XPro-Design: Rational Protein Engineering Framework Using Explainable AI

ICLR 2026 Conference Submission19715 Authors

Keywords: protein, engineering, thermostability, biotech, enzyme, optimization

TL;DR: XPro-Design is a novel explainable AI framework towards structure aware protein engineering

Abstract: Protein engineering seeks to rationally tailor proteins to achieve specific structural and functional objectives. These objectives encompass enhancing catalytic efficiency, modifying substrate specificity, improving binding affinity, reducing immunogenicity, and increasing stability under adverse conditions. A major bottleneck is protein instability, as elevated temperatures often drive degradation and compromise activity. Developing thermostable proteins is therefore a key objective in engineering efforts. Here, we present XPro-Design, an explainable AI driven framework for protein optimization that integrates amino acid-level explanations of functional impact into generative modeling. Our method captures epistatic interactions and the mutational landscape by training a low-rank matrix, which biases the generative model toward high-scoring regions of sequence space. This enables targeted generation of candidate variants optimized for thermostability, while remaining extensible to other objectives. XPro-Design further uses distribution tempering and annealing to effectively balance exploration vs exploitation without compromising on structural integrity. We demonstrate rational, causality driven design of protein variants with melting temperatures nearly 2x that of their wild-type counterparts, while preserving binding pocket integrity and domain architecture. Moreover, engineered variants show up to 38% lower folding free energy relative to wild-type, indicating significantly enhanced thermodynamic stability. XPro-Design establishes a generalizable strategy for explainable and controllable protein design, enabling multi-objective optimization beyond thermostability.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 19715