Causal-IQD-DTA: Counterfactual Interaction-Quality Disentanglement for Robust Drug–Target Affinity Prediction

khan huang, Chang Li

Keywords: drug--target affinity, protein--ligand learning, multimodal life-science models, causal representation learning, abundance shortcut robustness

Abstract:

Drug–target affinity (DTA) prediction estimates
binding strength between compounds and protein
targets, providing an early computational screen
before expensive biochemical assays. Structure-
aware DTA models now combine molecular, se-
quence, structural, and conformer inputs, but
complex-level affinity labels can entangle selec-
tive binding evidence with abundance shortcuts
such as ligand size, pocket size, and candidate
contact mass. The central goal is robustness under
abundance shortcuts: affinity predictions should
depend on selective binding evidence rather than
changes in contact opportunity alone. Causal-
IQD-DTA addresses this bottleneck by separating
these signals: zQ aggregates pocket-calibrated
atom–residue evidence for binding, and zN en-
codes size and contact-mass factors that describe
the candidate interaction map. We then perform
representation-level counterfactual intervention
on zN while holding zQ fixed, suppressing affinity
changes caused by nuisance abundance. Across
Davis, KIBA, and PDBbind, the model improves
random and transfer splits, with diagnostics show-
ing lower abundance-residual coupling and re-
duced counterfactual sensitivity.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 121