Fine-tuning sequence-to-expression models on personal genome and transcriptome data
- 1Department of Electrical Engineering and Computer Sciences, University of California Berkeley, Berkeley, CA, USA
- 2Center for Computational Biology, University of California Berkeley, Berkeley, CA, USA
- 3Chan Zuckerberg Biohub, San Francisco, CA, USA
Abstract
BACKGROUND: Genomic sequence-to-expression deep learning models, which are trained to predict gene expression and other molecular phenotypes across the reference genome, have recently been shown to have poor out-of-the-box performance in predicting gene expression variation across individuals based on their personal genome sequences. RESULTS: Here, we explore whether additional training (fine-tuning) on paired personal genome and transcriptome data improves the performance of such sequence-to-expression models. Using Enformer as a representative pre-trained model, we explore various fine-tuning strategies. Our results show that fine-tuning improves expression predictions on held-out individuals, including from held-out populations, for genes seen during fine-tuning, with comparable performance to variant-based linear models commonly used in transcriptome-wide association studies. However, fine-tuning does not improve model generalizability to held-out genes, which contain sequences and variants unseen during fine-tuning. CONCLUSIONS: Including individual-level genetic variation and paired expression data during the training of sequence-to-expression models improves their understanding of seen variants, enabling their application to held-out individuals. However, this strategy does not improve generalizability to unseen genes, highlighting a remaining open challenge in the field.
Related
- Comparative transcriptome analysis of galls induced by four closely related aphid species using the draft genome of the host plant (Hamamelis japonica) reveals gene expression patterns related to gall diversification
- gRely: Relyability for genome trained sequence-to-expression models
- A scalable approach to investigating sequence-to-function predictions from personal genomes
- Cross-subject EEG driving fatigue detection: early personal spectral reference and causal long-sequence residual encoding
- Skin transcriptome comparison of Changthangi Pashmina and Sirohi goats reveals fibre-barrier and skin remodeling expression signatures
BibTeX
@article{rastogi2026finetuning,
title = {{Fine-tuning sequence-to-expression models on personal genome and transcriptome data}},
author = {Ruchir Rastogi and Aniketh Janardhan Reddy and Ryan Chung and Nilah M. Ioannidis},
journal = {Genome biology},
year = {2026},
month = may,
doi = {10.1186/s13059-026-04091-1},
url = {https://doi.org/10.1186/s13059-026-04091-1},
note = {CC BY},
}
RIS
TY - JOUR
TI - Fine-tuning sequence-to-expression models on personal genome and transcriptome data
AU - Ruchir Rastogi
AU - Aniketh Janardhan Reddy
AU - Ryan Chung
AU - Nilah M. Ioannidis
PY - 2026
DA - 2026/05/25
JO - Genome biology
DO - 10.1186/s13059-026-04091-1
UR - https://doi.org/10.1186/s13059-026-04091-1
AB - BACKGROUND: Genomic sequence-to-expression deep learning models, which are trained to predict gene expression and other molecular phenotypes across the reference genome, have recently been shown to have poor out-of-the-box performance in predicting gene expression variation across individuals based on their personal genome sequences. RESULTS: Here, we explore whether additional training (fine-tuning) on paired personal genome and transcriptome data improves the performance of such sequence-to-expression models. Using Enformer as a representative pre-trained model, we explore various fine-tuning strategies. Our results show that fine-tuning improves expression predictions on held-out individuals, including from held-out populations, for genes seen during fine-tuning, with comparable performance to variant-based linear models commonly used in transcriptome-wide association studies. However, fine-tuning does not improve model generalizability to held-out genes, which contain sequences and variants unseen during fine-tuning. CONCLUSIONS: Including individual-level genetic variation and paired expression data during the training of sequence-to-expression models improves their understanding of seen variants, enabling their application to held-out individuals. However, this strategy does not improve generalizability to unseen genes, highlighting a remaining open challenge in the field.
SN - 1465-6906
SN - 1465-6914
SN - 1474-7596
SN - 1474-760X
ER -