Abstract
<jats:title>Abstract</jats:title> <jats:p> Protein-ligand affinity (PLA) prediction is central to AI-driven drug discovery, but precise interaction-based methods require costly conformation preparation and data encoding, limiting their throughput. To reconcile accuracy with efficiency, we first investigate whether pre-trained molecular representation models can replace complex encoders. A unified and diverse assessment of sequence-, graph-, and image-based representations reveals both strong overall performance and family-wise variability, delivering the first practical guidance for encoder selection in PLA tasks. Next, to achieve high computational efficiency without sacrificing expressiveness, we adopt the mixture-of-experts (MoE) strategy from large language models. Systematic ablation studies uncover key design principles for deploying MoE in molecular prediction. The resulting model, HydrAffinity, is an interaction-free, dynamic sparse model that uses pre-trained encoders and MoE for parameter-efficient learning. It outperforms all interaction-free methods and matches state-of-the-art interaction-based methods on CASF-2016. Routing analysis confirms that MoE develops distinct, family-specific activation patterns, providing interpretable evidence of dynamic parameterization across protein classes. Zero-shot tests on DUDE-Z and LIT-PCBA further show strong EF <jats:sup>5</jats:sup> <jats:sup>%</jats:sup> performance, making HydrAffinity a practical, scalable solution acting as an effective early-stage pre-filter. </jats:p> <jats:sec> <jats:title>Graphic abstract</jats:title> <jats:fig id="ufig1" position="float" orientation="portrait" fig-type="figure"> <jats:graphic xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="740495v1_ufig1" position="float" orientation="portrait"/> </jats:fig> </jats:sec>