Title: SpliceBind: Enhancing Binding Pocket Druggability Prediction with Isoform Awareness

Abstract:

Targeted kinase inhibitor therapy faces significant challenges, as splice-mediated drug resistance impacts up to 40% of patients. Current leading tools for assessing druggability are limited by their reliance on single structural inputs, rendering them unable to effectively compare distinct protein isoforms. To address this limitation, we present SpliceBind, a novel framework utilizing graph neural networks to predict druggability with isoform awareness.

Our approach not only enhances predictive performance—achieving an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.703 compared to 0.634 for P2Rank (p = 0.026)—but also tackles a deeper scientific inquiry: understanding the specific conditions under which structural methods are effective versus when they are inherently insufficient. Through a systematic analysis of six clinically validated variants across five mechanism classes, we have established a two-tier taxonomy of resistance.

This taxonomy distinguishes between structurally detectable changes, such as domain deletions (e.g., AR-V7, Delta = -18.39) and pocket disruptions, and mechanisms that remain fundamentally undetectable to pocket-centric approaches, such as allosteric changes (e.g., BRAF-p61). The latter represents a hard boundary that no amount of algorithmic refinement can overcome. Furthermore, our model’s learned embeddings successfully capture affinity-based resistance that geometry alone misses, as seen in ALK-L1196M (Delta_SB = -0.228 vs. Delta_P2Rank = -0.95), thereby partially bridging the gap between structural and biochemical data.

Evaluated on 229 kinase pockets across 25 families, SpliceBind demonstrates robust generalization to unseen families, achieving an AUROC of 0.761. This capability significantly impacts clinical workflows by allowing practitioners to immediately decide, upon identifying a splice variant, whether computational screening is adequate or if biochemical validation is strictly necessary. This distinction streamlines the path from variant discovery to therapeutic decision-making, substantially reducing time-to-treatment.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC