[1P-47*] トランスフォーマー深層学習モデルへの効率的な進化情報の取り込みによる高精度な変異導入効果予測
In protein engineering, amino acid mutations are introduced to alter functional proteins' properties, including enzymes and antibodies. While directed evolution approaches are prevalently used for improved functions, the experiments are often costly due to multiple library construction and selection rounds. Recently, machine learning techniques, including deep neural networks, are utilized to predict variant effects and to select variants for experimental verification. Several studies have shown that variant effect prediction accuracy can be improved by incorporating evolutionary information of an engineering target protein. However, there is no established protocol to incorporate evolutionary information into Transformer, a type of deep neural networks shown to achieve the best performance in natural language processing applications. This study proposes novel protocols to incorporate evolutionary information into Transformer-based variant effect predictors, mainly focusing on how to handle partial homology arising from multi-domain proteins. We show that the proposed protocols significantly improve prediction accuracy compared to Transformer without evolutionary information.