Deep generative models for analysis and engineering of functional proteins

Doctoral thesis, 2025

Proteins are essential biological molecules that sustain life through diverse functions, fromstructural support to catalyzing biochemical reactions. Their catalytic efficiency makes theminvaluable for industrial applications, where they often require optimization to function underspecific conditions. While experimental and computational approaches have made progress inprotein engineering, no universal method exists due to the complexity of protein structure andfunction. Recent advances in machine learning offer new possibilities by leveraging vastprotein sequence data. However, key challenges remain, including the limited availability anduneven distribution of high-quality labels describing essential properties like enzymaticactivity and thermal stability. Addressing these issues is critical for developing models capableof accurate trait selection. My work focuses on two key steps in protein engineering:diversification and selection. To improve selection, deep learning models were developed usingtransfer learning, data augmentation, and protein language models (pLMs) to predict physicaland functional properties such as melting temperature, enzymatic temperature, proteinabundance, and in vitro activity. These models not only enable precise trait selection but alsoprovide insights into the relationships between sequence, thermal adaptation, andconformational stability. For diversification, a deep generative model was created to capturenatural sequence diversity and extend it to generate novel variant libraries across proteinfamilies. This approach prioritizes functional sequences and allows for targeted engineering ofproteins with enhanced properties. Moving beyond general sequence generation, a frameworkwas developed to create variant pools optimized for specific traits, such as increased thermalstability. By integrating these advancements, we engineered functional protein variants fromdiverse wild-type sequences, achieving up to a 36C increase in melting temperature. This workhighlights the potential of generative machine learning to refine and accelerate the proteinengineering cycle, paving the way for more efficient and scalable biotechnologicalapplications.