Skip to main content
Back to timeline
Genesis Molecular AISource publication:

Ruby predicts potency without drawing a binding pose: at 25 ms per ligand it edges out Boltz-2 running 1,250 times slower

Synopsis

In drug discovery, free-energy methods are accurate but too costly and ligand-based QSAR is cheap but fails to generalize to new scaffolds, so computational routes tend to cover either screening or lead optimization. Ruby predicts potency directly from the protein-ligand pair representation of the cofolding model Pearl, without generating an explicit binding pose, and one checkpoint runs at four inference-effort settings. On seven public screening sets, its lowest setting scores a ligand in 25 milliseconds and enriches slightly better than Boltz-2 at about 30 seconds, while the Max setting at 4 seconds per ligand roughly doubles enrichment to 0.43.

Source-provided article image: Introducing RUBY: State-of-the-Art Potency Prediction Across Discovery Regimes
Genesis Molecular AI

Interpretation

The potency model Ruby no longer generates an explicit binding pose for each compound; it reads the representation of all relationships between protein and ligand tokens from the trunk of the cofolding model Pearl and predicts potency directly. Structure-inspired potency prediction typically generates an explicit pose and then scores it, but pose generation is costly, often inaccurate, and limits the scale of compounds that can be virtually screened. On seven public screening sets, Ruby's lowest setting at 25 milliseconds per ligand enriches slightly better than Boltz-2 at about 30 seconds, which is 1,250 times slower; raising effort to Max at 4 seconds per ligand roughly doubles enrichment to 0.43, ahead of every method evaluated.

Dropping pose generation lets the same model checkpoint work at four inference efforts, Low through Max, spanning more than two orders of magnitude of compute. The biggest part of the inference budget for diffusion-based cofolding is removed, so the same model can triage a million compounds and rank carefully within a series. Working with Nvidia to optimize the low-effort inference path, moving to NVFP4 kernels with Transformer Engine, cuBLAS BF16x9, CUDA Graphs and cuEquivariance attention cut warm inference on an example target from 44 to 15 milliseconds per ligand at identical model effort, roughly threefold more throughput from the kernel work alone.

On real program data, Ruby ranks potency usefully zero-shot, without any prior data from those specific assays. Public benchmarks are necessary but not sufficient; real programs are where performance actually matters. Across 8 assays over 5 diverse programs, zero-shot Ruby reaches an average Spearman correlation of 0.61, above all other benchmarked models; the Spearman correlation between two different experimental assays of the same target can be 0.70 or even lower; recalibrating on earlier within-assay measurements improves it further.

Perspective

The demonstrated scope is re-ranking on seven public screening sets, zero-shot comparison on four public potency benchmarks, and 8 assays from 5 diverse programs; inputs are protein sequence and ligand SMILES, with no other information used in the hit-identification evaluation.

The 0.61 on real programs is a zero-shot average Spearman correlation, and the correlation between two different experimental assays of the same target can itself be 0.70 or lower; recalibrating on earlier within-assay measurements improves it further, but the size of that improvement is not given in the material.