Novelty-stratified external validation of free in silico ADMET tools
Background and Purpose. Freely available in silico ADMET platforms are used routinely to triage compounds, yet each is validated by its own developers on its own training data, so reported accuracies are neither comparable across tools nor tested on novel chemistry. We asked whether any platform is generally superior, and whether structural novelty affects all endpoints alike. Experimental Approach. SwissADME, pkCSM, ADMETlab 3.0 and ADMET-AI were benchmarked on aqueous solubility, hERG blockade, blood–brain barrier penetration and drug-induced liver injury, using external datasets with identifiable training overlap removed by InChIKey. Compounds were stratified by nearest-neighbour ECFP4 Tanimoto similarity to approved drugs. Performance was computed overall, within strata and along a continuous novelty axis, tested by stratified bootstrap, and compared with a fingerprint Random Forest. Key Results. No platform was best across endpoints; the leader rotated between SwissADME, ADMETlab 3.0 and ADMET-AI. Contrary to expectation, novel chemistry was predicted no less accurately than drug-like chemistry for solubility, hERG or blood–brain barrier penetration (pooled AUROC change −0.04 and −0.02). Only drug-induced liver injury, the one endpoint lacking a structural determinant, lost accuracy (pooled AUROC −0.15), though this endpoint difference was not significant (p = 0.083). The fingerprint baseline matched or beat at least one published platform on every classification endpoint. Conclusion. These data do not support a universal collapse of accuracy on novel chemistry: three of four endpoints show none. What loss there is falls on the endpoint without a structural basis, consistently across tools but not significantly here. Tool selection should be endpoint-specific, and in silico hepatotoxicity predictions confirmed experimentally. © 2026 by the authors. This preprint is distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/4.0/). It has not been peer reviewed and is not under consideration at the journal for which it was originally formatted. Keywords
Authors
- Rahul Kumar (ORCID: https://orcid.org/0000-0002-2277-9831)
- Rakesh Kumar Dixit
- Shankar Gejji (ORCID: https://orcid.org/0009-0002-3032-3727)
- Sagar Agrawal (ORCID: https://orcid.org/0009-0007-7633-4233)
- Naveen Maurya
- Sagar Agrawal
Institutions
- University of Lucknow (IN)
- King George's Medical University (IN)
Publication Details
- Journal
- ChemRxiv
- Published
- 2026-11-16
- DOI
- https://doi.org/10.26434/chemrxiv.15012299/v1
- Primary Topic
- Computational Drug Discovery Methods
- Type
- preprint