Background

After the completion of the Human Genome Project, the greatest barrier faced by molecular biologists was the functional interpretation of the noncoding region, which accounts for 98% of the genome. The exon regions, which encode proteins, allow for relatively easy prediction of variant pathogenicity by directly comparing amino acid sequence changes. On the other hand, the interpretation system for promoters, enhancers, cleavage junction sites, and non-coding RNA regions responsible for transcriptional regulation was extremely opaque, despite the fact that a single base change could drastically alter disease susceptibility.

More than 90% of single nucleotide polymorphisms (SNPs) reported to be associated with diseases through Genome-Wide Association Studies (GWAS) are concentrated in these noncoding regions. However, existing bioinformatics tools have shown limitations in inferring human-specific regulatory variants or cell-line-specific epigenomic changes due to excessive reliance on multi-species sequence conservation scores. Although high-throughput mass screening experiments such as Massively Parallel Reporter Assays (MPRA) have been developed, verifying hundreds of millions of human variant combinations is accompanied by astronomical costs and technical constraints. This created an urgent need for computational models capable of quantifying the impact of noncoding variants on gene expression and chromatin structure using only sequence information.

Key Findings

In recent years, genomic deep learning technology has evolved beyond simple local Convolutional Neural Network (CNN)-based models into neural network architectures that capture long-range interactions spanning tens to hundreds of thousands of base pairs. If early DeepSEA or Basset models operated at the level of determining transcription factor binding or DNA nuclease sensitivity within sequences of a few thousand base pairs, recent models like Enformer and Borzoi combine Transformer architectures to compute even three-dimensional physical contacts between promoters and enhancers occurring over distances of several hundred thousand base pairs. was included in the scope.

These advanced models predict CAGE (Cap Analysis of Gene Expression) signals, RNA transcription levels, and histone modification (ChIP-seq) patterns for human cell lines with single-nucleotide accuracy. We have reached a stage where in silico mutagenesis can simulate how much the expression of downstream target genes increases or decreases when an artificial mutation is introduced at a specific noncoding position. The Sei model, which classifies noncoding variants into dozens of sequence classes based on chromatin state to prioritize pathogenic variants in a disease-specific manner, has also yielded such results.

However, the asymmetry of the model structure did not directly lead to full prediction accuracy. When evaluating rare variants, limitations remain in overfitting to the reference genome sequence, which leads to underestimating allele-specific expression differences, as well as in the ability to isolate subtle regulatory effects that occur only in specific tissues. The bias in the functional genomics datasets used for training is also cited as a key factor causing performance gaps between tissues.

Significance and Outlook

Genomic deep learning is forming a major turning point in elevating noncoding variants into the realms of precision medicine and drug target discovery. This provides the foundation to dramatically improve the diagnosis rate of rare, intractable genetic diseases, which previously remained as 'Variants of Uncertain Significance (VUS)' because their functional meaning could not be identified even as genomic decoding data from tens of thousands of patients accumulated. In cancer genomics, it is also beginning to serve as a key tool for uncovering the mechanisms of abnormal overexpression of oncogenes caused by somatic noncoding mutations.

The future challenge is to bridge the gap between computer simulation results and actual in vivo molecular dynamics. It is required to build Genomic Foundation Models that reflect genetic polymorphism across diverse ethnic groups and to develop training techniques that directly incorporate single-cell transcriptome and multi-omics empirical data into the loss function. To ensure reliability in clinical settings, model interpretability must be enhanced to clearly trace back the specific patterns of transcription factor motif disruption that cause diseases.