Structure and ligand modelling
We predict structures of proteins, nucleic acids and biomolecular complexes, including protein–ligand, protein–DNA and protein–RNA systems. We also model where ligands bind and estimate binding strength to help rank candidates before assay work. We check predictions for physical plausibility and refine them with physics-based methods when needed. We use open models that others can also inspect and use.
Example tools: Boltz-2 for structure prediction and affinity estimates, OpenFold3, Chai-1, and PLACER for catalytic and binding-site geometry.
Protein design and variant analysis
We design new protein backbones and binders for chosen targets, including antibodies and nanobodies, with CDR redesign, humanisation scoring and candidate-panel triage. Depending on screening capacity, we test designs individually or advance them to display-based screening. We also scaffold catalytic motifs into new backbones, design active-site variants, and filter them by catalytic geometry. Sequence-design tools account for ligands, metals and packed cores. Protein language models provide zero-shot variant scores and sequence embeddings; structural screening helps narrow mutation libraries before cloning.
Example tools: RFdiffusion3 (Foundry), BoltzGen, BindCraft, LigandMPNN, Germinal, RFantibody, AntiFold, Riff-Diff, ZymCTRL, ESM C and ESMFold2, and AMPLIFY.
Molecular simulation and structure analysis
We simulate protein stability, dynamics and conformational change, including complexes with substrates, cofactors and other ligands. We parameterise ligands with their targets and use machine-learned interatomic potentials when classical force fields are not accurate enough. We also compare structures across variants, interfaces and binding pockets, and create publication-quality figures.
Example tools: GROMACS, OpenMM, AmberTools, OpenMM-ML, PyMOL and ChimeraX.
What to expect
Designs are predictions and need experimental testing. Based on our experience so far and published benchmarks, a success rate on the order of one in ten is a reasonable expectation. A published RFdiffusion binder study reported about one in five designs binding in the experimental screen. Rates depend on how a project defines success.
Accuracy also depends on how well a model’s training data represents the target. Performance on well-represented targets may not carry over to unusual folds, novel ligand chemistry, or cryptic and allosteric pockets. We discuss this at the start of a project and treat predictions for novel targets as hypotheses to test, not results.