work that platforms exist to absorb.

Run the chain — fold, pocket, dock, profile — on the .bloom* platform.

References#

1. Lin, Z. et al. "Evolutionary-scale prediction of atomic-level protein structure with a language model." Science 379, 1123–1130 (2023). doi:10.1126/science.ade2574
2. Jumper, J. et al. "Highly accurate protein structure prediction with AlphaFold." Nature 596, 583–589 (2021). doi:10.1038/s41586-021-03819-2
3. Mirdita, M. et al. "ColabFold: making protein folding accessible to all." Nature Methods 19, 679–682 (2022). doi:10.1038/s41592-022-01488-1
4. Abramson, J. et al. "Accurate structure prediction of biomolecular interactions with AlphaFold 3." Nature 630, 493–500 (2024). doi:10.1038/s41586-024-07487-w
5. Passaro, S. et al. "Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction." bioRxiv (2025). doi:10.1101/2025.06.14.659707
6. Chai Discovery. "Chai-1: Decoding the molecular interactions of life." bioRxiv (2024). doi:10.1101/2024.10.10.615955
7. Corso, G. et al. "DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking." ICLR (2023). arXiv:2210.01776
8. Eberhardt, J. et al. "AutoDock Vina 1.2.0: New Docking Methods, Expanded Force Field, and Python Bindings." Journal of Chemical Information and Modeling 61, 3891–3898 (2021). doi:10.1021/acs.jcim.1c00203
9. McNutt, A.T. et al. "GNINA 1.0: molecular docking with deep learning." Journal of Cheminformatics 13, 43 (2021). doi:10.1186/s13321-021-00522-2
10. Buttenschoen, M. et al. "PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences." Chemical Science 15, 3130–3139 (2024). doi:10.1039/D3SC04185A
11. Krivák, R. & Hoksza, D. "P2Rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure." Journal of Cheminformatics 10, 39 (2018). doi:10.1186/s13321-018-0285-8
12. Swanson, K. et al. "ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries." Bioinformatics 40, btae416 (2024). doi:10.1093/bioinformatics/btae416
13. Daina, A. et al. "SwissADME: a free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules." Scientific Reports 7, 42717 (2017). doi:10.1038/srep42717
14. Pires, D.E.V. et al. "pkCSM: Predicting Small-Molecule Pharmacokinetic and Toxicity Properties Using Graph-Based Signatures." Journal of Medicinal Chemistry 58, 4066–4072 (2015). doi:10.1021/acs.jmedchem.5b00104
15. Fu, L. et al. "ADMETlab 3.0: an updated comprehensive online ADMET prediction platform." Nucleic Acids Research 52, W422–W431 (2024). doi:10.1093/nar/gkae236
16. Lee, S. et al. "GenMol: A Drug Discovery Generalist with Discrete Diffusion." arXiv (2025). arXiv:2501.06158
17. Noutahi, E. et al. "Gotta be SAFE: a new framework for molecular design." Digital Discovery 3, 796–804 (2024). doi:10.1039/D4DD00019F
18. Loeffler, H.H. et al. "Reinvent 4: Modern AI-driven generative molecule design." Journal of Cheminformatics 16, 20 (2024). doi:10.1186/s13321-024-00812-5
19. Dauparas, J. et al. "Robust deep learning-based protein sequence design using ProteinMPNN." Science 378, 49–56 (2022). doi:10.1126/science.add2187
20. Dauparas, J. et al. "Atomic context-conditioned protein sequence design using LigandMPNN." Nature Methods 22, 717–723 (2025). doi:10.1038/s41592-025-02626-1
21. Watson, J.L. et al. "De novo design of protein structure and function with RFdiffusion." Nature 620, 1089–1100 (2023). doi:10.1038/s41586-023-06415-8
22. Pacesa, M. et al. "One-shot design of functional protein binders with BindCraft." Nature 646, 483–492 (2025). doi:10.1038/s41586-025-09429-6
23. Dieckhaus, H. et al. "Transfer learning to leverage larger datasets for improved prediction of protein stability changes." PNAS 121, e2314853121 (2024). doi:10.1073/pnas.2314853121
24. Abanades, B. et al. "ImmuneBuilder: Deep-Learning models for predicting the structures of immune proteins." Communications Biology 6, 575 (2023). doi:10.1038/s42003-023-04927-7
25. Høie, M.H. et al. "AntiFold: improved structure-based antibody design using inverse folding." Bioinformatics Advances 5, vbae202 (2024). doi:10.1093/bioadv/vbae202
26. Genheden, S. et al. "AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning." Journal of Cheminformatics 12, 70 (2020). doi:10.1186/s13321-020-00472-1
27. Saigiridharan, L. et al. "AiZynthFinder 4.0: developments based on learnings from 3 years of industrial application." Journal of Cheminformatics 16, 57 (2024). doi:10.1186/s13321-024-00860-x
28. Tu, Z. et al. "ASKCOS: Open-Source, Data-Driven Synthesis Planning." Accounts of Chemical Research 58, 1764–1775 (2025). doi:10.1021/acs.accounts.5c00155
29. Maziarz, K. et al. "Re-evaluating retrosynthesis algorithms with Syntheseus." Faraday Discussions 256, 568–586 (2025). doi:10.1039/D4FD00093E
30. Wolf, F.A. et al. "SCANPY: large-scale single-cell gene expression data analysis." Genome Biology 19, 15 (2018). doi:10.1186/s13059-017-1382-0
31. Gayoso, A. et al. "A Python library for probabilistic analysis of single-cell omics data." Nature Biotechnology 40, 163–166 (2022). doi:10.1038/s41587-021-01206-w
32. Hao, Y. et al. "Dictionary learning for integrative, multimodal and scalable single-cell analysis." Nature Biotechnology 42, 293–304 (2024). doi:10.1038/s41587-023-01767-y
33. Domínguez Conde, C. et al. "Cross-tissue immune cell analysis reveals tissue-specific features in humans." Science 376, eabl5197 (2022). doi:10.1126/science.abl5197
34. The Galaxy Community. "The Galaxy platform for accessible, reproducible, and collaborative data analyses: 2024 update." Nucleic Acids Research 52, W83–W94 (2024). doi:10.1093/nar/gkae410