2026-04-12T00:20:16Z https://ipsj.ixsq.nii.ac.jp/oai

oai:ipsj.ixsq.nii.ac.jp:00194990 2025-01-19T23:15:33Z 1164:5352:9740:9741

グラフ畳み込みを用いたタンパク質予測立体構造の評価手法の開発 Protein prediction structure model quality assessment using Graph Convolution 佐藤, 倫 石田, 貴士 Rin, Sato Takashi, Ishida タンパク質の立体構造はタンパク質の機能に大きく関わり，創薬等の生命科学において重要な情報となる．実験的に立体構造を決定するのは時間的・金銭的にコストがかかるため，計算機を用いて立体構造を予測する手法が多く開発されてきた．で立体構造予測では，複数のテンプレート構造や複数の予測手法の組み合わせが用いられることが多く，そこで生成される予測立体構造の評価手法(Model Quality Assessment Program, MQAP)は有用な技術である．現在最も高精度なMQA手法の1つであるProQ3Dは，各残基ごとにクオリティ(局所スコア)し，その平均を取ることでタンパク質全体のクオリティ(大域スコア)を計算する．ここで局所スコアに機械学習アプローチを用いて大域スコアを予測することで精度の向上が期待されるが，タンパク質の残基の数は可変であるため，機械学習を導入することは困難である．この問題を解決するため本研究では，残基をノード，周辺残基との間にエッジとしたグラフ構造定義し，グラフ畳み込みとマルチタスク学習を用いて局所ラベルと大域ラベルを同時に学習する深層学習モデルを開発し，既存手法よりも高精度での予測に成功した． The three-dimensional structure of a protein is related to its function, and it is important in life science application such as drug discovery. Determination of three-dimensional structure is costly in terms of time and money, thus many methods for predicting three-dimensional structure using a computer have been developed. However, the accuracy is still insufficient. Thus, evaluation of the quality of a predicted model is required and such software is called model quality assessment program (MQAP). ProQ3D, which is currently one of the best MQA method, assesses the model quality of each residue (local score) using deep learning method. As results, the quality of whole protein structure (global score) was simply calculated as the mean value of local scores. Thus, if we can use a machine learning method to integrate those local scores for getting a global score, it may improve the accuracy compared with that by the mean value. However, it is difficult to use machine learning method because number of residues in a protein is not fixed. To deal with this problem, we developed a novel single model assessment method using graph convolution. We defined a graph structure on a protein, whose node is a residue and as edge means close residues. By using multi-task learning based on local and global score, proposed method achieved better accuracy than previous methods. 情報処理学会 2019-03-01 jpn technical report https://ipsj.ixsq.nii.ac.jp/records/194990 2188-8590 AA12055912 研究報告バイオ情報学（BIO） 2019-BIO-57 3 1 5 https://ipsj.ixsq.nii.ac.jp/record/194990/files/IPSJ-BIO19057003.pdf application/pdf 822.5 kB 2021-03-01