Advancing Data Usability, Activity Modeling, and Stability Optimization in Computational Enzyme Design

Abstract

A grand challenge of computational biology is to computationally design, in a single pass, a protein sequence that catalyzes an arbitrary chemical reaction at a high rate under specified conditions [1]. This work details advances in three essential areas on the path to that goal: data quality, activity prediction and modeling, and stability optimization. The structure and implementation of the Allotrope Simple Model (a FAIR data format for many scientific instruments) was examined in [2], setting the stage for training deep learning models on high-quality experimental datasets from diverse sources. In [3], the limits of physics-based and deep learning tools to rank the relative activities of variants of the hydrolase LipA towards native and non-native substrates were explored. In [4], an expert-guided inverse folding approach was used to raise the melting temperature of the luciferase NanoLuc by up to 7.2 ◦C while maintaining activity, without relying on guidance from multi-sequence alignments. Taken together, these publications describe modest progress towards zero-shot enzyme design and outline a direction for future work.