Fine-tuning genomic models on rare splicing events identifies a novel biomarker of TDP-43 pathology
- 1Biomedical Engineering, Johns Hopkins University, Baltimore, MD, USA; Center for Computational Biology, Johns Hopkins University, Baltimore, MD, USA
Abstract
Genomic foundation models (GFMs) are pretrained on large genomic datasets, but their ability to predict rare disease-associated events after fine-tuning remains unclear. We tested this using TDP-43 cryptic exons, well characterized splicing events linked to neurodegenerative disease that are poorly represented in standard annotations. We compared the state-of-the-art GFM AlphaGenome with OpenSpliceAI, a specialized splice site prediction model. Fine-tuned OpenSpliceAI detected held-out cryptic exons in human and mouse and captured allele-specific splicing effects at the disease-associated UNC13A locus. In contrast, a splicing prediction head trained on AlphaGenome's frozen pretrained backbone failed to predict cryptic exons. Unfreezing and fine-tuning all model weights, however, enabled AlphaGenome to generalize to held-out data. To investigate cryptic splicing in cell types that are difficult to study experimentally, we applied fine-tuned OpenSpliceAI across the genome to study cell type-specific genes. This led to the identification of new cryptic splicing events in CLDN11 and ARHGAP23, which were subsequently validated in FTLD-TDP brain tissue. Notably, the cryptic exon in CLDN11 is predicted to generate a neoepitope that could serve as an oligodendrocyte-specific biomarker of TDP-43 pathology. Our results show that even a state-of-the-art GFM can fail to generalize to rare splicing events, even with substantial pretraining. However, both foundation and specialized models that learn to generalize to rare biological events can have broad applications in biological discovery and medicine.
Related
- Integrated transcriptomic analysis identifies Paternally Expressed Gene 10 as a novel candidate biomarker associated with aggressive features in liposarcoma
- Altered splicing machinery in lung carcinoids unveils NOVA1, PRPF8 and SRSF10 as novel candidates to understand tumor biology and expand biomarker discovery.
- Genomic Engineering of Gene Dosage: A Generalizable Framework for Modeling Haploinsufficiency-Mediated Human Disorders through Splicing Modulation
- Genomic analysis of DDT resistance identifies importance of glutathione transferases in driving evolutionary response to vector control in the malaria vector Anopheles gambiae
- Genus-wide re-speciation of Corynebacterium and genomic characterisation of 87 sinus isolates resolves taxonomic conflicts and identifies four candidate novel species
BibTeX
@misc{martinlinares2026finetuning,
title = {{Fine-tuning genomic models on rare splicing events identifies a novel biomarker of TDP-43 pathology}},
author = {C. P. Martin-Linares and A. Peethambaran Mallika and P. S. Sandal and T. W. Martin and M. Morris and P. C. Wong and J. P. Ling},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.07.757366},
url = {https://www.biorxiv.org/content/10.64898/2026.10.07.757366v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Fine-tuning genomic models on rare splicing events identifies a novel biomarker of TDP-43 pathology
AU - C. P. Martin-Linares
AU - A. Peethambaran Mallika
AU - P. S. Sandal
AU - T. W. Martin
AU - M. Morris
AU - P. C. Wong
AU - J. P. Ling
PY - 2026
DA - 2026/10/07
PB - bioRxiv
DO - 10.64898/2026.10.07.757366
UR - https://www.biorxiv.org/content/10.64898/2026.10.07.757366v1
AB - Genomic foundation models (GFMs) are pretrained on large genomic datasets, but their ability to predict rare disease-associated events after fine-tuning remains unclear. We tested this using TDP-43 cryptic exons, well characterized splicing events linked to neurodegenerative disease that are poorly represented in standard annotations. We compared the state-of-the-art GFM AlphaGenome with OpenSpliceAI, a specialized splice site prediction model. Fine-tuned OpenSpliceAI detected held-out cryptic exons in human and mouse and captured allele-specific splicing effects at the disease-associated UNC13A locus. In contrast, a splicing prediction head trained on AlphaGenome's frozen pretrained backbone failed to predict cryptic exons. Unfreezing and fine-tuning all model weights, however, enabled AlphaGenome to generalize to held-out data. To investigate cryptic splicing in cell types that are difficult to study experimentally, we applied fine-tuned OpenSpliceAI across the genome to study cell type-specific genes. This led to the identification of new cryptic splicing events in CLDN11 and ARHGAP23, which were subsequently validated in FTLD-TDP brain tissue. Notably, the cryptic exon in CLDN11 is predicted to generate a neoepitope that could serve as an oligodendrocyte-specific biomarker of TDP-43 pathology. Our results show that even a state-of-the-art GFM can fail to generalize to rare splicing events, even with substantial pretraining. However, both foundation and specialized models that learn to generalize to rare biological events can have broad applications in biological discovery and medicine.
ER -