From Predicted Ki to Surrogate IC50: Similarity-Guided Empirical Calibration of Drug–Target Affinity Predictions
TL;DR
A DeepPurpose model to estimate inhibition constants (Ki) from molecular graphs and protein sequences was trained and whether those predictions could be aligned empirically with measured IC50 values without treating Ki and IC50 as interchangeable was asked.
Abstract
Drug target affinity models return the endpoint on which they are trained, whereas medicinal chemistry decisions are often made with a different assay readout. Here, we trained a DeepPurpose model to estimate inhibition constants (Ki) from molecular graphs and protein sequences and asked whether those predictions could be aligned empirically with measured IC50 values without treating Ki and IC50 as interchangeable. A BindingDB-trained checkpoint retained useful cross-target ranking on the Davis kinase benchmark without Davis training data (concordance index 0.860). We then rebuilt the Ki training set from ChEMBL 37 records coded as Binding assays (assay_type = 'B'), after removing censored records and targets with poor replicate reproducibility. This reduced median fold error from 16.35x to 9.23x on a leakage-cleaned kinase panel and from 15.30x to 7.90x on a 14-target non-kinase panel before any IC50 calibration. Similarity-guided leave-one-out calibration further reduced the non-kinase panel median error to 3.12x for the original checkpoint and 3.22x for the ChEMBL checkpoint at Tanimoto T = 0.6. Because retraining removed a substantial part of the apparent correction, we interpret the calibration as a target- and chemistry-dependent empirical offset between model output and IC50 assay space, not as a mechanistic Ki-to-IC50 conversion. In a separate project-level stratification of public records, median replicate variability was 2.41x for the subset classified as biochemical Ki and 3.30x for biochemical IC50; a small-sample correction placed the Ki variability nearer 2.8x. These values provide an empirical scale for the remaining calibration error rather than a theoretical performance limit. Similarity, rather than the number of calibrators, governed the main accuracy coverage trade-off. The resulting values are surrogate IC50 estimates for cross-target triage; within-target ranking remains a limitation of the present architecture.
