Year: 2026
Type: preprint
Abstract: Predicting molecular potency from chemical structure is an important task in drug discovery, and is almost universally framed as scalar regression on the half-maximal inhibitory concentration (IC50) derived from a dose–response experiment. This formulation collapses the full experimental signal into a single number and implicitly assumes well-behaving data, even though real assay measurements might be imbalanced, noisy, and potentially censored. Because public bioactivity datasets report only the summary pIC50, the alternative of predicting the entire dose–response curve has remained essentially unexplored. Here, using an in-house dataset containing the full curves, we present a novel probabilistic approach to potency prediction that models the entire dose–response curve as the cumulative distribution function of a latent concentration variable. We benchmark it against two alternative curve-prediction models—a multi-task regression over the concentration grid and a parametric four-parameter logistic model—and against descriptor-and graph-based scalar baselines. The probabilistic model achieves the best performance on all metrics, demonstrating the competitiveness of curve-level modeling over the scalar formulation. Its advantage is two-fold. First, it exposes useful curve information, such as slope, asymptotes, and shape, that is otherwise lost in the scalar collapse. Second, it learns directly from raw measurements rather than a fitted summary statistic that itself carries uncertainty and noise. Altogether, these results show that modeling the full dose–response curve offers a robust and information-preserving alternative to conventional scalar potency prediction.
Source: ChemRxiv
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7212230795