NeuroFold: A Multimodal Approach to Generating Novel Protein Variantsin silico

The generation of high-performance enzyme variants with desired physicochemical and functional properties presents a formidable challenge in the field of protein engineering. Existing in silico design methods are limited by inadequate training data, insufficient diversity within datasets, and suboptimal sampling techniques. Here, we introduce a novel approach that addresses these limitations and significantly improves the efficiency of generating functional enzyme variants. Using a multimodal approach, NeuroFold can leverage sequence, structural, and homology data during both sampling and discrimination phases, thereby enabling more diverse and informed sampling of the sequence space. Our model demonstrated a 40-fold increase in Spearman rank correlation as compared to large language models (LLMs) such as ESM-1v and empowers the rapid creation of high-quality enzyme variants, such as the β-lactamase variants generated by NeuroFold in this study, which demonstrated increased thermostability and varying levels of activity. This pipeline represents a promising advancement in the field of enzyme engineering, offering a valuable tool for the development of novel enzymes with enhanced performance and desired chemical properties.