OmniBind: Proteome-Wide Promiscuity Predictions for Early-Stage Drug Screening
1. The paper presents OmniBind, a message‑passing neural network that estimates a compound’s mean predicted binding affinity across 15,405 human proteins directly from its SMILES string, achieving roughly one thousand predictions per second—orders of magnitude faster than full proteome profiling.
2. Promiscuity is defined as the unweighted average predicted pKD over the proteome; the authors show this metric correlates almost perfectly (Spearman ρ > 0.96) with tissue‑specific, abundance‑weighted formulations, indicating global binding propensity is the dominant factor.
3. OmniBind’s promiscuity scores agree with experimental binding data from ChEMBL, PDSP, and Safety Pharmacology databases at a level approaching assay reproducibility limits, and display a stronger negative correlation with maximum unbound plasma concentration (Cmax) than molecular weight alone, validating its pharmacokinetic relevance.
4. By combining target affinity (from Ligand‑Transformer or other predictors) with the proteome‑wide promiscuity to compute a specificity Z‑score, ranking drugs by specificity consistently outperforms affinity‑only ranking, improving BEDROC and enrichment factors across all thresholds tested.
5. The specificity advantage is robust to the underlying affinity predictor; similar enrichment gains are observed when using SSM‑DTA or Boltz‑2 predictions, and cross‑validation with the Eurofins SafetyScreen panel confirms the generality of the promiscuity signal.
6. FDA‑approved drugs show lower OmniBind promiscuity scores than a matched set of random small molecules (mean 5.75 pKD vs 5.88 pKD, p < 0.001), supporting the idea that early‑stage screening should penalize promiscuity to reduce drug attrition.
7. A publicly accessible web server returns promiscuity, standard deviation, and optional specificity scores within seconds, enabling seamless integration of proteome‑wide off‑target assessment into high‑throughput virtual screening or generative chemistry pipelines.
📜Paper:
biorxiv.org/content/10.64898…
#DrugDiscovery #ComputationalChemistry #MachineLearning #Proteomics #OmniBind