Biology Models

25 Milliseconds per Compound, 1,250x More Screened per GPU-Hour: Genesis's Ruby Skips the 3D Pose, but the Model Isn't Open and the Evaluation Is Retrospective

November 21, 20261 views阅读中文原文
25 Milliseconds per Compound, 1,250× More Screened per GPU-Hour: Genesis's Ruby Skips the 3D Pose, but the Model Isn't Open and the Evaluation Is Retrospective

25 Milliseconds per Compound, 1,250× More Screened per GPU-Hour: Genesis's Ruby Skips the 3D Pose, but the Model Isn't Open and the Evaluation Is Retrospective

The short version

  • Who: Genesis Molecular AI, a drug-discovery AI company in San Mateo, California (formerly Genesis Therapeutics), behind the co-folding foundation model Pearl and the GEMS platform. On 20 November 2026 it published a technical report introducing Ruby, a small-molecule potency predictor.
  • What it does: predicts how strongly a compound will act on a target. It does not generate a 3D binding pose and then score it. Instead it consumes the protein–ligand token pair representation produced inside Pearl's trunk and regresses potency directly. Inputs are just a protein sequence and a ligand SMILES.
  • How fast: 25 milliseconds per compound at the lowest setting (25.11 ms), 4.06 seconds at the highest. Against Boltz-2's reference setting, which Genesis puts at roughly 30 seconds per ligand, that works out to 1,250×; against Nesso-1, 42×.
  • How accurate: across seven public screening targets, mean nEF@1% climbs from 0.24 at the lowest setting to 0.43 at the highest, versus 0.23 for Boltz-2 on the same targets. It posts the best average correlation on all four public potency benchmarks: OpenFE 0.74, CASP16 0.77, FEP+4 0.86, OpenBind 0.57.
  • What was released: a 14-page technical report, limitations included, plus GenesisPotency-2026, a roughly 5.6-million-row dataset under CC BY-SA 4.0 with row-level provenance and audit annotations.
  • What was not: no weights, no API, no trial access, no model code on GitHub. Only the dataset repository is public. Every evaluation is retrospective.

This article is compiled from multiple media reports. Figures follow Genesis's own technical report and press release, cross-checked against independent rewrites in Korean and Chinese outlets.

Bottom line first

What you can actually take away from this launch is not "a stronger drug-discovery model." It is a report that spells out where the method fails, plus a 5.6-million-row audited dataset. You cannot download the model, and the benchmarks were never run prospectively on live programs. But the methodology for cleaning up public potency data is usable today by any team building an in-house activity model.

What Ruby actually does

Skipping the pose is the whole point

Structure-based potency prediction normally runs in two steps: predict the protein–ligand complex geometry, then score that geometry. The first step is the problem. Pose generation is expensive, and its accuracy varies a lot by method. If you plan to screen a billion compounds, step one eats the entire budget.

Ruby takes a different cut. Before Pearl's trunk emits atomic coordinates, it already holds an intermediate representation of the relationships between protein and ligand tokens. Ruby reads that representation straight into a potency regression and never generates an explicit pose for any compound.

Two consequences follow:

  1. The most expensive part of the inference budget disappears, which is what makes virtual-screening-scale throughput possible at all.
  2. A single checkpoint can be dialled across four inference efforts — Low, Medium, High, Max — spanning more than two orders of magnitude of compute. The same weights triage a million-compound library and then rank closely related analogues inside a chemical series.

The trade-off is explicit: you lose the "how did this molecule bind" artifact. The report concedes that poses still matter for interpreting structure–activity relationships and suggests generating them only for the handful of compounds worth a closer look.

The speed–enrichment trade-off is a table you can read yourself

On seven public screening targets — three from large-library docking campaigns and four from CACHE challenges — Ruby re-ranked the compounds those campaigns actually sent for testing and measured early enrichment:

  • Low: 25.11 ms per ligand, mean nEF@1% = 0.24
  • Medium: 180.77 ms, 0.35
  • High: 348.48 ms, 0.39 (Nesso-1 sits at 0.18, and Ruby is 3× faster)
  • Max: 4.06 s, 0.43 (Boltz-2's reference setting: 0.23, p = 0.007)

The 1,250× figure is not rhetorical. The report states that the GPU time needed to score 10,000 compounds with Boltz-2's reference setting would cover roughly 12.5 million compounds with Ruby Low. It is a relative throughput number, not an absolute speed, and it moves with the hardware.

Hardware is where a second story sits. Genesis worked with NVIDIA to optimise the low-effort inference path for Blackwell: NVFP4 kernels in Transformer Engine, cuBLAS BF16x9, CUDA Graphs, and cuEquivariance attention. On one example target, warm inference fell from 44.2 ms to 14.6 ms per ligand. That is roughly 3× more throughput from kernel work alone — same model, same maths, different implementation.

How to read the accuracy numbers: four gaps of 0.0x are not dominance

Zero-shot results on the public benchmarks, each stated as Ruby versus the best comparator:

  • OpenFE (58 source series across 39 targets, Pearson r): 0.74, versus IsoDDE 0.73 and FEP+ 0.72
  • CASP16 (Pearson r): 0.77, versus IsoDDE 0.75 and Nesso-1 0.65
  • FEP+4 (four targets, Pearson r): 0.86, versus IsoDDE 0.85 and Nesso-1 0.81
  • OpenBind (494 compounds, Spearman ρ): 0.57, versus ApherisAffinity 0.54 and Nesso-1 0.51

Three things need saying.

  1. The first three are Pearson r; OpenBind is Spearman ρ. Mixing them is an error, and the report keeps them separate.
  2. The gap to IsoDDE is consistently 0.01 to 0.02. Statistically that reads as the same tier, not a generational jump. The real separation appears against open models such as Boltz-2 and Nesso-1.
  3. A meaningful share of the comparator scores are cited, not re-run. The report states plainly that IsoDDE, ApherisAffinity, TerraBind, and Nesso-1's OpenBind result are published summary values, because per-compound predictions were unavailable for pairing. That is why those error bars are missing from the figure. It is not fabrication, but it does mean the paired comparison is not strict.

The more persuasive evidence is the partner-programme set: eight biochemical and cellular assays across five ongoing Genesis–Incyte programmes. With no prior data from those assays, Ruby reached a mean Spearman of 0.61 and Pearson of 0.59, against 0.46 and 0.45 for Boltz-2, the strongest aggregate comparator. The report offers a useful anchor: for many campaigns, the Spearman correlation between two different experimental assays of the same target can be 0.70 or even lower. On that scale, 0.61 is approaching the reproducibility ceiling of the experiments themselves.

The dataset may be worth more than the model

GenesisPotency-2026 is the most concrete output of this launch, because it is something you can use immediately. Its construction is itself evidence for how messy public potency data is:

  • Raw material: ChEMBL 37, BindingDB release 202607, and a PubChem HTS pull from October 2025 — roughly 157 million rows before filtering.
  • Quality control and sampling (ligand and assay eligibility, grouping by source document, target and readout, restricting to measurements dated before 1 January 2023) cut that to 5.9 million potency data points.
  • Audit scope: 2.51 million ChEMBL and BindingDB records, of which 1.24 million belonged to documents that could actually be read. That covered 8,995 documents and 115,235 sampled compounds.
  • Audit outcome: 699,742 entries retained unchanged, 52,575 flagged for correction, 486,600 removed.
  • Published result: 5.41 million audited public-source data points (the website and press release say "5.6 million rows"; the report's Figure 1c gives 5.41M — both figures are in the source material, and we follow the report).
  • Coverage against AQAffinity's released training set: 3,110 proteins versus 2,691 (1.16×), 1.00 million compounds versus 0.82 million (1.23×), and 2.78× as many protein–compound pairs. Neither set contains the other.

Two audit findings are worth remembering:

  • In one patent, 885 measurements of FLAP inhibitors had been assigned — independently, in two separate databases — to flap endonuclease 1, an unrelated protein. Two databases "agreeing" turned out to be the same transcription error travelling twice.
  • In another document, 10 of 15 sampled structures carried potency values belonging to different compounds.

That is why the report insists that agreement between databases is not independent confirmation. All audit annotations ship with row-level provenance, so you can disagree with their criteria and re-filter yourself.

The dataset lives at Genesis-Molecular-AI/genesis-potency-dataset on GitHub under CC BY-SA 4.0.

How to use it, and who it suits

What you can do now:

  1. Train your own model on the data. CC BY-SA 4.0 permits commercial use, provided derivative datasets stay open under the same terms. If your team has been stuck on "public data is too dirty and cleaning costs too much," 5.41 million audited rows removes a large chunk of that work.
  2. Use the report as a template for evaluation design. It separates early enrichment (nEF), within-series Spearman, and post-calibration MAE cleanly, and it distinguishes public benchmarks from live partner assays. That framework can be lifted into an internal review process as-is.
  3. Cost out your own pipeline against it. If you currently run dock-then-score, the 25 ms versus 30 s gap is the number to plug into a build-versus-buy estimate.

What you cannot do:

  • You cannot get Ruby's weights or call an API. Access means talking to Genesis about the GEMS platform.
  • You cannot reproduce those 0.74 or 0.86 numbers on your own targets. Every figure comes from Genesis's own runs.

Who should care: computational chemistry and ML teams working on hit identification and lead optimisation; smaller companies that want an in-house potency model but lack clean data; and anyone doing data governance, for whom the audit methodology may outvalue the model.

Five ways to misread this launch

  • Don't treat "0.43 versus 0.23" as universal. It is a mean nEF@1% over seven specific public targets. Change the targets or the chemical space and the ordering can move.
  • Don't treat comparator scores as like-for-like. Several IsoDDE and Nesso-1 numbers are cited published summaries with no paired per-compound predictions.
  • Don't skip the word "retrospective." The hit-identification benchmarks re-rank compounds that were already experimentally tested, so they do not directly establish the same enrichment over the enormous libraries those compounds came from. Rolling temporal evaluation preserves the order of assay measurements but cannot show how Ruby-guided choices would have changed what was synthesised or how a programme unfolded.
  • Don't assume the audit is complete. The report states that only records with openly accessible source material could be audited, and that sampling happened within documents (up to 15 compounds each), so retained records may still contain undetected errors.
  • Don't soften "not open" into "open-sourced." Genesis is a commercial company and the model runs through GEMS. Releasing the methodology and the data was their choice; describing it as an open model would be wrong.

Cost and time

  • Dataset: free, CC BY-SA 4.0, with the share-alike obligation for derivatives.
  • Model: no public pricing. Genesis has not disclosed API or on-premise terms; that is a sales conversation.
  • Reference hardware: timings were measured on NVIDIA B200, and the 44.2 → 14.6 ms improvement comes from Blackwell-specific kernel work. Expect different numbers on other accelerators.
  • Effort to test the idea yourself: validating "regress potency from a co-folding model's intermediate representation, no pose" on your own data is a few weeks of prototype work using public data and an open co-folding model. Reaching the report's four effort settings plus kernel optimisation is a team-scale investment.

Where to start

  1. Pull genesis-potency-dataset and inspect its schema and audit annotation fields to see whether it aligns with your existing data.
  2. Pick a target where you hold wet-lab data, train a baseline on public data, and compare nEF@1% against your current dock-and-score flow. Re-rank the same tested compounds, or the comparison is meaningless.
  3. If you screen with Boltz-2, measure your per-compound time and multiply by the 1,250× order of magnitude to estimate whether switching paths pays off.
  4. Watch two things: whether Ruby becomes available through GEMS, and whether anyone independently reproduces that 0.61 on their own targets.

Sources

  • Genesis announcement: https://www.genesis.ml/news/introducing-ruby-state-of-the-art-potency-prediction-across-discovery-regimes
  • Technical report (PDF): https://cdn.prod.website-files.com/69d6baace8258c456e8a0f87/6aff5b9933da7ef317571829_c9db249aad6de30e829dd56f7ffea826_genesis-ruby-2026.pdf
  • GlobeNewswire release: https://www.globenewswire.com/news-release/2026/11/20/3407667/0/en/genesis-molecular-ai-announces-potency-prediction-breakthrough-with-ruby-releases-potency-training-dataset.html
  • Dataset repository: https://github.com/Genesis-Molecular-AI/genesis-potency-dataset

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...