MLDE

MLDE applies machine learning to predict fitness across combinatorial protein-variant libraries to guide directed evolution and efficiently identify high-fitness sequences such as variants of protein G domain B1 (GB1).

Key Features:

• Path-Independent Optimization: Overcomes path-dependent constraints of single-step greedy optimization by enabling in silico screening of full combinatorial libraries to capture cooperative mutation effects and mutation-order independence.
• Focused Training Protocol (ftMLDE): Implements a focused training protocol (ftMLDE) that reached the global maximum up to 92% of the time with a screening burden of 470 variants on an epistatic, hole-filled four-site combinatorial fitness landscape of protein G domain B1 (GB1).
• Efficiency and Effectiveness: Compared to minimal-screening-burden single-step greedy optimization (which reached the global maximum 1.2% of the time on the GB1 landscape), ftMLDE matched a minimal screening burden of 80 total variants and achieved the global optimum up to 9.6% of the time with a 49% higher expected maximum fitness.
• Design Considerations: Evaluated different encoding strategies, integrated new models and training procedures tailored to protein engineering, and implemented training-set design strategies that avoid information-poor low-fitness "holes".

Scientific Applications:

• Protein variant optimization: Accelerates identification and optimization of high-performance protein variants in directed evolution experiments.
• Screening reduction: Reduces experimental screening burden and associated costs by prioritizing variants for testing.
• Epistasis analysis: Enhances characterization and understanding of epistatic interactions within combinatorial mutant libraries.

Methodology:

Uses machine learning techniques to simulate and predict fitness outcomes and to perform in silico screening of combinatorial variant libraries; implements a focused training protocol (ftMLDE) and training-set selection strategies that avoid low-fitness "holes".