学术报告：Transfer Learning for Disease-Specific Molecular Prediction

报告时间：8月31日（星期一）上午10:00-11:00

报告地点：学院南路校区，主教107研讨室

主持人：梁超教授

报告人：龙泉，University of Calgary，教授

报告摘要：Large-scale representation learning can capture generalizable structure, but domain-specific datasets are often much smaller. Transfer learning addresses this mismatch by reusing representations learned from large reference datasets and adapting them to related target tasks with limited samples. From a statistical perspective, this reduces the effective complexity of the target prediction problem by constraining estimation around an informative pretrained representation rather than learning the full mapping from scratch.

We illustrate this idea with two applications built on Enformer, a large-scale model built on biological genomics and transcriptomic data. First, we adapt pretrained sequence representations to disease-specific transcription-factor prediction in breast and prostate cancer, improving tissue-specific regulatory prediction and downstream variant prioritization. Second, we develop TL-Prot, which transfers Enformer representations to the prediction of tissue-specific protein abundance using biologically constrained spatial aggregation. Together, these examples show how transfer learning can support accurate molecular prediction in small-sample, high-dimensional settings by leveraging information learned from large-scale reference data.

报告人简介：Quan Long is a Professor at the University of Calgary. His training spans mathematics, computer science, and computational biology, with research experience across China, the UK, Austria, USA and Canada. His research focuses on statistical genetics and high-dimensional data, particularly focusing on small-sample machine learning. His work has been cited more than 40,000 times. In this talk, he will discuss how to carry out small-sample based disease-specific molecular prediction using transfer learning.

撰稿人：刘洁

审稿人：邓露