DrugForm-TAS: Target-Agnostic Selectivity as Proteome-wide Binding Propensity Estimation

Assessing the selectivity of binding a small molecule with proteins remains a fundamental yet unresolved challenge in computational drug discovery. Conventional strategies for estimating ligand selectivity depend on exhaustive cross-prediction of binding affinities across thousands of potential off-target proteins-an approach that is computationally prohibitive for large-scale virtual screening or de novo molecular design. Here, we present DrugForm-TAS (Target-Agnostic Selectivity), the first model capable of directly predicting an unconditional, quantitative measure of small-molecule proteome-wide binding propensity without requiring any prior knowledge of target proteins or affinity thresholds. Built upon our previously developed DrugForm-DTA model, which enables affinity prediction for arbitrary protein targets, and trained on a thoroughly curated dataset derived from BindingDB, DrugForm-TAS employs a lightweight transformer-like neural network that operates solely on a ligand's SMILES representation. The model acts as a fast pre-screening filter that can greatly reduce the candidate set, as evidenced by the correlation between its predictions and experimental observations. By eliminating the need for target-specific computations, DrugForm-TAS enables target-agnostic nonspecificity profiling. When combined with drug-target affinity calculation, it provides a novel tool for fast pre-screening selectivity estimation in early-stage drug design.