Anthropic just launched an AI drug discovery project that puts neglected diseases first, and the framing sharpens a question researchers have been asking for a year: which AI tools are actually usable for early-stage drug discovery, not just paper demos. The category has grown fast. AlphaFold moved from an academic marvel to a working tool. Open-source alternatives like RoseTTAFold caught up. New molecular-design libraries lower the bar for a chemist to run virtual screens on a workstation. We tested seven AI drug discovery apps for desktop that cover the different jobs: structure prediction, molecular design, screening, and workflow.
Every option here runs on Windows, macOS, or Linux — either as a locally installed library or as a browser-based workbench.
What to look for in an AI drug discovery tool
The pipeline breaks into distinct phases and different tools shine in each:
- Structure prediction. Predict a protein’s 3D fold from its sequence. AlphaFold 3, RoseTTAFold All-Atom, Chai-1, Boltz-1, and OpenFold.
- Molecular design. Propose small molecules for a target, filter by drug-like properties. DeepChem.
- Virtual screening / docking. Match candidate ligands against a receptor structure. AutoDock Vina and its GPU forks.
- Property prediction. ADMET (absorption, distribution, metabolism, excretion, toxicity) modelling. DeepChem, plus commercial add-ons.
The right stack usually chains three: predict the target’s structure, screen a compound library against it, and score the hits for drug-like properties. Neglected disease projects lean heavily on open source for cost reasons.
Quick comparison
| Tool | Best for | Free plan | Setup | Standout |
|---|---|---|---|---|
| AlphaFold 3 Server | Structure prediction, ligand complexes | Yes, quota | Web | 20 jobs/day, complexes with small molecules and ions |
| RoseTTAFold All-Atom | Open-source structure prediction | Yes | Self-host | Full atom-level modelling of ligand-protein complexes |
| DeepChem | Full ML pipeline for chemistry | Yes | Python library | Datasets, models, and tutorials in one package |
| Chai-1 | Open-weights AlphaFold 3 successor | Yes | Self-host | Antibody-antigen modelling, multi-chain complexes |
| Boltz-1 | MIT open-weights structure model | Yes | Self-host | AlphaFold 3-tier accuracy, permissive licence |
| OpenFold | AlphaFold reimplemented in PyTorch | Yes | Self-host | Trainable from scratch, extensible |
| AutoDock Vina | Molecular docking / virtual screening | Yes | Self-host | 20+ years of docking, GPU forks (Vina-GPU) available |
The 7 AI drug discovery apps we tested
1. AlphaFold 3 Server — best hosted structure prediction
AlphaFold 3 Server by Google DeepMind and Isomorphic Labs is the fastest way to get a state-of-the-art structure prediction without setting up any local compute. It predicts protein-only structures, protein-nucleic acid complexes, and protein-ligand complexes for small molecules and ions. The web UI accepts sequences and small-molecule SMILES; results include the PAE plots and pLDDT confidence colouring.
Where it falls short: Non-commercial use only via the server. Job quota (~20/day) is capped. Commercial teams need to license through Isomorphic Labs.
Pricing: Free for non-commercial research. Commercial licensing on request.
Platforms: Web (Windows, macOS, Linux).
Download: alphafoldserver.com
Bottom line: The pick when you want AlphaFold 3 without owning a GPU. Non-commercial only; layer Chai-1 or Boltz-1 if you need to ship a product.
2. RoseTTAFold All-Atom — best open-source structure prediction
RoseTTAFold All-Atom from the Baker Lab extends the RoseTTAFold family to model everything in the biological system at atomic resolution: proteins, nucleic acids, small molecules, ions, and covalent modifications. Open source and permissively licensed, it’s the pick for commercial teams who need a locally deployable state-of-the-art predictor.
Where it falls short: GPU requirements are non-trivial (24GB+ recommended). Setup is heavier than the hosted alternatives.
Pricing: Free, open-source.
Platforms: Windows (via WSL), macOS (limited), Linux.
Download: github.com/baker-laboratory/RoseTTAFold-All-Atom
Bottom line: The pick for commercial and industrial drug discovery pipelines that need a self-hosted state-of-the-art model.
3. DeepChem — best full pipeline library
DeepChem is the open-source Python library for democratising deep learning in chemistry, materials science, and biology. It bundles datasets, model implementations, tutorials, and pre-trained checkpoints so a chemistry team can go from “we have a target” to “we have a screening pipeline” in a week. Community-supported and actively maintained.
Where it falls short: Broad rather than deep in any one area. Some models trail the state of the art from single-purpose libraries.
Pricing: Free, open-source.
Platforms: Windows, macOS, Linux (Python).
Download: deepchem.io · GitHub
Bottom line: The pick when you want a full pipeline in one library. Best entry point for a team new to ML for drug discovery.
4. Chai-1 — best open-weights AlphaFold 3 successor
Chai-1 from Chai Discovery released with open weights and delivered AlphaFold 3-competitive performance on the CASP benchmarks. Strong on antibody-antigen and multi-chain complexes, which matters for biologics work. Runs locally on a workstation GPU.
Where it falls short: Newer than the RoseTTAFold family; community integrations still catching up. Documentation is thinner.
Pricing: Free, open-weights (see licence).
Platforms: Windows (via WSL), Linux; macOS via CPU or MPS.
Download: chaidiscovery.com · GitHub
Bottom line: The pick when you need AlphaFold 3-tier accuracy locally and permissive licensing. Best for antibody design.
5. Boltz-1 — MIT's open-weights structure model
Boltz-1 from MIT’s Jameel Clinic released as an open-weights AlphaFold 3-tier structure predictor with a permissive licence. Strong on protein-ligand complexes and easy to fine-tune, which lets research groups adapt the model to their target class.
Where it falls short: Youngest of the open-weights options. Community best practices are still consolidating.
Pricing: Free, open-weights (MIT licence).
Platforms: Linux; Windows via WSL. macOS via CPU.
Download: github.com/jwohlwend/boltz
Bottom line: The pick when you want the most permissively licensed AlphaFold 3-tier model and plan to fine-tune.
6. OpenFold — AlphaFold in PyTorch
OpenFold by the OpenFold Consortium reimplements AlphaFold in PyTorch, retrains it on public data, and releases everything: weights, training code, dataset preparation. If you want to train your own version, adapt the architecture, or study the model, OpenFold is where you start.
Where it falls short: Not as accurate as AlphaFold 3-tier successors on the newest benchmarks. Training from scratch requires substantial compute.
Pricing: Free, open-source.
Platforms: Linux; Windows via WSL.
Download: github.com/aqlaboratory/openfold
Bottom line: The pick for research groups that want to extend the model, not just consume its predictions.
7. AutoDock Vina — best molecular docking tool
AutoDock Vina has been the workhorse of virtual screening for two decades, and the ecosystem around it (PyRx for batch screens, Vina-GPU for accelerated runs, custom scoring functions) keeps it competitive with commercial tools. Fast, well-documented, and the reference tool many published pipelines still cite.
Where it falls short: Scoring function is dated relative to modern ML-based scorers. Setup for large screens takes effort.
Pricing: Free, open-source.
Platforms: Windows, macOS, Linux.
Download: vina.scripps.edu · GitHub
Bottom line: The pick for virtual screening. Pair with a structure predictor above for the full pipeline.
How to pick the right one
- If you need one structure prediction without owning a GPU: AlphaFold 3 Server.
- If you need commercial-safe self-hosted structure prediction: RoseTTAFold All-Atom, Chai-1, or Boltz-1.
- If you want a full open-source pipeline in one library: DeepChem.
- If you plan to extend or train the model yourself: OpenFold.
- If you need virtual screening today: AutoDock Vina, ideally the GPU fork.
For a neglected-disease project (following Anthropic’s positioning), the open-source stack — RoseTTAFold or Chai-1 for structure, DeepChem for the ML pipeline, AutoDock Vina for screening — keeps costs low and provenance clear.
FAQ
What is the best free AI drug discovery tool? AlphaFold 3 Server for hosted structure prediction, DeepChem for the full pipeline in one library, AutoDock Vina for virtual screening. All three are free.
Can I use AlphaFold 3 commercially? Not through the free server. Commercial use goes through Isomorphic Labs licensing. The open-weights alternatives (RoseTTAFold All-Atom, Chai-1, Boltz-1) have permissive licences.
Do these run offline? Yes for all seven except AlphaFold 3 Server. Local runs require GPU compute; a 24GB VRAM workstation covers most workflows.
What hardware do I need? For structure prediction locally: 24GB+ GPU VRAM recommended for the AlphaFold 3-tier models. DeepChem and AutoDock Vina run on modest hardware, though GPU acceleration helps large screens.
Does Anthropic’s Claude help with drug discovery workflows? Claude Science, launched separately from the neglected-disease programme, is a workbench that can drive many of these tools programmatically. It doesn’t replace the underlying models; it orchestrates them.
