AI drug discovery pipeline: from a target protein to ranked compounds
FoldRx folds your target, finds its pockets, docks a compound library and weighs docking against ADMET, pose checks and off-target selectivity in one ranking. Build it node by node in the AI workspace, choose the model for every stage, and see exactly what each one found. No code, no cluster.
Runs are on the Pro plan: 10 a month from $24, or 50 on Max. Finished demo screens are open to everyone. See pricing
The stages, and the models that can fill them
- 1
Target
Built-in check
Paste a protein, or DNA or mRNA, whose longest open reading frame is translated for you. Or start from a saved sequence in your library.
- 2
Fold
ESMFold (default) · ESMFold2 · Boltz-2 · a structure you bring · your own model
Predict the 3D structure with per-residue confidence. Bring a structure from AlphaFold Server or the PDB instead, or plug in your own model endpoint, such as AlphaFold 3 run under your licence.
- 3
Pockets
P2Rank (default) · fpocket
Find druggable cavities on the structure, ranked, with centres ready to dock into.
- 4
Compounds
Approved drugs · fragments · natural products · your SMILES
Choose a curated library, paste your own compounds, or screen the analogs the previous round designed.
- 5
Dock
AutoDock Vina (default) · Uni-Dock (GPU)
Dock every compound into the chosen pocket and keep the best poses, with scores in kcal/mol.
- 6
ADMET
ADMET-AI · RDKit rules
Predicted absorption, distribution, metabolism, excretion and toxicity, plus drug-likeness rules. Predictions, labelled as such.
- 7
Pose checks
PoseBusters · Boltz-2 affinity
Reject physically implausible poses, and rescore the shortlist with Boltz-2's binding affinity head.
- 8
Selectivity
Off-target panel
Dock the shortlist against common anti-targets and flag compounds that bind them as well as your target.
- 9
Rank
Weighted rank
Weigh docking, ADMET, pose checks and selectivity into one ordered table. Export it as CSV.
- 10
Design
CReM analogs
Propose analogs of this round's best hits, then run them as the next round.
Bring your own model, including AlphaFold 3
Any stage that runs a model can run yours instead. Register an endpoint that accepts the stage's input and answers in its shape, and it appears in that stage's model list, for your account only. That is how a lab with an AlphaFold 3 licence folds its screens with AlphaFold 3, and how a team uses a fine-tuned in-house model.
The model contract and the AlphaFold 3 adapter- Talindrew sends your endpoint exactly what the built-in model receives.
- Long jobs answer 202 and are polled, the protocol NVIDIA NIMs use, so a NIM registers as it is.
- Your token is stored encrypted and never shown again; private network addresses are refused.
- A test button folds a 35-residue protein through your endpoint before your first run.
See it in 0:47
GFP folded, pLDDT colours, notebook, FoldRx.
FoldRx questions
Not on Talindrew's servers: AlphaFold 3's model parameters are licensed by Google DeepMind for non-commercial use, so no service may run them for you commercially. If you hold that licence, run AlphaFold 3 on your own hardware behind Talindrew's open-source adapter and add it as your own model; it then folds your FoldRx runs like any built-in model. AlphaFold Server structures can also be uploaded as the structure you bring.
ESMFold (fast, single sequence), ESMFold2 (Biohub's 2026 successor, single-sequence Fast variant) and Boltz-2 (co-folding with a ligand, with binding affinity) run on Talindrew's GPUs. You can also bring a structure file, or register a model endpoint you run, such as AlphaFold 3, ESMFold2 or a fine-tuned model, as your own model for the fold stage.
A few minutes to about ten for a typical target and a few hundred compounds. Each stage shows its progress, and a stage you have not changed is reused when you start the next round.
Talindrew's cloning suite and notebook are free. FoldRx runs use GPUs, so they are on the Pro plan (10 runs a month from $24) and Max (50 runs a month). Finished demo screens, including the one in the film, are open to everyone.
Docking scores, ADMET predictions and affinities rank candidates; none of them is a measurement. Results are for research and do not constitute medical advice.
Open a finished screen, then run your own target
Demo screens are free to explore. Runs are on Pro and Max.