Unveiling Phenotype–Genotype Interplay with Deep Learning Foundation Models for scRNA-seq: A Quantitative Perspective

Abstract

Foundation models have emerged as powerful tools for analyzing single-cell RNA sequencing (scRNA-seq) data, leveraging large-scale pretraining to capture complex gene expression patterns. However, a comprehensive quantitative framework for understanding the interplay between phenotypes and genotypes remains underdeveloped. Such a framework is critical not only for validating model performance but also for uncovering previously unrecognized biological relationships. In this work, we present both traditional and deep learning-based quantitative analysis pipelines for PolyGene [1], a transformer-based scRNA-seq foundation model, aimed at disentangling the complex phenotype–genotype relationship. First, we implement a top-k classification and entropy evaluation pipeline to serve as a primary validation framework. Our results demonstrate that the pretrained PolyGene [1] is robust in top-k classification metrics and provides meaningful insights into the entropy landscape of human cells across different life stages. Second, we propose a novel deep learning gradientbased gene selection method designed to address limitations in traditional feature selection approaches, such as poor scalability and sensitivity to heterogeneity in high-dimensional data. Through empirical evaluations on benchmark scRNA-seq datasets, we show that our method enhances model interpretability and improves downstream performance, offering a more scalable and biologically relevant alternative to existing techniques. Overall, this work introduces a set of quantitative analysis tools that fill a critical gap in evaluating and interpreting scRNA-seq foundation models, contributing to a deeper understanding of the genotype–phenotype interplay through modern deep learning techniques.