Skip to main content
Back to timeline
Scientific ReportsSource publication:

ADMET multitask grouping stops relying on trial and error: TAG-lite uses shared-encoder gradient affinity to split largely disjoint endpoints into seven groups

Synopsis

ADMET endpoints are usually measured on different compounds, so existing task-compatibility measures that require shared compounds do not transfer directly. TAG-lite instead compares the direction of two tasks' gradients in the parameter space of a shared AttentiveFP encoder, estimating affinity from one multitask training configuration and clustering 27 TDC endpoints into seven groups. Affinity measured before group-wise training tracked group-level transfer (r = 0.881, permutation P = 0.006), but stopped tracking transfer magnitude below roughly 5% compound overlap.

Source-provided article image: Gradient affinity from a shared encoder guides multitask grouping of ADMET endpoints measured on largely disjoint compounds
Figure 1 · Page 3

Interpretation

Multitask ADMET models can now be grouped by measured gradient affinity rather than by biological category or physicochemical similarity, and that affinity can be computed when two endpoints share no compounds at all. Earlier task-compatibility measures either required both tasks' losses on a common batch or grouped endpoints by biological category, physicochemical similarity or pairwise co-training, without measuring how the model shares information across endpoints. Across 27 Therapeutics Data Commons endpoints, the 351 endpoint pairs shared a median Jaccard overlap of only 0.99%, and 292 pairs fell below 5%; an affinity matrix from one shared AttentiveFP encoder, averaged over three seeds, gave seven groups and one singleton under Ward clustering.

Affinity is measured before group-wise training begins, so it can indicate which endpoints belong together before any multitask training is committed. The direction of transfer is unknown until the joint model has been trained, and training conflicting endpoints produces negative transfer through gradient conflict. Group mean affinity correlated with group mean relative gain at Pearson r = 0.881 (95% CI 0.380–0.982, permutation P = 0.006); the task-level correlation was r = 0.469 (permutation P = 0.016), and treating transfer direction as a binary outcome gave an area under the curve of 0.745, though that discrimination was no longer significant when whole groups were resampled (P = 0.082), so the task-level result is reported as exploratory.

Grouping by affinity harms fewer endpoints than grouping without measured affinity, and the difference comes mainly from the grouping rather than from multitask training itself. Anti-affinity grouping, random grouping, pharmacokinetic-category grouping and task-type grouping all made multitask training a net-negative intervention under the same protocol. Across 26 endpoints, TAG-lite with PCGrad gave a mean relative gain of −0.43%, against −13.10% for anti-affinity, −12.81% for random grouping, −11.54% for pharmacokinetic category and −10.78% for task type; paired Wilcoxon tests put TAG-lite above random and anti-affinity (P = 0.005 and 0.011) and above both knowledge-based groupings (P = 0.009 and 0.010), but each alternative partition was realised once, with no estimate of between-partition variance.

Gradient affinity reads transfer magnitude reliably only at the group level, loses magnitude information at very low compound overlap, and still retains directional information there. Prior work argued that gradient alignment becomes uninformative below roughly 30% sample overlap, while most ADMET endpoint pairs sit far below that level. Stratifying by within-group mean Jaccard overlap at 5%, the affinity–gain correlation was r = +0.700 (P = 0.0037) in the high-overlap stratum (n = 15) and r = −0.186 (P = 0.583) in the low-overlap stratum (n = 11); direction discrimination gave areas under the curve of 0.759 and 0.750 in the same strata.

Perspective

The conclusions rest on 27 TDC ADMET endpoints, an AttentiveFP encoder and TDC scaffold splits; the affinity threshold of 0.155 and the 0.5% overlap threshold are both in-sample operating points chosen on the same 26 endpoints and have not been validated on held-out endpoints or another encoder.

Group mean affinity takes only seven distinct values, and the lower bound of the task-level area-under-the-curve interval approaches chance; pairwise affinity ordering is reproducible across seeds (Spearman ρ = 0.62–0.67), but the discrete partition at k = 8 is not (mean adjusted Rand index 0.16). HydFE was excluded after its outcome was observed, the 0.5% overlap condition rests on a single informative case, and whether discarding gradient magnitude caused that failure remains untested.