Zejuan Li

Pathology foundation models trained from whole-slide images alone treat tissue morphology as an autonomous visual phenotype, leaving learned representations only weakly anchored to the molecular processes that generate tissue architecture. Here we present Fuji, a multimodal pathology foundation model that grounds morphology in transcriptomic and mutational supervision through a two-stage pretraining strategy: multi-teacher contrastive distillation that harmonizes formalin-fixed paraffin-embedded and fresh-frozen specimens, followed by multimodal multi-task learning that jointly optimizes masked image reconstruction, transcriptomic embedding reconstruction, and attention-aligned mutational signature inference. Trained on 60,546 whole-slide images paired with 45,675 bulk RNA-seq profiles and 22,683 whole-genome sequencing samples drawn from TCGA, GTEx, and CPTAC, Fuji yields a protocol-invariant, genome-informed embedding that we treat as a quantitative coordinate system for probing how genomic states project into tissue architecture. Genome-scale instability phenotypes, microsatellite instability, chromosomal instability, homologous recombination deficiency, and whole-genome duplication, are recoverable from histology with AUCs up to 0.98, consistently exceeding existing pathology foundation models. Extrachromosomal DNA amplification leaves a distinct morphological footprint that stratifies survival independently of tissue lineage. Etiologic mutational processes such as tobacco exposure and mismatch repair deficiency localize to discrete architectural neighborhoods, validated in two independent population-based cohorts, whereas clock-like signatures remain diffuse, delineating the scope and limits of genomic visibility from routine histology. Fuji further enables direct inference of tumor mutational burden, neoantigen load, immunotherapy response, and FDA-actionable driver mutations from hematoxylin and eosin slides. By coupling pretraining to molecular and mutational supervision, Fuji establishes an engineering basis for genome-informed inference from the most widely performed assay in oncology.

Spatial transcriptomics links gene expression with tissue morphology, however, current tools often prioritize genomic analysis, lacking integrated image interpretation. To address this, we present Thor, a comprehensive platform for cell-level analysis of spatial transcriptomics and histological images. Thor employs an anti-shrinking Markov diffusion method to infer single-cell spatial transcriptome from spot-level data, effectively combining gene expression and cell morphology. The platform includes 10 modular tools for genomic and image-based analysis, and is paired with Mjolnir, a web-based interface for interactive exploration of gigapixel images. Thor is validated on simulated data and multiple spatial platforms (ISH, MERFISH, Xenium, Stereo-seq). Thor characterizes regenerative signatures in heart failure, screens breast cancer hallmarks, resolves fine layers in mouse olfactory bulb, and annotates fibrotic heart tissue. In high-resolution Visium HD data, it enhances spatial gene patterns aligned with histology. By bridging transcriptomic and histological analysis, Thor enables holistic tissue interpretation in spatial biology.

Impact of long-term plasma storage on biomarker analysis is critical for ensuring data reliability. Cell-free DNA (cfDNA) epigenetic markers, including 5-hydroxymethylcytosine (5hmC), have emerged for disease detection, prognosis, and treatment response. However, the effects of prolonged storage on 5hmC analysis remain unclear. We evaluated the quantity and quality of cfDNA and 5hmC sequencing analyses in 1070 plasma samples stored for up to 14 years from patients with solid tumors and acute myeloid leukemia (AML) and non-cancer individuals. In long-term stored plasma samples, cfDNA yield remained largely stable; however, uniquely mapped reads (UMRs) from 5hmC sequencing were significantly reduced in solid tumor and control samples. Notably, prolonged plasma storage independently contributed to increased genomic DNA (gDNA) contamination in solid tumor and AML samples and significantly correlated with decreased UMRs in control samples. Across all groups, samples with gDNA contamination exhibited significantly reduced UMRs. Furthermore, gDNA contamination independently compromised cfDNA fragment integrity, decreased sequencing library success in solid tumors, and reduced 5hmC sequencing UMRs across all groups. Therefore, extended plasma storage contributes to increased gDNA contamination, compromising cfDNA and 5hmC sequencing quality. Implementing measures to minimize gDNA contamination in long-term plasma storage is crucial for improving downstream cfDNA analysis reliability.

Immune checkpoint inhibitors (ICIs) drastically improve therapeutic outcomes for lung cancer, but accurately predicting individual patient responses to ICIs remains a challenge. We performed the genome-wide profiling of 5-hydroxymethylcytosine (5hmC) in 85 plasma cell-free DNA (cfDNA) samples from lung cancer patients and developed a 5hmC signature that was significantly associated with progression-free survival (PFS). We built a 5hmC predictive model to quantify the 5hmC level and validated the model in the validation, test, and control sets. Low weighted predictive scores (wp-scores) were significantly associated with a longer PFS compared to high wp-scores in the validation [median 7.6 versus 1.8 months; p = 0.0012; hazard ratio (HR) 0.12; 95% confidence interval (CI), 0.03–0.54] and test (median 14.9 versus 3.3 months; p = 0.00074; HR 0.10; 95% CI, 0.02–0.50) sets. Objective response rates in patients with a low or high wp-score were 75.0% (95% CI, 42.8–94.5%) versus 0.0% (95% CI, 0.0–60.2%) in the validation set (p = 0.019) and 80.0% (95% CI, 44.4–97.5%) versus 0.0% (95% CI, 0.0–36.9%) in the test set (p = 0.0011). The wp-scores were also significantly associated with PFS in patients receiving single-agent ICI treatment (p < 0.05). In addition, the 5hmC predictive signature demonstrated superior predictive capability to tumor programmed death-ligand 1 and specificity to ICI treatment response prediction. Moreover, we identified novel 5hmC-associated genes and signaling pathways integral to ICI treatment response in lung cancer. This study provides proof-of-concept evidence that the cfDNA 5hmC signature is a robust biomarker for predicting ICI treatment response in lung cancer.

Accurate prognostic markers are essential for guiding effective lung cancer treatment strategies. The level of 5-hydroxymethylcytosine (5hmC) in tissue is independently associated with overall survival (OS) in lung cancer patients. We explored the prognostic value of cell-free DNA (cfDNA) 5hmC through genome-wide analysis of 5hmC in plasma samples from 97 lung cancer patients. In both training and validation sets, we discovered a cfDNA 5hmC signature significantly associated with OS in lung cancer patients. We built a 5hmC prognostic model and calculated the weighted predictive scores (wp-score) for each sample. Low wp-scores were significantly associated with longer OS compared to high wp-scores in the training [median 22.9 versus 8.2 months; p = 1.30 × 10⁻¹⁰; hazard ratio (HR) 0.04; 95% confidence interval (CI), 0.00–0.16] and validation (median 18.8 versus 5.2 months; p = 0.00059; HR 0.22; 95% CI: 0.09–0.57) sets. The 5hmC signature independently predicted prognosis and outperformed age, sex, smoking, and TNM stage for predicting lung cancer outcomes. Our findings reveal critical genes and signaling pathways with aberrant 5hmC levels, enhancing our understanding of lung cancer pathophysiology. The study underscores the potential of cfDNA 5hmC as a superior prognostic tool for guiding more personalized therapeutic strategies for lung cancer patients.

Measurable residual disease (MRD) is an important biomarker in acute myeloid leukemia (AML). However, MRD cannot be detected in many patients using current methods. We developed a highly sensitive 5-hydroxymethylcytosine (5hmC) signature in cell-free DNA by analyzing 115 AML patients and 86 controls. The 5hmC method detected MRD in 20 of 29 patients with negative MRD by multiparameter flow cytometry and 11 of 14 patients with negative MRD by molecular methods. MRD detection by the 5hmC method was significantly associated with relapse-free survival. This novel method can be used in most AML patients and may significantly impact AML patient management.

Epigenetic abnormality is a hallmark of acute myeloid leukemia (AML), and aberrant 5-hydroxymethylcytosine (5hmC) levels are commonly observed in AML patients. As epigenetic subgroups of AML correlate with different clinical outcomes, we investigated whether plasma cell-free DNA (cfDNA) 5hmC could categorize AML patients into subtypes. We profiled the genome-wide landscape of 5hmC in plasma cfDNA from 54 AML patients. Using an unbiased clustering approach, we found that 5hmC levels in genomic regions with a histone mark H3K4me3 classified AML samples into three distinct clusters that were significantly associated with leukemia burden and survival. Cluster 3 showed the highest leukemia burden, the shortest overall survival of patients, and the lowest 5hmC levels in the TET2 promoter. 5hmC levels in the TET2 promoter could represent TET2 activity resulting from mutations in DNA demethylation genes and other factors. The novel genes and key signaling pathways associated with aberrant 5hmC patterns could add to our understanding of DNA hydroxymethylation and highlight the potential therapeutic targets in AML. Our results identify a novel 5hmC-based AML classification system and further underscore cfDNA 5hmC as a highly sensitive marker for AML.

Background: Early cancer detection can dramatically improve clinical outcomes and greatly reduce economic burden of patients. Plasma cell–free DNA (cfDNA) 5‐hydroxymethylcytosine (5hmC) is an emerging epigenetic marker for cancer diagnosis. However, the utility of such marker has not been investigated in many common cancers yet. The purpose of this study was to evaluate 5hmC in plasma cfDNA for an early detection of common cancers.Methods: We used a highly sensitive nano‐5hmC‐Seal method and profiled the genome‐wide distribution of 5hmC in plasma cfDNA from 384 patients with bladder, breast, colorectal, kidney, lung or prostate cancer and 221 controls. Genes and signalling pathways with differential hydroxymethylation were analysed. We used machine learning to develop 5hmC signatures for cancer detection and cancer‐origin determination in the training sets and validated the signatures in the validation sets.Results: We identified genes and signalling pathways with aberrant DNA hydroxymethylation in six cancers. We discovered a pan‐cancer signature that detected all six cancers with a sensitivity of 68.6% and a specificity of 96.6% and cancer–specific signatures with a sensitivity of 80.0% for breast cancer, 88.9% for kidney cancer, 94.1% for lung cancer and 96.4% for prostate cancer, and a specificity of 100% for all except lung (96.2%). The sensitivity of cancer–specific signatures was 89.3%–100.0% for early‐stage cancers. The lung cancer–specific signature achieved a sensitivity of 98.0% and a specificity of 82.3% in an independent patient cohort with different ethnic backgrounds. Additionally, we discovered a 5hmC signature that could accurately determine cancer origin.Conclusions: We demonstrated that plasma cfDNA 5hmC is a highly sensitive biomarker for common cancer detection. Our genome‐wide analysis of 5hmC in six cancers reveals new target genes and signalling pathways with therapeutic potential for common cancers.

Aberrant changes in 5-hydroxymethylcytosine (5hmC) are a unique epigenetic feature in many cancers including acute myeloid leukemia (AML). However, genome-wide analysis of 5hmC in plasma cell-free DNA (cfDNA) remains unexploited in AML patients. We used a highly sensitive and robust nano-5hmC-Seal technology and profiled genome-wide 5hmC distribution in 239 plasma cfDNA samples from 103 AML patients and 81 non-cancer controls. We developed a 5hmC diagnostic model that precisely differentiates AML patients from controls with high sensitivity and specificity. We also developed a 5hmC prognostic model that accurately predicts prognosis in AML patients. High weighted prognostic scores (wp-scores) in AML patients were significantly associated with adverse overall survival (OS) in both training (P = 3.31e−05) and validation (P = 0.000464) sets. The wp-score was also significantly associated with genetic risk stratification and displayed dynamic changes with varied disease burden. Moreover, we found that high wp-scores in a single gene, BMS1 and GEMIN5 predicted OS in AML patients in both the training set (P = 0.023 and 0.031, respectively) and validation set (P = 9.66e−05 and 0.011, respectively). Lastly, our study demonstrated the genome-wide landscape of DNA hydroxymethylation in AML and revealed critical genes and pathways related to AML diagnosis and prognosis. Our data reveal plasma cfDNA 5hmC signatures as sensitive and accurate markers for AML diagnosis and prognosis. Plasma cfDNA 5hmC analysis will be an effective and minimally invasive tool for AML management.

To investigate germline predisposition in lymphoma, we performed whole-exome sequencing and discovered a novel variant (c.817-1G>T) in programmed cell death 1 ligand 2 (PD-L2) in a family with early-onset lymphomas and other cancers. The variant was present in the proband with follicular lymphoma and his son with Hodgkin's lymphoma. It was in the terminal splice acceptor site of PD-L2 and embedded in a putative enhancer of Janus kinase 2 (JAK2) and programmed cell death 1 ligand (PD-L1). We also found that gene expression of PD-L2, PD-L1, and JAK2 was significantly increased. Using 3' rapid amplification of cDNA ends (3 ' RACE), we detected an abnormal PD-L2 transcript in the son. Thus, the c.817-1G>T variant may result in the elevated PD-L2 expression due to the abnormal PD-L2 transcript and the elevated PD-L1 and JAK2 expression due to increased enhancer activity of PD-L1 and JAK2. The PD-L2 novel variant likely underlies the genetic etiology of the lymphomas in the family. As PD-L2 plays critical roles in tumor immunity, identification of PD-L2 as a germline predisposition gene may inform personalized immunotherapy in lymphoma patients.