Morgan Fingerprint Classification Prioritizes Autophagy Modulators While Transcriptome Regression Fails to Distinguish Opposing Modulators
Abstract
Autophagy is a conserved lysosomal degradation and recycling pathway with established roles in aging, neurodegeneration, cancer, and metabolic disease. Computational prioritization of autophagy modulators is difficult: autophagy modulators have a variety of mechanisms and the analysis of perturbational transcriptomes could confound pathway-specific effects with generic stress responses. Here, we evaluated four formulations of the prioritization task: full transcriptome regression, gene-signature matching, binary modulator classification, and activator-versus-inhibitor direction prediction, using LINCS L1000 transcriptomics, the Human Autophagy Modulator Database (HAMdb), and a 99,433-compound ChEMBL library. We found that a Uni-Mol neural network trained to predict 12,328-gene expression responses assigned near-identical profiles to Rapamycin and Wortmannin, two compounds with opposing effects on autophagy (predicted Spearman ρ = +0.97 versus +0.43 observed), and that genome-wide diagnostics detected no association between molecular structure and expression. In contrast, XGBoost on 2,048-bit Morgan fingerprints distinguished modulators from non-modulators with compound-level AUROC 0.845 using L1000 context and 0.84 to 0.88 (95% bootstrap CI) from structure alone, ranking held-out canonical controls at the top of the score distribution, and enriched autophagy and mTOR/PI3K annotations up to ~36× and ~58× over baseline on the ChEMBL library. Direction prediction remained near chance (macro F1 0.532), limited by the lack of samples in the HAMdb compounds that are marked as autophagy inhibitors (only 132 inhibitors). Taken together, these results suggest that binary structure-based classification can prioritize autophagy modulators, while transcriptome regression and structure-based direction prediction are not supported by current data.
Related
- BigFingerprint: A Topology-aware Extended-connectivity Fingerprint for Polymers
- Auditing frozen chemical language model embeddings for drug-delivery screening: fingerprint baselines, random-initialization controls, and scaffold-memorization control tasks
- Loss-Engineered Terahertz Pixelated Metasensors for Enhanced Molecular Vibrational Fingerprint Sensing.
- RINSE: A reciprocal-space fingerprint for comparing crystal structures
- Broadband Terahertz Fingerprint Metasurfaces via Localized Surface Plasmon Resonance and Electromagnetically Induced Transparency Synergy.
BibTeX
@misc{alipour2026morgan,
title = {{Morgan Fingerprint Classification Prioritizes Autophagy Modulators While Transcriptome Regression Fails to Distinguish Opposing Modulators}},
author = {Abolfazl Alipour},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010770/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010770/v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Morgan Fingerprint Classification Prioritizes Autophagy Modulators While Transcriptome Regression Fails to Distinguish Opposing Modulators
AU - Abolfazl Alipour
PY - 2026
DA - 2026/10/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010770/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010770/v1
AB - Autophagy is a conserved lysosomal degradation and recycling pathway with established roles in aging, neurodegeneration, cancer, and metabolic disease. Computational prioritization of autophagy modulators is difficult: autophagy modulators have a variety of mechanisms and the analysis of perturbational transcriptomes could confound pathway-specific effects with generic stress responses. Here, we evaluated four formulations of the prioritization task: full transcriptome regression, gene-signature matching, binary modulator classification, and activator-versus-inhibitor direction prediction, using LINCS L1000 transcriptomics, the Human Autophagy Modulator Database (HAMdb), and a 99,433-compound ChEMBL library. We found that a Uni-Mol neural network trained to predict 12,328-gene expression responses assigned near-identical profiles to Rapamycin and Wortmannin, two compounds with opposing effects on autophagy (predicted Spearman ρ = +0.97 versus +0.43 observed), and that genome-wide diagnostics detected no association between molecular structure and expression. In contrast, XGBoost on 2,048-bit Morgan fingerprints distinguished modulators from non-modulators with compound-level AUROC 0.845 using L1000 context and 0.84 to 0.88 (95% bootstrap CI) from structure alone, ranking held-out canonical controls at the top of the score distribution, and enriched autophagy and mTOR/PI3K annotations up to ~36× and ~58× over baseline on the ChEMBL library. Direction prediction remained near chance (macro F1 0.532), limited by the lack of samples in the HAMdb compounds that are marked as autophagy inhibitors (only 132 inhibitors). Taken together, these results suggest that binary structure-based classification can prioritize autophagy modulators, while transcriptome regression and structure-based direction prediction are not supported by current data.
ER -