Multi-Omic Integration by Machine Learning (MIMaL) Reveals Protein-Metabolite Connections and New Gene Functions

Metabolomics and proteomics generate large, complex datasets that reflect the state of a biological system. Multi-omics is the integration of these disparate methods and data to gain a clearer picture of the biological state. Multi-omic studies of the proteome and metabolome are becoming more common as mass spectrometry technology continues to be democratized. However, knowledge extraction through integration of these data remains challenging. Here we show that connections between these omic layers can be discovered through a combination of machine learning and model interpretation. We find that SHAP values connecting proteins to metabolites are valid experimentally, and reveal also largely new connections. Further, clustering the magnitudes of protein control over all metabolites enabled prediction of gene five gene functions, each of which was validated experimentally. We accurately predicted that two uncharacterized genes in yeast modulate mitochondrial translation, <em>YJR120W</em> and <em>YLD157C</em>.We also predict and validate functions for several incompletely characterized genes, including <em>SDH9</em>, <em>ISC1</em>, and <em>FMP52</em>. Our work demonstrates that multi-omic analysis with machine learning (MIMaL) is a new lens that reveals new insight from multi-omic data that would not be possible using any omic layer alone.

代谢组学（Metabolomics）与蛋白质组学（Proteomics）能够生成反映生物系统状态的大型复杂数据集。多组学（Multi-omics）指整合这些不同的研究方法与数据集，以更清晰地刻画生物系统的状态。随着质谱技术持续普及，针对蛋白质组与代谢组的多组学研究正愈发常见。然而，通过整合此类数据集开展知识提取仍颇具挑战。本研究表明，可通过机器学习与模型可解释性相结合的手段，挖掘不同组学层级间的关联。我们发现，连接蛋白质与代谢物的SHAP值具备实验有效性，且揭示了大量全新的关联。此外，通过对蛋白质调控所有代谢物的强度进行聚类分析，我们得以预测五种基因的功能，且所有功能均通过实验得到了验证。我们精准预测出酵母中两个未表征基因——YJR120W与YLD157C——可调控线粒体翻译过程。此外，我们还对多个表征不完全的基因进行了功能预测与实验验证，包括SDH9、ISC1及FMP52。本研究证明，结合机器学习的多组学分析（Multi-omic analysis with machine learning, MIMaL）是一种全新的研究视角，可从多组学数据中挖掘出单一组学层级分析无法获得的全新认知。