Can sequence-only AI speed drug hit selection without replacing assays?
An unreviewed model trained on about 82 million protein–ligand records improved early virtual-screening enrichment and ranked compounds in milliseconds. Its lab-in-the-loop result was simulated on existing datasets, not a prospective drug programme.
The short version
- What happened
- PSICHIC+ was trained on about 82 million protein–ligand records covering 4,900 proteins and 6.1 million ligands, mixing quantitative affinities with weaker screening labels.
- Who it affects
- Faster prioritisation could let smaller research teams test more targets before spending on synthesis and assays. The risk is shifting scarce experiments toward compounds favoured by training data while overlooking unfamiliar chemistry or downstream safety problems.
- What remains unproven
- Unreviewed bioRxiv preprint; processed training and evaluation data are promised upon publication.
Keep in mind
Unreviewed bioRxiv preprint; processed training and evaluation data are promised upon publication.
The active-learning campaigns were retrospective replays of existing labels, not prospective laboratory experiments.
Affinity evaluation covered 461 pairs across 16 assays, and runtime comparisons combine measured and published configurations.
Two authors are NVIDIA employees; NVIDIA hardware and BioNeMo tooling were used for optimization.
Binding affinity and enrichment do not measure selectivity, toxicity, pharmacokinetics or clinical benefit.
Source analysis · 1 linked source. Primary analysed source date: 31 October 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
One model combines precise and coarse bioactivity evidence
An Australian-led bioRxiv preprint describes PSICHIC+, a model that ranks compounds from a protein's amino-acid sequence and a molecule's SMILES string, without requiring a three-dimensional protein structure. The training corpus contained about 82 million protein–ligand records spanning 4,900 proteins and 6.1 million ligands. It mixed direct affinity measurements, functional potency, high-throughput active or inactive calls and synthetic negatives, while assigning weaker sources broader predictive distributions rather than pretending every label had equal precision.
The team retrained the model separately for two virtual-screening benchmarks after removing proteins sharing at least 90% sequence identity with each target. Across 102 DUD-E targets, PSICHIC+ achieved an enrichment factor of 13.3 in the top 5% of ranked compounds and BEDROC 0.51. A point-regression version reached 5.6 and 0.10; a classification version reached 10.8 and 0.34. On 81 DEKOIS 2.0 targets, its top-5% enrichment factor was 8.0, close to 8.1 and 8.2 for two leading structure-based comparators.[1]
Speed improved, but physics still led the affinity benchmark
For hit-to-lead ranking, the authors excluded benchmark proteins and ligands from pretraining and evaluated 461 protein–ligand pairs across 16 target assays. Sequence-only PSICHIC+ produced mean within-target Pearson correlation 0.52; Boltz-2 reached 0.60 and the physics-based FEP+ reference 0.74. The model's distribution concentration tracked error with correlation 0.52, so the team used it to adjust an ensemble with Boltz-2. That combination reached 0.68 and was the only tested method not significantly outperformed by FEP+, although the two reported comparisons were not adjusted for multiple testing.
Runtime is the proposed advantage. On one NVIDIA A100, the optimized sequence model processed a pair in a reported 1.34 milliseconds, around 15,000 times faster than the cited 20-second Boltz-2 configuration. The ensemble kept Boltz-2's approximate 20-second cost, compared with about one day per pair reported for FEP+. Those ratios mix the authors' own benchmark with published comparator runtimes and different computational workflows. They show an operating-point trade-off, not that one millisecond prediction replaces molecular simulation or a wet-lab measurement.[1]
The distinction that matters
Three levels of evidence
The study measures ranking under increasingly realistic conditions, but stops before prospective drug discovery.
Virtual screening
102 DUD-E and 81 DEKOIS targets
Better early enrichment than model variants; competitive with structure-based scores
Affinity ranking
461 pairs across 16 assays
Correlation 0.52 alone, 0.68 with Boltz-2, versus 0.74 for FEP+
Active learning
Existing labels replayed on four datasets
Promising recovery under a 360-label budget; no prospective assay
The lab loop was a replay, not a prospective discovery
The paper also simulated repeated propose-label-retrain cycles on four curated datasets: 665 compounds for SARS-CoV-2 main protease, 4,535 for USP7, 2,502 for dopamine D2 receptor and 9,997 for TYK2. Batches of 30 were selected until 360 labels had been revealed. On USP7, active retraining recovered 64.4% of the dataset's true top 1% compounds, versus 35.6% for a static zero-shot ranking. Across targets and batch sizes, the model usually recovered more top compounds than random forest, multilayer perceptron and Gaussian-process baselines.
Despite the phrase lab-in-the-loop, the study replayed labels that already existed; TYK2 used calculated relative binding free energies rather than experiments. No new compound was synthesized or prospectively assayed, and no candidate entered animal or human testing. The preprint is unreviewed, processed data are promised upon publication, and two authors work for NVIDIA, whose hardware and software support the reported speed optimization. What would change the assessment is a blinded prospective campaign with frozen budgets, novel chemistry, measured assay failures, independent replication and evidence that better ranking shortens discovery without narrowing chemical diversity or missing toxicity.[1]