CAFNet-DG: Task-Specific Modeling and Prevalence-Aware Evaluation for Adverse Drug Reaction Prioritization with Ordinal-Frequency Estimation

Adverse drug reaction (ADR) prioritization and ordinal-frequency estimation are related but distinct prediction objectives, while aggregate ranking performance in sparse annotation matrices can be strongly influenced by frequently annotated side effects. Within CAFNet-DG, CAFNet-D uses task-specific association and ordinal-frequency functions built on shared molecular–relational representations, while annotation prevalence is evaluated explicitly within each training fold. On a SIDER-derived dataset comprising 750 drugs and 994 side effects, replacing a shared output with task-specific functions produced the largest joint change in staged drug-disjoint ablation, increasing mAP from 0.340 to 0.382 and reducing RMSE from 1.754 to 1.142. A prevalence-only ranking achieved an mAP of 0.410, close to the CAFNet-DG value of 0.413, indicating that aggregate ADR ranking is strongly associated with the benchmark annotation distribution. Prevalence-matched, prevalence-stratified, frequently annotated side-effect removal, and per-drug analyses were therefore used to distinguish drug-conditioned ranking variation from global annotation-prevalence effects. An equal-weight same-architecture ensemble control yielded no significant differences from CAFNet-DG across seven ranking metrics after Holm correction; generic variance reduction therefore could not be excluded as an explanation for the modest score-integration gain. Scaffold-disjoint and independent-data evaluations showed modest transfer, while ordinal-frequency agreement on CT-ADE remained limited. The results support the CAFNet-D task-specific formulation and prevalence-aware evaluation, while CAFNet-DG provides the score-integration setting examined under matched ensemble controls.