Fine-tuned DNA sequence language models for predicting gene expression in admixed population: A comparison with variant-based methods

Predicting gene expression from whole genome sequencing data is essential for elucidating regulatory mechanisms and interpreting functional consequences of non-coding variants to complex diseases. While traditional approaches rely on variant-based linear models, DNA sequence-based language models have gained increasing popularity for their ability to learn regulatory features directly from whole genome sequence data. Previous studies have evaluated models such as Enformer, but newer sequence-to-function models, including Borzoi and AlphaGenome, remain less well characterized for individual-level gene expression prediction, particularly in non-European and admixed populations. Here, we evaluated and fine-tuned Borzoi and AlphaGenome using personal genomes paired with whole-blood RNA-seq data from the Hispanic Community Health Study/Study of Latinos (HCHS/SOL). Fine-tuning substantially improved the performance of both sequence-based models compared with their pretrained versions; however, their performances remained similar to that of the variant-based linear models. Leveraging the ancestral diversity of HCHS/SOL, we found that Borzoi and AlphaGenome showed similar performance across Amerindian-, European-, and African-enriched groups. Fine-tuned Borzoi, fine-tuned AlphaGenome and elastic net performed modestly better when trained and evaluated in the same ancestry-enriched group. Finally, we explored strategies to improve prediction performance and found that incorporating predicted epigenetic features improved prediction near the promoter region. Significant progress remains needed before the full potential of sequence-based models is attained. Incorporating more biology-guided training approaches could be helpful to achieving better performance.

Authors

Institutions

Publication Details

Journal
bioRxiv (Cold Spring Harbor Laboratory)
Published
2026-11-18
DOI
https://doi.org/10.64898/2026.11.17.767377
Primary Topic
Genetic Associations and Epidemiology
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Fine-tuned DNA sequence language models for predicting gene expression in admixed population: A comparison with variant-based methods

Minoli A. Perera, Jiecong Lin, Tamar Sofer, Yu‐Jyun Huang et al.
3 links
bioRxiv (Cold Spring Harbor Laboratory)
Genetic Associations and Epidemiology
preprint

Fine-tuned DNA sequence language models for predicting gene expression in admixed population: A comparison with variant-based methods

Minoli A. Perera, Jiecong Lin, Tamar Sofer, Yu‐Jyun Huang, Anne E. Justice, Kari E. North, Xiaonan Liu, Junwei Lu, Xihong Lin
preprint en

Abstract

Predicting gene expression from whole genome sequencing data is essential for elucidating regulatory mechanisms and interpreting functional consequences of non-coding variants to complex diseases. While traditional approaches rely on variant-based linear models, DNA sequence-based language models have gained increasing popularity for their ability to learn regulatory features directly from whole genome sequence data. Previous studies have evaluated models such as Enformer, but newer sequence-to-function models, including Borzoi and AlphaGenome, remain less well characterized for individual-level gene expression prediction, particularly in non-European and admixed populations. Here, we evaluated and fine-tuned Borzoi and AlphaGenome using personal genomes paired with whole-blood RNA-seq data from the Hispanic Community Health Study/Study of Latinos (HCHS/SOL). Fine-tuning substantially improved the performance of both sequence-based models compared with their pretrained versions; however, their performances remained similar to that of the variant-based linear models. Leveraging the ancestral diversity of HCHS/SOL, we found that Borzoi and AlphaGenome showed similar performance across Amerindian-, European-, and African-enriched groups. Fine-tuned Borzoi, fine-tuned AlphaGenome and elastic net performed modestly better when trained and evaluated in the same ancestry-enriched group. Finally, we explored strategies to improve prediction performance and found that incorporating predicted epigenetic features improved prediction near the promoter region. Significant progress remains needed before the full potential of sequence-based models is attained. Incorporating more biology-guided training approaches could be helpful to achieving better performance.

bioRxiv (Cold Spring Harbor Laboratory)
Northwestern University (US), University of North Carolina at Chapel Hill (US), Beth Israel Deaconess Medical Center (US), Harvard University (US)
National Institutes of Health, National Institute on Aging
Good health and well-being
Genetic Associations and Epidemiology
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon