Multitask Learning for Membrane Permeability Prediction across Assays with Prediction Reliability Assessment

Membrane permeability is a critical determinant of drug exposure and distribution; however, experimental permeability data are unevenly available across assay systems, thereby restricting the robustness and predictive utility of the in silico models. In this study, we curated five apparent permeability (P_app) data sets (Caco-2, MDCK, Ralph Russ canine kidney (RRCK), LLC-PK1, and parallel artificial membrane permeability assay (PAMPA)) and developed a multitask graph convolutional neural network (MT-GCN) to predict P_app jointly across assays, benchmarking it against a fingerprint-based random forest (RF), single-task graph convolutional network (GCN) (ST-GCN), and transfer-learning GCN (TL-GCN). In 10-fold cross-validation, MT-GCN achieved the best overall performance for Caco-2 and MDCK (R² values of 0.503 and 0.500, respectively). It also delivered the clearest gains in data-limited assays, where the ST-GCN did not exceed the RF baseline. MT-GCN achieved R² = 0.558 for LLC-PK1 and R² = 0.449 for RRCK, whereas PAMPA remained an exception in which RF performed comparably or better (RF R² = 0.532). A Caco-2 subsampling study showed that MT-GCN improves data efficiency, outperforming RF and ST-GCN, even with a relatively small number of training samples. Furthermore, we introduced an applicability domain framework to quantify the prediction confidence at the compound level. We trained assay-specific error models using ensemble dispersion, learned-space similarity, and local-consistency metrics and defined a reliability score as the probability that a prediction falls within a 2-fold error. These scores were consistent with the observed errors and enabled practical accuracy coverage control by filtering out predictions with low-reliability scores. The findings of this study indicate that integrating multitask learning with applicability domain assessment facilitates precise and practical permeability prediction across assays, enhancing performance and decision confidence for data-limited end points.