报告主题：Advancing Microbial Genomics Through Algorithmic Innovation and Machine Learning

主 讲 人：廖和睿 博士（麻省理工学院）

主 持 人：李 猛 教授

时 间：2025年12月30日（二）14：30

地 点：致知楼706

嘉宾简介：

廖和睿，麻省理工学院-诺和诺德人工智能博士后。2018年毕业于大连理工大学生物信息学专业，曾任职于华大基因，2024年获香港城市大学计算生物学博士学位，师从孙燕妮教授。研究方向为微生物基因组学与计算生物学，专注于开发高精度、可解释的计算工具用于微生物株水平分析、变异检测和疾病预测，结合传统算法与深度学习方法解决微生物组学中的关键问题。已发表第一作者/共同第一作者论文6篇，包括Genome Biology、Microbiome, Bioinformatics等领域知名期刊。所开发生信工具在GitHub获超80 stars，Bioconda累计下载超20,000次，被国际同行广泛使用。获MIT-Novo Nordisk AI博士后奖学金、香港科技300种子基金等荣誉。担任Nature、Nucleic Acids Research、Microbiome等顶级期刊审稿人。

报告摘要：

Microbial genomics plays an important role in modern biology and medicine, with transformative applications across multiple domains. For example, microbial strain identification is essential for tracking pathogen transmission and understanding functional diversity within species. Microbiome-based disease prediction enables early diagnosis and precision medicine. Accurate microbial SNV calling reveals adaptive evolution and intra-host diversity. Furthermore, antimicrobial resistance (AMR) prediction and bacterial colonization modeling are critical for clinical decision-making and infection control. Despite their importance, these applications face significant computational challenges stemming from genomic complexity, data heterogeneity, and limited computational frameworks.

This talk presents a suite of algorithmic and machine learning innovations addressing these challenges. For strain-level profiling, VirStrain employs greedy k-mer covering algorithm to distinguish highly similar viral genomes, demonstrating accuracy in SARS-CoV-2 variant identification. StrainScan extends this framework to bacteria through hierarchical tree-based indexing and informative k-mers, achieving 20% improvement in bacterial multi-strain detection. For microbiome-based disease prediction, GDmicro integrates graph convolutional networks with domain adaptation to overcome limited labeled data and cross-cohort variability, improving inflammatory bowel disease classification from 0.783 to 0.949 AUC. For variant detection, AccuSNV leverages deep learning to eliminate heuristic rules, enabling robust and accurate bacterial SNV calling. Finally, brief insights are provided into ongoing work on AI-based AMR prediction and a new hierarchical microbial foundation model for colonization prediction.