Tissue-specific Transfer Learning Improves Functional Variant and Therapeutic Target Discoveries in Breast and Prostate Cancer
Qing Li, Dinghao Wang, Zilong Zhang, Deshan Perera, Zhishan Chen, Wanqing Wen, M. Ethan MacDonald, Weijia Cai, Jun Yan, Xiao-Ou Shu, Wei Zheng, Xingyi Guo, Quan Long
Abstract
DNA foundation models trained on large-scale genomic and epigenetic datasets have shown promise for regulatory variant interpretation, yet their application to tissue-specific contexts remain limited. Here, we present a transfer learning (TL) framework to adapt Enformer, a deep neural network trained on 5,313 multi-omics tracks, to breast and prostate cancer using 275 and 357 tissue-specific transcription factor (TF) ChIP–seq tracks, respectively. We computed tissue-specific cis-regulatory activity (tCRA) scores for millions of single-nucleotide variants (SNVs) in genome-wide association study (GWAS) datasets and prioritized high-impact SNV subsets (1M, 1.5M, and 2M).
Introduction
Machine learning (ML), particularly deep learning, has transformed biological data analysis by enabling the integration of multi-scale omics data and the modeling of complex regulatory systems. However, developing accurate ML models typically requires extensive training on large, high-quality datasets. In many specialized domains, such as tissue-specific or disease-focused analyses, such datasets are often unavailable or underpowered, limiting the applicability and relevance of general-purpose models. Moreover, Models trained on broadly heterogeneous datasets may capture signals not relevant to a particular tissue or disease, thereby underperforming in specialized applications.
Materials and Methods:
Our previous studies identified multiple risk TFs with central roles in regulating gene expression and significantly associated with breast and prostate cancer. Building on these findings, we systematically queried the Cistrome database and curated high-quality ChIP-seq tracks corresponding to risk TFs. We retained only tracks that passed quality control thresholds suggested by Cistrome (FastQC score > 25; uniquely mapped read ratio > 0.6; PCR bottleneck coefficient > 0.8; PeaksFoldChangeAbove10 ≥ 500; fraction of reads in peaks (FRiP) > 0.01; union DNase I hypersensitive (DHS) ratio > 0.7).
Discussion:
In this study, we applied transfer learning (TL) framework to redirect Enformer models to tissue-specific TL models and enhanced regulatory genetic variants and cancer suspicious genes discovery. Using the TL-derived tissue-based cis-regulatory activity (tCRA) scores, we prioritized millions of variants and defined high-impact SNV subsets for functional evaluation and gene discovery.
Acknowledgments:
The computational infrastructure was partly supported by a Canada Foundation for Innovation JELF grant (36605) to Q.L. (Q. Long).
Citation: Li Q, Wang D, Zhang Z, Perera D, Chen Z, Wen W, et al. (2026) Tissue-specific transfer learning improves functional variant and therapeutic target discoveries in breast and prostate cancer. PLoS Genet 22(5): e1012145. https://doi.org/10.1371/journal.pgen.1012145
Editor: Xiang Zhou, Yale University, UNITED STATES OF AMERICA
Received: September 16, 2025; Accepted: April 25, 2026; Published: May 6, 2026.
Copyright: © 2026 Li et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Code availability The transfer learning and calculation code for tCRAs are available from Github website: https://github.com/theLongLab/Transfer-Learning. Data availability GWAS summary statistic data for breast cancer were downloaded from the BCAC website (https://www.ccge.medschl.cam.ac.uk/breast-cancer-association-consortium-bcac/data-data-access/summary-results). GWAS summary statistic data for prostate cancer were downloaded from PRACTICAL(https://practical.icr.ac.uk/?page_id=8164). Variants scores from Enformer were downloaded from: https://github.com/deepmind/deepmind-research/tree/master/enformer Cistrome TF ChIP-seq bed peaks files were downloaded from: http://cistrome.org/. Roadmap project 15 chromatin states were downloaded from https://egg2.wustl.edu/roadmap/web_portal/chr_state_learning.html#core_15state. Bed files for nine cell lines with 15 chromatin states from Roadmap project were downloaded from: https://egg2.wustl.edu/roadmap/data/byFileType/chromhmmSegmentations/ChmmModels/coreMarks/jointModel/final/. Bed files for prostate cell lines with 15 chromatin states: https://ngdc.cncb.ac.cn/omix/release/OMIX237. ClinVar variants annotation summary are downloaded from: https://ftp.ncbi.nlm.nih.gov/pub/clinvar/tab_delimited/variant_summary.txt.gz. Five diseases genes were downloaded from DisGeNET: https://www.disgenet.org/search with the following concept Unique Identifier (CUI): C0678222 (Breast Carcinoma), C3539878 (Triple Negative Breast Neoplasms), C0006142 (Malignant neoplasm of breast), C4722518 (Triple-Negative Breast Carcinoma) are used for breast cancer and C0600139 (Prostate carcinoma) and C0376358 (Malignant neoplasm of prostate) for prostate cancer. Gene essentiality scores were downloaded from DepMap: https://depmap.org/portal/. Drugs and their targets are retrieved from DrugBank (https://go.drugbank.com/), ChEMBL (https://www.ebi.ac.uk/chembl/) and therapeutic target databases (https://idrblab.net/ttd/). Enformer model weights: https://www.kaggle.com/models/deepmind/enformer.
Funding: This research was supported by the grant from US National Institutes of Health grant R37 CA227130 and R01 CA269589 to X.G. and a New Frontiers in Research Fund (NFRFE-2023-00291) and NSERC Discovery Grant (RGPIN-2024-04679) to Q.L. (Q. Long), D.P. was supported by an Alberta Innovates and an Eyes High scholarship. D.W. was supported by an Alberta Innovates scholarship. W.C. was supported by start-up funding from Southern Illinois University School of Medicine. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
