Machine learning‐based genetic evolution of antitumor proteins containing unnatural amino acids by integrating chemometric modeling and cytotoxicity analysis

Antitumor proteins (ATPs) are small oligoproteins or peptides that have been recognized as new and promising therapeutics against a variety of human tumors and cancers. In order to extend the structural diversity space of ATPs, the unnatural amino acids were incorporated into naturally occurring ATPs by using a chemometrics‐based genetic evolution strategy. Based on hundreds of ATPs derived from animals, plant and microbes statistical regression models were developed, optimized, and validated with a systematic combination of 5 widely used machine learning methods and 3 sophisticated unnatural amino acid descriptors. The best regression predictor was employed to guide genetic evolution of a large oligoprotein population. In the evolution procedure, a number of unnatural amino acids with desired physicochemical properties were introduced, resulting in an evolution‐improved population, from which few oligoprotein candidates with top scores, containing 1 to 3 unnatural amino acids, and having diverse structures were successfully prepared, and their antitumor potency against 2 cancer cell lines was analyzed with biological assays. It was found that the high‐activity ATPs are preferentially structured in partial α‐helix or β‐sheet with an alternative sequence pattern of polar, charged, and hydrophobic amino acids, while the intrinsically disordering oligoproteins usually have low or no antitumor activity against tested cancer cell lines. A machine learning‐based genetic evolution strategy is used to extend the structural diversity space to improve the biological activity of antitumor proteins by incorporating unnatural amino acids. Results show that high‐activity antitumor proteins are preferentially structured in partial α‐helix or β‐sheet with an alternative sequence pattern of polar, charged, and hydrophobic amino acids.