A scalable approach to investigating sequence-to-function predictions from personal genomes
4 affiliations
- 1Paul G. Allen School of Computer Science and Engineering, University of Washington, Seattle, WA, USA
- 2Department of Computational and Systems Biology, University of Pittsburgh, Pittsburgh, PA, USA
- 3Genentech, South San Francisco, CA, USA. [email protected]
- 4Paul G. Allen School of Computer Science and Engineering, University of Washington, Seattle, WA, USA. [email protected]
Abstract
A key promise of sequence-to-functon(S2F) models is their ability to evaluate arbitrary sequence inputs, providing a robust framework for understanding genotype-phenotype relationships. However, despite strong performance across genomic loci, S2F models struggle with predicting inter-individual variation in gene expression. Training a model to make genotype-dependent predictions at a single locus—an approach we call personal genome training—offers a potential solution. We introduce SAGE-net, a scalable framework and software package for training and evaluating S2F models using personal genomes. Leveraging its scalability, we conduct extensive experiments on model and training hyperparameters, demonstrating that training on personal genomes improves predictions for held-out individuals. However, the model achieves this by identifying predictive variants rather than learning a cis-regulatory grammar that generalizes across loci. This failure to generalize persists across a range of hyperparameter settings. These findings highlight the need for further exploration to unlock the full potential of S2F models in decoding the regulatory grammar of personal genomes. Scalable software and infrastructure development will be critical to this progress.
Related
- Long‐Read Reference Transcriptomes for Five Species of Rosaceae Tribe Spiraeeae: A Resource for Investigating Sequence Divergence, Species‐Associated Expression Variation, and Leaf‐Developmental Candidate Genes
- Rational design of disordered proteins for sequence–function investigation
- Extracting Latent Sequence-Function Information from Directed Evolution to Design Improved C-H Azidases
- Gaugefixer: overcoming parameter non-identifiability in models of sequence-function relationships.
- A multispecies foundation model for genome function prediction, gene annotation and sequence design
BibTeX
@article{spiro2026scalable,
title = {{A scalable approach to investigating sequence-to-function predictions from personal genomes}},
author = {Anna E. Spiro and Xinming Tu and Yilun Sheng and Alexander Sasse and Rezwan Hosseini and Maria D. Chikina and Sara Mostafavi},
journal = {Nature Methods},
year = {2026},
month = jun,
doi = {10.1038/s41592-026-03124-8},
url = {https://doi.org/10.1038/s41592-026-03124-8},
note = {CC BY},
}
RIS
TY - JOUR
TI - A scalable approach to investigating sequence-to-function predictions from personal genomes
AU - Anna E. Spiro
AU - Xinming Tu
AU - Yilun Sheng
AU - Alexander Sasse
AU - Rezwan Hosseini
AU - Maria D. Chikina
AU - Sara Mostafavi
PY - 2026
DA - 2026/06/08
JO - Nature Methods
DO - 10.1038/s41592-026-03124-8
UR - https://doi.org/10.1038/s41592-026-03124-8
AB - A key promise of sequence-to-functon(S2F) models is their ability to evaluate arbitrary sequence inputs, providing a robust framework for understanding genotype-phenotype relationships. However, despite strong performance across genomic loci, S2F models struggle with predicting inter-individual variation in gene expression. Training a model to make genotype-dependent predictions at a single locus—an approach we call personal genome training—offers a potential solution. We introduce SAGE-net, a scalable framework and software package for training and evaluating S2F models using personal genomes. Leveraging its scalability, we conduct extensive experiments on model and training hyperparameters, demonstrating that training on personal genomes improves predictions for held-out individuals. However, the model achieves this by identifying predictive variants rather than learning a cis-regulatory grammar that generalizes across loci. This failure to generalize persists across a range of hyperparameter settings. These findings highlight the need for further exploration to unlock the full potential of S2F models in decoding the regulatory grammar of personal genomes. Scalable software and infrastructure development will be critical to this progress.
SN - 1548-7091
SN - 1548-7105
ER -