うつ病とその症状の生物学的背景を解明するための機械学習アルゴリズムの開発

概要

令和4年度

京都大学化学研究所　スーパーコンピュータシステム　利用報告書

うつ病とその症状の生物学的背景を解明するための機械学習アルゴリズムの開発

The development of machine learning algorithms to decipher the biological background
of major depression and its symptoms
Bioinformatics Center (Mamitsuka laboratory), Institute for Chemical Research, Kyoto
University, Peter Petschner
研究成果概要

During FY2022/2023 we used a large, UK Biobank derived dataset to test a machine
learning model and identify relevant genetic factors behind depression. The dataset
contained 168,096 individuals, 2 environmental (phenotypic variables) factors,
depression score as output and 3,792,532 genetic factors, in the form of so called singlenucleotide polymorphisms (SNPs).
Our original network architecture followed a feed-forward neural network with
extensions to increase interpretability. Run of this model and its extension using side
information about the ca. 3.5M SNPs needs the utilization of the SDF nodes of the
supercomputer. After optimization steps data-loading and calculation times take about
7-9 minutes per 1000 individuals, memory use is 900GB-1.3TB. Thus, obtaining results
still requires around 3 months of run time for a single run. We started up to December
more than 17 runs, but many was killed by job manager. After we found solution with a
system engineer jobs were running smoothly. The shutdown at the beginning of January
caused small delays, due to modification of the code to allow for restarting of a job.
From a scientific perspective preliminary runs showed an average increase of
21.56%, if SNPs were added to sex and age, in the prediction performance of depression
and aligned well with previous SNP-based heritability estimates of depression. The
current interpretation method yielded 22 genes with consistent positive signal for
depression. Among these, 3 (13.6%) showed overlap with the most recent genome-wide
association study, giving 19 newly discovered genes.
The above indicate the potential of the new method, however, additional runs
are necessary. Therefore, we continue the research and plan to extensively use the SC for
the same purpose in the next fiscal year as well. Publications are expected in the
future. ...