Qing Li

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). These TL-prioritized variants demonstrated consistently greater enrichment in tissue-specific enhancers, cancer GWAS risk variants, and ClinVar pathogenic variants compared to the original Enformer model. Transcriptome-wide association studies (TWAS) using TL-based SNVs identified more cancer-relevant genes, many of which exhibited functional essentiality (DepMap), therapeutic tractability (drug databases), and disease relevance (DisGeNET). Notably, TL models outperformed the base model in identifying genes enriched for drug targets and clinically relevant disease associations. Our results show that TL-derived tCRA scores enhance regulatory variant prioritization and improve susceptibility gene discovery in a tissue-specific manner. Our study provides a generalizable framework for tailoring foundation models to disease-relevant contexts, with implications for variant interpretation, therapeutic target discovery, and precision medicine.

Susceptibility transcription factors (TF) whose DNA bindings are altered by genetic variants regulating colorectal cancer (CRC) risk genes remain poorly defined. Using generalized linear mixed models, we analyze 218 TF ChIP-Seq datasets alongside GWAS data from 100,204 CRC cases and 154,587 controls of East Asian and European ancestries. We identify 51 TFs and TF-cofactor interactions, including VDR-cofactors, as key regulators of CRC risk. Integrating these TF insights with transcriptome-wide association studies (TWAS), we further evaluate associations between genetically predicted gene expression, alternative splicing, and alternative polyadenylation with CRC risk, using RNA-seq data from 364 Asian-ancestry and 707 European-ancestry individuals. Multi-ancestry TWAS identify 222 risk genes, including 95 novel genes and 48 potentially druggable targets. Single-cell analysis provides additional functional evidence supporting ~45% of these genes, and experimental validation confirms oncogenic roles for RHPN2, IRS2, and TXN. Our findings elucidate key TF–gene regulatory networks and uncover novel CRC risk genes.

Transcriptome-wide association study (TWAS) is an emerging model leveraging gene expressions to direct genotypephenotype association mapping. A key component in TWAS is the prediction of gene expressions; and many statistical approaches have been developed along this line. However, a problem is that many genes have low expression heritability, limiting the performance of any predictive model. In this work, hypothesizing that appropriate denoising may improve the quality of expression data (including heritability), we propose AE-TWAS, which adds a transformation step before conducting standard TWAS. The transformation is composed of two steps by first splitting the whole transcriptome into co-expression networks (modules) and then using autoencoder (AE) to reconstruct the transcriptome data within each module. This transformation removes noise (including nonlinear ones) from the transcriptome data, paving the path for downstream TWAS. We showed two inspiring properties of AE-TWAS: (1) After transformation, the transcriptome data enjoy higher expression heritability at the low-heritability spectrum and possess higher connectivity within the modules. (2) The transferred transcriptome indeed enables better performance of TWAS; and moreover, the newly formed highly connected genes (i.e., hub genes) are more functionally relevant to diseases, evidenced by their functional annotations and overlap with TWAS hits.

Identifying risk protein targets and their therapeutic drugs is crucial for effective cancer prevention. Here, we conduct integrative and fine-mapping analyses of large genome-wide association studies data for breast, colorectal, lung, ovarian, pancreatic, and prostate cancers, and characterize 710 lead variants independently associated with cancer risk. Through mapping protein quantitative trait loci (pQTL) for these variants using plasma proteomics data from over 75,000 participants, we identify 365 proteins associated with cancer risk. Subsequent colocalization analysis identifies 101 proteins, including 74 not reported in previous studies. We further characterize 36 potential druggable proteins for cancers or other disease indications. Analyzing >3.5 million electronic health records, we conducted analyses of emulated trials for 11 drugs across 290 comparisons and identified three drugs significantly associated with reduced colorectal cancer risk: caffeine vs. paroxetine (HR, 0.51; 95% CI, 0.41–0.64), haloperidol vs. prochlorperazine (HR, 0.47; 95% CI, 0.33–0.68), and trazodone hydrochloride vs. paroxetine (HR, 0.49; 95% CI, 0.38–0.63). Conversely, caffeine was associated with increased cancer risk in comparisons with finasteride (colorectal cancer) and fluoxetine (breast cancer). Meta-analysis identified six drugs significantly associated with cancer risk, including acetazolamide, which was associated with reduced colorectal cancer risk (HR, 0.79; 95% CI, 0.72–0.87). This study identifies novel protein biomarkers and candidate drug targets across six major cancer types and highlights several approved drugs with potential chemopreventive effects.