Back

A DNA foundation model predicts osteoporosis risk genes without proximity bias

Regep, C.; Kapourani, C.-A.; Sofyali, E.; Dobrowolska, A.; Loukas, G.; Anighoro, A.; Canale, E.; Gross, T.; Licciardello, M.; Gupta, R.; Maciuca, S.; Desai, T.; Del Vecchio, A.; Field, C.; Gemayel, K.; Javer, A.; Zhang, Z.; Tsujikawa, R.; Inoue, F.; Hessel, E.; Taylor-King, J.; Whittaker, J.; Roblin, D.; McIntyre, R.; Edwards, L.

2026-03-12 genetics
10.64898/2026.03.09.707383 bioRxiv
Show abstract

Targets supported by human genetic associations are more than twice as likely to progress from clinical development to approval. Genome-wide association studies are the largest source of genetic evidence for disease risk but linking non-coding variants to effector genes remains a significant barrier to identifying causal targets. Current gene-mapping approaches suffer from proximity bias, largely ignoring distal genes. Here we introduce Rosalind, a DNA foundation model fine-tuned on human genetic variation from GTEx, that directly predicts variant-gene regulatory relationships from sequence without relying on nearest-gene heuristics. We demonstrate Rosalinds accuracy through extensive benchmarking, apply it to multiple complex traits to establish broad utility, and provide experimental validation in osteoporosis using a translational osteoblast assay. We demonstrate that genes distal to osteoporosis risk variants were significantly more likely to alter a bone formation phenotype than nearest genes. Together, these results highlight deep learning-based regulatory models as a general and scalable framework for translating novel genetic insights to drug discovery.

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

1
Nature Communications
○ 5641 papers in training set
Top 14%
12.9%
Similar papers in this journal
2
Genome Biology
○ 637 papers in training set
Top 0.6%
12.9%
Similar papers in this journal
3
Cell Genomics
○ 172 papers in training set
Top 0.2%
11.9%
Similar papers in this journal
4
The American Journal of Human Genetics
○ 234 papers in training set
Top 0.6%
7.9%
Similar papers in this journal
5
Nature Genetics
○ 286 papers in training set
Top 1%
5.5%
Similar papers in this journal
50% of probability mass above
6
eLife
● 5828 papers in training set
Top 26%
4.9%
7
Proceedings of the National Academy of Sciences
● 2444 papers in training set
Top 13%
4.0%
8
Science
● 477 papers in training set
Top 2%
4.0%
9
PLOS Genetics
● 862 papers in training set
Top 4%
3.2%
10
Science Advances
● 1243 papers in training set
Top 14%
2.4%
11
Bioinformatics
◐ 1204 papers in training set
Top 6%
2.1%
12
Molecular Systems Biology
○ 162 papers in training set
Top 1%
2.1%
13
Nucleic Acids Research
◐ 1281 papers in training set
Top 8%
1.9%
14
Human Genetics and Genomics Advances
○ 84 papers in training set
Top 1%
1.7%
15
Cell
○ 431 papers in training set
Top 6%
1.7%
16
Nature
○ 645 papers in training set
Top 7%
1.5%
17
PLOS Computational Biology
● 1863 papers in training set
Top 15%
1.5%
18
Nature Methods
○ 385 papers in training set
Top 5%
1.1%
19
Cell Reports
○ 1498 papers in training set
Top 26%
0.9%
20
Genome Medicine
○ 183 papers in training set
Top 5%
0.9%
21
PLOS ONE
● 5266 papers in training set
Top 58%
0.8%
22
Cell Systems
○ 201 papers in training set
Top 5%
0.8%
23
Genome Research
● 468 papers in training set
Top 6%
0.8%
24
Nature Biotechnology
○ 172 papers in training set
Top 5%
0.6%
25
Nature Neuroscience
○ 252 papers in training set
Top 5%
0.6%

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.