Gradient affinity from a shared encoder guides multitask grouping of ADMET endpoints measured on largely disjoint compounds
Year: 2026
Type: article
Abstract: Multi-task learning can improve ADMET prediction, but choosing which endpoints to learn together often relies on trial and error. Existing task-compatibility measures compare tasks on shared compounds, an assumption rarely met in ADMET data, where endpoints are measured on different compounds. TAG-lite, a gradient-based workflow, estimates task compatibility even when tasks share no compounds. Across 27 Therapeutics Data Commons endpoints, whose pairs shared a median 0.99% of compounds, affinities from one shared AttentiveFP encoder, averaged over three seeds, defined seven endpoint groups. Measured after a brief warm-up, before group-wise training, affinity tracked group-level transfer ( r = 0.881, permutation P = 0.006); in an exploratory task-level analysis it predicted transfer direction (area under the curve 0.745, 95% CI 0.515–0.945). Below approximately 5% compound overlap it no longer tracked magnitude, while task-level direction discrimination was preserved. Affinity-guided grouping outperformed single random, anti-affinity, pharmacokinetic-category and task-type partitions. The only large negative-transfer case among affinity-admitted endpoints shared almost no compounds with its partners, motivating a screening rule combining affinity and compound overlap; after this post-hoc exclusion, gradient surgery numerically improved every endpoint in the admitted groups.
Source: Scientific Reports
Authors: Seoyoon Koo, Sungmin Ko, Seung-Yong Seong
Institution: Seoul National University
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7212016982
