Deep learning on genome sequence to identify epistatic effects in complex diseases

2022 Fiscal Year Final Research Report

• Project/Area Number: 20K15773
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 43050:Genome biology-related
• Research Institution: The University of Tokyo
• Principal Investigator: Koido Masaru 東京大学, 大学院新領域創成科学研究科, 助教 (40787561)
• Project Period (FY): 2020-04-01 – 2023-03-31
• Keywords: ディープラーニング / 機械学習 / 遺伝子発現制御 / エピスタシス

Outline of Final Research Achievements

In this study, I demonstrated that the machine learning method for DNA sequence patterns, MENTR, accurately predicts causal variants of transcriptional regulation. MENTR identified the causal variants and their target transcripts (or transcriptional regulations) for diseases such as asthma, atopic dermatitis, and ossification of the posterior longitudinal ligament. Leveraging game theory revealed that MENTR utilizes distant nonlinear effects in its predictions, suggesting the learning of epistasis effects for transcription. Revising the deep learning model in MENTR led to an 80% reduction in model parameters at the expense of a 5% accuracy trade-off.

Free Research Field

ゲノミクス

Academic Significance and Societal Importance of the Research Achievements

MENTRの原因多型の予測に関する精密さ（特に真陰性予測能の高さ）は多型の組み合わせ効果（エピスタシス効果）を検証するための必須の特性である。本研究でエピスタシス効果を自ずと学習していることが示唆されたMENTRとその軽量モデルの活用により、大規模ゲノム解析から見出される疾患感受性多型の再解釈が進展し、エピスタシス効果を含む新たな生物学的知見の発見が期待される。