F Foldworks
Foldworks · v0.1 preview

Design proteins that bind to any target
In an afternoon.

Most drugs and lab tools work by latching onto one target protein to block it, tag it, or switch it off. Foldworks designs those binders for you. Point it at a target and let AI build new proteins, antibodies, or small molecules that stick to it, then rank the ones most likely to work. Hours, not months.

See what it does
~15 min per design batch·auto binding-site finder·PDB / SDF / CSV files
Demo demo / demo · 5 design jobs/day · no signup
Cartoon rendering of SARS-CoV-2 main protease (PDB 6LU7) in complex with an inhibitor
PDB · 6LU7 · MOL*
6LU7· SARS-CoV-2 Mpro· 2.16 Å X-rayRCSB ↗
Built on
AlphaFoldRFdiffusionRFantibodyProteinMPNNAbMPNNBindCraftmberESMFoldBoltz-2Chai-1AutoDock VinaGNINADiffDock-LAiZynthFinderREINVENT 4ADMET-AIPoseBustersThermoMPNNPROSSDeepFRIPDBFixerOpenMMPyRosettaMeekoRDKitMol*UniProtRCSB PDBGTExOpen TargetsHuman Protein AtlasAlphaFoldRFdiffusionRFantibodyProteinMPNNAbMPNNBindCraftmberESMFoldBoltz-2Chai-1AutoDock VinaGNINADiffDock-LAiZynthFinderREINVENT 4ADMET-AIPoseBustersThermoMPNNPROSSDeepFRIPDBFixerOpenMMPyRosettaMeekoRDKitMol*UniProtRCSB PDBGTExOpen TargetsHuman Protein Atlas

Pipelines

Built around
binder design.

Designing a new protein binder is the main job. But antibodies, small-molecule drugs, and fluorescent tags all work the same way, grabbing onto a spot on your target, so they run on the same platform, against the same target and the same site. Switch between them without starting over, and jobs run in the background with live progress.

★ Main pipeline

Protein binder design

Invent a new protein from scratch, design its sequence, then check it folds and binds.

Choose the spot on your target you want to hit, or let Foldworks find it. AI invents brand-new mini-protein shapes that fit there; about 16% of lab-tested designs actually bind (RFdiffusion, Watson 2023). A second model picks an amino-acid sequence for each shape, and ~52% express cleanly in E. coli vs. ~9% for older Rosetta designs (ProteinMPNN, Dauparas 2022). A third folds the complex to confirm they stick (ESMFold). BindCraft is an alternate route with ~10× the hit rate on short helical binders (Pacesa 2024).

STEP 01
RFdiffusion
Invent the shape
~16% bind (Watson 2023)
STEP 02
ProteinMPNN
Pick the sequence
~52% E. coli expression
STEP 03
ESMFold
Confirm it binds
interface pLDDT · ipTM
STEP 04
Rank
Best designs first
~15 min on a 4090
RFdiffusion · BindCraft · ProteinMPNN · ESMFoldWatson 2023 ↗Pacesa 2024 ↗Dauparas 2022 ↗
Also supports

Antibody + nanobody design

Design the grabbing loops of an antibody, or a tiny single-piece nanobody from scratch.

Antibodies grip their target with a handful of floppy loops, and here AI redesigns those loops on a standard antibody frame; about 17% bind in the lab (RFantibody + AbMPNN, Bennett 2024). Or design a nanobody, a single small antibody-like protein, from scratch (mber, checked with NanoBodyBuilder2). Experimental: no drug-readiness scoring yet.

RFantibodyAbMPNNmber
Bennett 2024 ↗Manifold Bio 2025 ↗
Also supports

Small-molecule drug screening

Test thousands of drug-like compounds for fit, then flag the ones that look safe and drug-worthy.

Foldworks fits a curated library of drug-like molecules into your target's pocket to see which sit there best, a step called docking (AutoDock Vina). It re-scores the poses with a neural network (GNINA), then double-checks the top 5 for binding strength (Boltz-2). Every hit comes with predicted absorption, toxicity, and other drug-likeness scores (ADMET-AI, 10 measures), plus standard medicinal-chemistry flags (RDKit: Ro5, Veber, PAINS).

AutoDock VinaGNINABoltz-2ADMET-AI
Trott 2010 ↗McNutt 2021 ↗Swanson 2024 ↗
Also supports

Sequence redesign + fluorescent probes

Improve an existing protein's sequence, or find a glowing tag for imaging.

Keep a protein's shape but redesign its amino-acid sequence, which helps when it won't dissolve or express. ProteinMPNN recovers ~52% of the native sequence and expresses 6× better than older Rosetta designs, and a soluble-trained variant rescues ones that won't fold. The same docking engine also screens fluorescent dyes for imaging and reports their excitation/emission colors (Ex/Em).

ProteinMPNNsoluble-MPNNFluorophore lib
Dauparas 2022 ↗
The foundation

Finding the binding site

The spot a drug can grip, where two proteins touch, or where an antibody lands, matched to what you're designing.

Before designing, Foldworks finds the best spot to aim at. fpocket finds ~80% of known drug pockets; P2Rank's top-3 picks cover ~90% (COACH420). DoGSite3 scores how druggable a pocket is, FreeSASA finds where two proteins touch, and DiscoTope-3 predicts where an antibody would land. If you don't pick a spot yourself, the software chooses one.

fpocketP2RankDoGSite3DiscoTope-3

Built for chemists

Make every chemist a
computational chemist.

These design tools normally take a specialist to set up and string together by hand. Foldworks turns them into one-click workflows with sensible defaults, documented methods, and plain-file output. If you can read a binding curve, you can design a binder.

01

One-click design runs.

Pick what you want: a protein binder, an antibody, a small-molecule screen, or a probe. Foldworks handles all the fiddly setup and runs it on the right hardware. No config files, no command line, no software licences to wrangle.

02

Defaults you can trust.

Every workflow uses settings taken straight from the published papers: Watson 2023, Dauparas 2022, Trott 2010. They work out of the box for the common cases, and every method is documented so you can cite it.

03

Plain files, no lock-in.

Results come out as standard PDB / SDF / CSV files, with a full record of how they were made. Open them in ChemDraw, Excel, or your lab's own tools. Nothing is locked in a proprietary format, and you can re-run anything later.

Target data

Every target carries
its own context.

Before you design anything, Foldworks pulls together what's already known about your target: its sequence and variants, where in the cell and which tissues it appears, how druggable it looks, and its known or predicted 3-D structures. Straight from the original databases, each with a citation, never a black-box middleman.

Access

Start free.
Scale on request.

The Researcher demo is free for everyone, forever. For higher limits, shared team workspaces, or running Foldworks inside your own private network, email info@foldworks.bio. Pricing depends on usage and setup.

ResearcherFreeNo signup · demo / demo
  • 5 design jobs a day
  • Unlimited structure viewer
  • Every design type + safety checks
  • Plain PDB / SDF / CSV files
TeamContact usHigher quotas, shared workspaces
  • Per-seat or per-job billing
  • Projects + shared structures
  • Priority GPU queue
  • Email + Slack support
Email info@foldworks.bio →
Enterprise · self-hostedContact usIn your private network, no data leaves
  • Docker / Helm deployment
  • Custom libraries + private targets
  • SSO, audit logs, role-based access
  • Direct researcher support
Email info@foldworks.bio →

Methods

Every Foldworks workflow is a thin wrapper around a published, peer-reviewed method, and the models, weights, and code are all documented and citable. Output is plain PDB / SDF / CSV, so it drops into whatever tools you already use.

Backbone diffusionRFdiffusion · RFantibody
VHH / nanobody designmber (Manifold Bio)
Structure predictionESMFold · AlphaFold DB · Boltz-2 · Chai-1
Small-mol dockingAutoDock Vina · GNINA · DiffDock-L
Pocket detectionfpocket · P2Rank · DoGSite3
ADMET / PK-PDADMET-AI (Chemprop) · RDKit triage
RetrosynthesisAiZynthFinder · rule-based templates
Function annotationDeepFRI (GO terms, EC numbers)
Structure prepPDBFixer · OpenMM
Hallucination designBindCraft (AF2 + ColabDesign)
Sequence designProteinMPNN · AbMPNN · soluble-MPNN
Affinity predictionBoltz-2 · Chai-1 (AF3-class)
De novo small-molREINVENT 4 (RL-based)
Epitope predictionFreeSASA · DiscoTope-3
Pose validationPoseBusters (~20 geometry checks)
Stability predictionThermoMPNN · PROSS
Post-design refinementPyRosetta FastRelax
VisualizationMol* (RCSB)