Precision Medicine group Progress Report

Abstract

This progress report comprises three parts. In part 1, we report a web-based graphical user interface (GUI) built on LogitDA, called ImmunoResponse Predictor. We introduced a novel objective response rate (ORR)-informed cosine distance metric (iCosineDist) that quantifies the applicability of LogitDA to future mUC/mRCC datasets. Using independent datasets from the PCD4989g trial, LogitDA achieved prediction AUCs of 0.75 (mUC) and 0.72 (mRCC), with corresponding accuracies of 0.71 and 0.83, and percentages of support of 81% and 72%, respectively. The ImmunoResponse Predictor enhances the clinical applicability of LogitDA. We have also obtained some results on integrating bulk RNA-seq and single-cell RNA-seq data, to identify gene signatures associated with immune checkpoint inhibitor (ICI) response in mUC. 

In part 2, for early risk prediction of pancreatic cancer, we developed a model using serum samples from Taiwanese participants, in which metabolite concentrations were carefully quantified. The final model incorporates 11 key metabolites and achieves 95% accuracy in distinguishing high-risk individuals from normal controls. When applied to stage I and II PDAC patients, the model yielded encouraging results, suggesting that it effectively captures essential metabolic alterations associated with tumorigenesis. We are currently investigating and validating the underlying biological mechanisms driving these associations. The next step is to apply and adapt this model to the UK Biobank (UKBB) cohort to evaluate its predictive performance for cancer risk.

In part 3, we present an interpretable deep learning model that integrates genotype, imaging, and phenotype data through a mediation-based framework to reveal genotype–image–phenotype associations and enhance disease diagnosis.The method (1) extracts imaging features strongly linked to genotypes, (2) identifies phenotype associations through these features, and (3) integrates direct genotype–phenotype and image–phenotype associations to capture additional signals missed by mediation. Applied to Alzheimer’s disease (AD) diagnosis and validated on the PASCAL VOC, a synthetic toy dataset, and the ADNI dataset, the integrated model achieves superior performance over single-modality or partially integrated approaches.

線上視訊請點選連結