Skip to content
DraconiticResearch
Search Sources API
226,742 studies31 sourcesUpdated 43 min ago
Peer-reviewedCC BY 4.0International Journal of Molecular Sciences

Target-Specific Benchmarking of Potency Prediction on Activity-Cliff Compounds Across 33 Human Protein Targets Using Inhibition Constants

6 affiliations
  1. 1SilicoScientia Private Limited (Registered Office), Nagananda Commercial Complex, No. 07/3, 15/1, 18th Main Road, Jayanagar 9th Block, Bengaluru 560 041, India
  2. 2SilicoScientia Private Limited (Corporate Office), Centre for Cellular and Molecular Platforms (C-CAMP), GKVK Campus, Bellary Road, Bengaluru 560 065, India
  3. 3School of Pharmacy and Research, Dnyaan Prasad Global University, Pune 411 018, India
  4. 4Dr. D. Y. Patil Institute of Pharmaceutical Sciences and Research, Pune 411 018, India
  5. 5Department of Pharmaceutical and Biological Chemistry, School of Pharmacy, University College London, London WC1N 1AX, UK
  6. 6SR University, Warangal, Telangana 506 371, India

Abstract

Activity cliffs are pairs of closely related compounds whose potencies differ far more than their structures suggest, and the compounds involved in them remain a leading cause of local failure in ligand-based potency prediction. The MoleculeACE benchmark established that models predict these compounds poorly and that the gap differs between datasets. What has been missing is a panel built on one activity endpoint, a per-target account of how large the gap is, and a measurement of how far it depends on the way the model is chosen. We curated inhibition constant (Ki) data for 33 human protein targets from ChEMBL release 37 and BindingDB and annotated activity cliffs using the MoleculeACE criteria. Sixteen descriptor and algorithm combinations per target were trained to predict potency and scored separately on cliff compounds, with the winner chosen by five-fold cross-validation inside the training set. Cliff frequency ranged from 3.95% to 61.34%. Error on cliff compounds exceeded global error for 26 of 33 targets, by 12.8% on average (Wilcoxon p = 3.4 × 10−5), with a range from an 8.0% reduction to a 122.0% increase. Choosing the model on the test set instead inflated apparent cliff accuracy by 0.119 pKi units and the number of targets that showed lower cliff-specific error than global error fell from 14 to seven once selection was corrected. Cliff-specific error has to be measured per target and quoted with the size of the subset it rests on.

Related

  1. Short-read sequencing benchmarking with donor-specific assemblies
    Sean R. McGee, Joshua D. Smith, Christian D. Frazar +8
    Genome BiologyPeer-reviewedCC BY 4.0Molecular biology & gene regulationGenetics & genomicsCrossref
  2. Inhibition of JEV infection using β-Catenin specific inhibitor, iCRT-14
    A. Datey, S. Ghosh, S. Chatterjee +4
    bioRxivPreprintAll rights reservedFull textMolecular biology & gene regulationbioRxiv
  3. Benchmarking graph-based models for in-silico toxicity prediction in drug discovery
    Noel Suarez-Barro, Manuel Lama, Juan C. Vidal
    arXivPreprintarXiv licenceFull textBiophysicsBioinformatics & computational biologyarXiv
  4. Prediction of bacteria-specific drug metabolism using a cross-attention-based model.
    Jaehyeon Park, Si Eun Kim, Heeyoung Yang +3
    Briefings in bioinformaticsPeer-reviewedCC BYFull textMolecular biology & gene regulationEurope PMC
  5. Sequence-based coevolutionary prediction of species-specific interactomes
    Miguel Fernandez-Martin, Camila Pontes, Victoria Ruiz-Serra +1
    PLOS Computational BiologyPeer-reviewedCC BY 4.0Full textProteomics & metabolomicsMicrobiologyPLOSPMC
BibTeX
@article{alamri2026targetspecific,
  title = {{Target-Specific Benchmarking of Potency Prediction on Activity-Cliff Compounds Across 33 Human Protein Targets Using Inhibition Constants}},
  author = {Mohammed Musallam Mohammed Alamri and Shovonlal Bhowmick and Ritesh P. Bhole and Rupesh Chikhale and Payal R. Kute and Md Ataul Islam},
  journal = {International Journal of Molecular Sciences},
  year = {2026},
  month = nov,
  doi = {10.3390/ijms272210230},
  url = {https://www.mdpi.com/1422-0067/27/22/10230},
  note = {CC BY 4.0},
}
RIS
TY  - JOUR
TI  - Target-Specific Benchmarking of Potency Prediction on Activity-Cliff Compounds Across 33 Human Protein Targets Using Inhibition Constants
AU  - Mohammed Musallam Mohammed Alamri
AU  - Shovonlal Bhowmick
AU  - Ritesh P. Bhole
AU  - Rupesh Chikhale
AU  - Payal R. Kute
AU  - Md Ataul Islam
PY  - 2026
DA  - 2026/11/14
JO  - International Journal of Molecular Sciences
DO  - 10.3390/ijms272210230
UR  - https://www.mdpi.com/1422-0067/27/22/10230
AB  - Activity cliffs are pairs of closely related compounds whose potencies differ far more than their structures suggest, and the compounds involved in them remain a leading cause of local failure in ligand-based potency prediction. The MoleculeACE benchmark established that models predict these compounds poorly and that the gap differs between datasets. What has been missing is a panel built on one activity endpoint, a per-target account of how large the gap is, and a measurement of how far it depends on the way the model is chosen. We curated inhibition constant (Ki) data for 33 human protein targets from ChEMBL release 37 and BindingDB and annotated activity cliffs using the MoleculeACE criteria. Sixteen descriptor and algorithm combinations per target were trained to predict potency and scored separately on cliff compounds, with the winner chosen by five-fold cross-validation inside the training set. Cliff frequency ranged from 3.95% to 61.34%. Error on cliff compounds exceeded global error for 26 of 33 targets, by 12.8% on average (Wilcoxon p = 3.4 × 10−5), with a range from an 8.0% reduction to a 122.0% increase. Choosing the model on the test set instead inflated apparent cliff accuracy by 0.119 pKi units and the number of targets that showed lower cliff-specific error than global error fell from 14 to seven once selection was corrected. Cliff-specific error has to be measured per target and quoted with the size of the subset it rests on.
SN  - 1422-0067
ER  -