LIMPACAT : Multi-Omics Attention Transformer for Immune Prediction in Liver Cancer Using Whole-Slide Imaging

Abstract

This study introduces LIMPACAT (Liver Immune Microenvironment Prediction and Classification Attention Transformer), a framework that leverages whole-slide images (WSIs) to predict immune cell levels associated with prognosis in liver cancer. Since direct immune cell composition data is unavailable in the TCGA-LIHC dataset, we inferred immune cell levels using liver-specific single-cell RNA sequencing (scRNA-seq) data to simulate bulk RNA-seq profiles, allowing estimation of immune compositions within the tumor microenvironment. To ensure consistency, we tested three normalization methods--log normalization, canonical correlation analysis (CCA), and SCTransform--for scRNA-seq data integration and preprocessing, which facilitated reliable predictions of immune cell distributions in bulk RNA-seq. By applying this scRNA-seq-informed cell composition deconvolution model to real bulk RNA-seq data from liver cancer samples, we confirmed alignment in immune cell composition estimates between bulk RNA-seq and scRNA-seq. Using a multiple instance learning (MIL) framework with an attention transformer, LIMPACAT achieved approximately 80% accuracy in classifying immune cell levels relevant to patient prognosis. These findings highlight the feasibility of integrating WSIs and multi-omics data to enhance immune profiling and prognostic predictions in liver cancer research.