Self-Supervised Learning in Genomic Data Analysis

Abstract

Genomic data analysis faces a fundamental asymmetry: sequencing costs have plummeted, generating petabytes of unlabelled DNA, RNA, and protein sequences, while functional annotation -- the labelled data that supervised learning requires -- remains expensive, slow, and incomplete. Only 1-2% of protein sequences in UniProt have experimentally verified function; fewer than 5% of non-coding genomic variants have characterised regulatory effects; and single-cell RNA-seq datasets contain millions of cell profiles but limited ground-truth cell-type annotations. Self-supervised learning (SSL) addresses this asymmetry by learning representations from the structure of unlabelled sequences themselves - masked language modelling over DNA and protein sequences, contrastive learning over single-cell expression profiles, and autoregressive generation of genomic sequences. We present the Genomic Self-Supervised Learning Assessment Framework (GSLAF), evaluating five SSL paradigms -- masked language models (MLMs), autoregressive sequence models, contrastive learning, variational autoencoders (VAEs), and multi-modal foundation models -- across four genomic tasks (protein function prediction, variant effect estimation, single-cell type annotation, and regulatory element classification). Our Genomic SSL Performance Score (GSPS) measures downstream task accuracy, representation quality, data efficiency, computational cost, and biological interpretability. Multi-modal foundation models achieve the highest GSPS (0.928) by integrating DNA sequence, RNA expression, protein structure, and epigenomic signals into unified representations, while masked language models achieve the highest data efficiency (0.960) for protein function prediction through transfer learning from pre-trained protein language models.