Researchers have developed DepFix, a multitask machine learning framework designed to jointly predict chemical compound activity across the serotonin, dopamine, and norepinephrine transporters while deriving relative pharmacological selectivity. Monoamine transporters represent well-established therapeutic targets in antidepressant pharmacology, but publicly accessible bioactivity data across these three proteins remain notably sparse and unevenly distributed. By modeling activity differences simultaneously across the target family, the framework addresses data limitations to assist researchers in prioritizing compounds with distinct pharmacological profiles for early antidepressant discovery.
The computational architecture combines a multitask directed message-passing neural network with target-specific Morgan fingerprint random forest models. Trained on a curated dataset of 6,679 chemical compounds, DepFix achieved scaffold-test macro-averaged R² values of 0.560 for overall activity and 0.553 for activity-derived selectivity. Methodological comparisons demonstrated that blending these two distinct modeling branches delivered the clearest predictive improvement, whereas incorporating gated task adapters and an auxiliary loss yielded smaller performance gains under random splitting benchmark conditions.
Model robustness testing demonstrated that DepFix’s true-label predictive performance exceeded results from all 100 Y-randomization benchmark tests, confirming that its predictions reflect genuine chemical relationships rather than statistical artifacts. In additional analyses, temporal validation and functional-assay evaluations indicated that model transfer was endpoint-dependent across experimental contexts. To test the approach on large chemical libraries, the investigators screened 24,385 unique structures from the HERB database, effectively prioritizing candidate molecules characterized by serotonin-preferential and triple-balanced activity patterns.
Structural modeling incorporating molecular docking and molecular dynamics simulations further characterized plausible ligand poses inside the transporter binding pockets. To examine biological responses, the authors investigated an exploratory four-compound panel in retinoic-acid-differentiated SH-SY5Y cell cultures. In these cell-based assays, pterostilbene induced predominantly serotonergic alterations, whereas dihydro-N-caffeoyltyramine and 3,5-dihydroxy-1-(3,4-dihydroxyphenyl)-7-(4-hydroxyphenyl)heptane modulated all three monoamine systems, providing empirical support for the computationally predicted multi-target activity profiles.
Collectively, the study demonstrates that DepFix can jointly predict binding activity and relative selectivity across multiple monoamine transporters despite sparse training data. Although temporal and functional transfer analyses highlight important endpoint-dependent limitations in generalizability across distinct assay settings, the machine learning framework provides valuable computational decision support for screening and prioritizing promising multi-target small molecules during early antidepressant development.
Subject of Research: Machine learning prediction of monoamine transporter activity and selectivity
Article Title: DepFix: a multitask machine learning framework for joint prediction of monoamine transporter activity and selectivity
Article References: Liang, H., Ma, J., Li, R., Feng, J., & Liu, Y. (2026). DepFix: a multitask machine learning framework for joint prediction of monoamine transporter activity and selectivity. Molecular Diversity, 31(1), Article 1. https://doi.org/10.1007/s11030-026-11747-4
Image Credits: AI Generated
DOI: 10.1007/s11030-026-11747-4
Keywords: DepFix, machine learning, monoamine transporters, antidepressant pharmacology, serotonin transporter, dopamine transporter, norepinephrine transporter, directed message-passing neural network, Morgan fingerprints, random forests, scaffold test, HERB database, molecular dynamics simulations, molecular docking, pterostilbene, SH-SY5Y cells, multitask learning
Cite Scienmag News
Blake Davidson. (November 22, 2026). Machine Learning Model Predicts Monoamine Transporter Activity and Target Selectivity. Scienmag. https://scienmag.com/machine-learning-model-predicts-monoamine-transporter-activity-and-target-selectivity/
Blake Davidson. "Machine Learning Model Predicts Monoamine Transporter Activity and Target Selectivity." Scienmag, 22 November 2026, https://scienmag.com/machine-learning-model-predicts-monoamine-transporter-activity-and-target-selectivity/. Accessed 22 November 2026.
Blake Davidson. "Machine Learning Model Predicts Monoamine Transporter Activity and Target Selectivity." Scienmag. November 22, 2026. https://scienmag.com/machine-learning-model-predicts-monoamine-transporter-activity-and-target-selectivity/








