Recently, Huiyan Sun's team has made research progress at the intersection of causal learning and biomedicine, with two papers published in the journalBriefings in Bioinformatics.
Research Work 1
Siwei Wu, a 2022-cohort master's student supervised by Research Fellow Huiyan Sun, has proposed CPCG, a new method for identifying key cancer prognostic genes based on causal feature selection.
Title: Identifying cancer prognosis genes through causal learning
Journal:Briefings in Bioinformatics
Link:https://doi.org/10.1093/bib/bbae721
First author: Siwei Wu (main research interests: bioinformatics, causal inference, etc. He co-authored the bookCausal Inference: Principles, Analysis and Practiceand will continue to pursue a doctoral degree.)
Overview
Accurate identification of key cancer prognostic genes is essential for guiding tumor treatment. Owing to confounding bias and selection bias, biomarkers identified by conventional feature selection methods based on generalized correlation often have unstable predictive performance. In addition, clinical data are frequently characterized by small sample sizes and abundant censored data, further aggravating the challenge of identifying effective biomarkers; causal features, by contrast, offer significant advantages in these respects thanks to their stability. To address these issues, this study proposes a new computational framework, CPCG, which first evaluates the impact of gene expression on patient prognosis by integrating parametric and semi-parametric proportional hazards models to determine a set of candidate genes; it then constructs, through causal inference, a causal skeleton of the relationships between candidate genes and patient overall survival; and finally takes the first-order neighbor genes directly related to overall survival as the core genes for cancer prognosis. Experimental results show that CPCG can significantly eliminate spurious associations between genes and patient prognosis, and directly identify from high-dimensional transcriptomic data a compact, reliable, interpretable, generalizable and robust set of key cancer prognostic genes. CPCG also provides a reference for robust biomarker identification in other medical tasks.
Research Work 2
Haolong Zeng, a 2022-cohort master's student supervised by Research Fellow Huiyan Sun, has proposed ICGI, a new framework for cancer gene identification that integrates large language models with causal inference.
Title: Cancer gene identification through integrating causal prompting large language model with omics data-driven causal inference
Journal:Briefings in Bioinformatics
Link:https://doi.org/10.1093/bib/bbaf113
First author: Haolong Zeng (main research interests: causal machine learning, large language models, etc. He is an intern in an AI algorithm role at Tencent and will join Huawei as an AI engineer on the Ascend team. As a core contributor, he participated in the GitHub LLM Cookbook project, which has more than 16,000 stars.)
Overview
Identifying cancer genes from a multi-omics perspective is essential for understanding cancer mechanisms and improving treatment strategies. Because of issues such as ignored confounders, selection bias and the black-box nature of models, cancer genes identified by deep learning methods are prone to spurious, biased and unexplainable predictions. Taking into account both the emergent capabilities of LLMs and the advantages of causal inference, this study proposes ICGI, a new framework for cancer gene identification that integrates large language models with causal inference. Through chain-of-thought prompting and retrieval-augmented generation, combined with causal prompting and causal learning mechanisms, ICGI can identify cancer genes from variations in the genome, transcriptome and even other omics, distinguish cancer from normal samples, and provide explanations of biological mechanisms, offering a reference for tapping the potential of LLMs in biomedical prediction tasks. In addition, this study developed and deployed a web application interface (https://huggingface.co/spaces/ICGI/ICGI) to promote the dissemination and application of this work.
