Work

Transfer learning across the EC classification hierarchy overcomes data scarcity in drug-target interaction prediction

Year: 2026
Type: preprint
Abstract: Predicting drug-target interactions (DTIs) with deep learning offers opportunities to accelerate drug discovery, yet performance is often constrained by the scarcity of target-specific training data. Transfer learning relieves this limitation but requires choosing a relevant source task, a choice usually made ad hoc. Here we use the Enzyme Commission (EC) classification hierarchy as a systematic criterion for source selection, ascending the EC tree from the target enzyme until a related family carries sufficient ChEMBL bioactivity data to pre-train a deep feedforward network, which is then fine-tuned on the scarce target. We applied this approach to SHMT2 and MTHFD2, two pharmacologically understudied enzymes of the mitochondrial one-carbon (1C) pathway, which is implicated in cancer and fibrosis and supports fibroblast collagen production through glycine synthesis. With 29 active compounds for MTHFD2 and only 6 for SHMT2 which has no ligands of its own, the six coming from its paralogue SHMT1 in the same EC class. EC-guided transfer learning produced usable models in a system where training from scratch failed (Matthews correlation coefficient 0.00 for SHMT2). Virtual screening of the DrugBank library identified six candidates, three of which, carbimazole, crizotinib, and GSK-2018682, reduced TGF-β-induced collagen production and glycine accumulation in human lung fibroblasts. The methodology, datasets, and code are released as a reusable template for enzyme targets with scarce bioactivity coverage.
Language: English
Subfield: Drug Discovery
Open Access status: green
OpenAlex ID: W7236417267
Cites: 34
Cited by:
Related to: 9