Abstract 2444: A deep learning framework for the reconstruction of gene expression profiles derived from formalin-fixed paraffin-embedded samples

Abstract Formalin-fixed paraffin-embedded (FFPE) tissue samples often suffer from RNA degradation due to fixation, embedding, and storage processes, affecting the integrity of transcripts for RNA sequencing (RNA-seq). Herein, we developed a novel deep learning model to reconstruct gene expression profiles from FFPE-derived RNA-seq data. We first compared the transcript integrity numbers (TINs) of two in-house datasets derived from paired fresh frozen (FF) and FFPE tissues, with different fixation durations and library preparation strategies. We found that tissue samples fixed for 24 hours and subjected to the rRNA depletion-based library preparation strategy exhibited higher TINs. Using RNA-seq data for 832 FFPE samples across seven cancer types, we identified genes with consistently low TINs on the genome-wide scale. Interestingly, the signature genes of immune cells are more prone to degradation, compared to other cell types such as tumor epithelial cells. To reconstruct gene expression profiles, we developed a deep learning framework, based on 9568 FF tumor samples across 28 cancer types in the TCGA database. The framework demonstrated remarkable performance on both simulated data (672 transcriptomes) and in-house cancer cohorts with RNA-seq data, including one FFPE and three FF-FFPE-matched cohorts. On the simulated data with noise levels ranging from 10% to 60%, the reconstructed data achieved consistently high correlation coefficients (> 0.85) with the ‘clean’ data without noise. For the FF-FFPE-matched datasets, compared to the original FFPE-derived data, the reconstructed gene expression profiles significantly improved the correlation with matched FF profiles, even for immune cell signature genes (all P < 0.05). In our application to cancer subtyping, the gene expression profiles reconstructed improved the classification accuracy from 67% to 92%. Furthermore, the reconstructed profiles recapitulated the subtype-specific biological properties lost in the FFPE-derived data and showed more significant association with patient survival. As the first deep learning model to rectify FFPE-derived RNA-seq data, our framework enables more efficient use of archived FFPE tissue samples that are widely available in the clinic. Citation Format: Lingli He, Kai Song, Ying Li, Celia Yujuan Dong, Cherry Y. Wong, Lin Qi, Xianrui Zhang, Yufeng Chen, Louis Vermeulen, Shentao Zhou, Simon Siu Ng, Xin Wang. A deep learning framework for the reconstruction of gene expression profiles derived from formalin-fixed paraffin-embedded samples [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2444.