Multi-Task Learning for Joint Prediction of Acetylcholinesterase and Butyrylcholinesterase Inhibition

Cholinesterase inhibition can disrupt cholinergic signaling and is therefore an important endpoint for neurotoxicity screening. Nevertheless, experimental inhibition data on acetylcholinesterase (AChE) and butyrylcholinesterase (BuChE) remain unavailable for many chemicals, and existing regression models have generally considered the two enzymes separately. This study developed a multi-task learning framework for jointly predicting AChE and BuChE inhibitory activities. Bioactivity records were curated from ChEMBL, yielding 2281 compounds with paired the base-10 logarithm of the half-maximal inhibitory concentration (logIC50) measurements. Five individual molecular representations were evaluated using single-task learning, single-representation multi-task learning, and multi-representation fusion multi-task learning (MRF-MTL). The best-performing MRF-MTL model achieved coefficients of determination on the test set (R2test) of 0.622 and 0.678 for AChE and BuChE, respectively. A structure–activity landscape-based applicability domain (ADSAL) was subsequently applied. Under the strict ADSAL settings, the R2test values reached 0.793 for AChE and 0.859 for BuChE, with 128 and 129 test compounds retained, respectively. Parameter-sensitivity analysis showed that the similarity-weighting parameter strongly influenced prediction performance, whereas the activity-inconsistency threshold was the principal determinant of domain coverage. SHapley Additive exPlanations (SHAP) analysis showed that fingerprint-derived structural features accounted for most of the feature attribution and revealed distinct patterns for the two endpoints. The proposed framework integrates joint prediction, applicability-domain assessment, and model interpretation, supporting efficient neurotoxicity screening and sustainable chemical management.