EPIC-Seq
epicseq.stanford.eduInfers tumor gene expression from the fragmentation patterns of cell-free DNA at gene promoters.
Paper: Inferring gene expression from cell-free DNA fragmentation profiles
EPIC-Seq website ↗Tools
Techniques, assays, and computational methods developed by the lab and its principal investigator, with the publications that describe them.
Published
Infers tumor gene expression from the fragmentation patterns of cell-free DNA at gene promoters.
Paper: Inferring gene expression from cell-free DNA fragmentation profiles
EPIC-Seq website ↗Assigns classic Hodgkin lymphoma cases to the H1 and H2 genomic subtypes from noninvasive profiling.
Paper: Distinct Hodgkin lymphoma subtypes defined by noninvasive genomic profiling
Hodgkin Lymphoma Clusterer website ↗Early, noninvasive identification of patients who benefit from immune checkpoint inhibition.
Paper: Noninvasive early identification of therapeutic benefit from immune checkpoint inhibition
DIREct website ↗Continuous Individualized Risk Index — dynamic outcome prediction from serial tumor biomarkers.
Paper: Dynamic risk profiling using serial tumor biomarkers for personalized outcome prediction
CIRI website ↗Lung Cancer Likelihood in Plasma — integrates genomic features for early lung cancer detection.
Paper: Integrating genomic features for non-invasive early lung cancer detection
CLiP website ↗In development
Explainable AI that classifies hematologic malignancies from targeted-sequencing genomic profiles, clonal dominance, and immune-repertoire features, with rule-based rationales for every call.
Comprehensive classification of hematological malignancies by integrating genomic and immune repertoire profiles with explainable AI
Blood 146 (Suppl. 1), ASH Annual Meeting 2025
Background. Accurate classification of hematologic malignancies remains challenging due to overlapping morphological, immunophenotypic, genomic, and immunological features. Shared mutational patterns among subtypes of lymphoid and myeloid malignancies further complicate accurate molecular classification. Additionally, distinguishing between clonal driver mutations, subclonal events, and clonal hematopoiesis of indeterminate potential (CHIP) complicates interpretation of genomic profiles, often requiring extensive expert input and delaying definitive diagnoses. Developing robust and interpretable machine learning (ML) frameworks capable of integrating genomic and immune repertoire features will greatly benefit efforts to accelerate and standardize diagnostics.
Methods. We curated a cohort of 11,090 specimens from 7,543 patients profiled using the Stanford Actionable Mutation Panel for Hematopoietic and Lymphoid Neoplasms (Heme-STAMP) between Oct. 2020 and Jun. 2025. Heme-STAMP is a targeted NGS panel curated for clinical relevance in myeloid and lymphoid malignancies and employing deep sequencing (~2000X coverage). Cases were initially annotated with 138 detailed diagnostic categories, and harmonized into 15 major malignant classes. Around 30% of cases lacked an established diagnosis at the time of testing and therefore were excluded. Samples without clinical suspicion of malignancy comprised the benign class. Genomic features used for the classifier training encompassed somatic mutations, canonical fusions/translocations, and focal copy number alterations (CNAs). To prioritize truncal mutations, we introduced an allele frequency-rank feature, capturing clonal dominance within each sample. TCR and IG sequencing data from 1,001 samples were used to derive immune repertoire features. Finally, we developed a comprehensive suite of ML models as well as advanced neural networks, coupled with an explainable AI framework that generates transparent, rule-based rationales alongside predictions, and called this framework Heme-xAIGen.
Results. After excluding samples labeled as unknown malignancy, the training (80%) and validation (20%) cohorts comprised diverse malignancies, including AML (15.6%), MPN (15.5%), PCNs (11.3%), CLL/SLL (8.6%), MDS (8.9%), DLBCL (3.9%), and FL (2.8%). Benign cases accounted for ~13% of the dataset. TET2 (29.6%), ASXL1 (26.7%), and DNMT3A (20.5%) were the most frequently mutated genes in myeloid neoplasms; in mature lymphoid neoplasms we observed high frequencies of TP53 (17%), CREBBP (12.7%), MYD88 (10.9%), and BCL2 (5.2%). Benign samples showed a significant correlation between patient age and CHIP-associated mutation burden (R = 0.42, P < 0.05), and TCL cases exhibited a significantly higher frequency of positive TCR clonality than mature B-cell neoplasms (80.5% vs. 16.5%; P < 0.001). The ML pipeline accurately classified hematological malignancies, with the best-performing model achieving an overall AUC of 0.89 (AML 0.80; MPN 0.91; CLL/SLL 0.89; DLBCL 0.78). All feature types contributed, with mutations and CNAs having the greatest impact. Heme-xAIGen slightly improved discrimination (AUC = 0.90) while generating biologically coherent rules, for example a CCND1::IGH fusion with absence of FLT3 and NPM1 mutations for MCL. Longitudinal application in patients with antecedent myeloid conditions assigned higher AML risk scores to patients who subsequently progressed to AML (P < 0.05).
Conclusions. We introduce Heme-xAIGen, an explainable AI framework that accurately classifies hematologic malignancies by integrating genomic profiles, clonal dominance through VAF ranking, and immune repertoire data. Our approach demonstrates substantial predictive accuracy and delivers transparent diagnostic rationales. By distinguishing complex lymphoid and myeloid malignancies and identifying disease progression, the model addresses critical diagnostic gaps, supports clinical decision-making, and facilitates patient monitoring.
Machine-learning estimation of telomere length from short-read whole-genome sequencing, down to low-pass plasma cfDNA, with a Telomere Ambiguity Index that captures heterogeneity in telomere maintenance.
A pan-cancer atlas of telomere length in tissues and plasma cell-free DNA
Cancer Research 86 (7 Suppl.), AACR Annual Meeting 2026, Abstract 6892
Background. Telomeres shorten with age because of incomplete replication and limited telomerase activity, and their dysfunction promotes genomic instability in cancer. Existing estimators rely on strict repeat counts and often miss error-affected or biologically shortened fragments. This is problematic in plasma cfDNA where most fragments are ∼170 bp and ctDNA is even shorter. Nanopore-based assays provide precise measurement, but an accurate, error-tolerant short-read approach remains unmet.
Methods. We developed VERACITY, a machine-learning method that estimates telomere length (TL) from short-read WGS by modeling telomeric-repeat patterns and sequencing-error profiles. VERACITY also computes a Telomere Ambiguity Index (TAI) that quantifies heterogeneity in telomere maintenance. We applied VERACITY to 10,363 TCGA WGS samples (4,847 tumors; 5,089 matched normals; ages 14–85) across 27 cancer types, 3 deep (∼200×) cfDNA genomes with systematic down-sampling, and 462 low-pass (2–3×) cfDNA genomes from the DELFI cohort (233 cancers; 229 healthy), generating an atlas of telomere biology in tissues and plasma.
Results. Down-sampling showed reliable TL estimation down to ∼2.5× coverage, supporting low-pass cfDNA application. Tissue and cfDNA trends were highly concordant. Across cancers, TL declined with age in tumors and matched normals, with tumors showing far steeper erosion. Male tumors showed greater loss than female tumors. Age-associated trajectories varied widely: GBM, TGCT, and STAD showed steep nonlinear declines, whereas CESC, KICH, and LGG showed minimal or positive slopes. Although stomach tissues show the weakest TL–age correlation, gastric cancers exhibited one of the strongest declines in both tissue and cfDNA. Several cancers, including PRAD and TGCT, displayed elevated TAI, suggesting diverse or unstable maintenance strategies.
Conclusions. VERACITY provides an accurate and scalable framework for TL estimation from standard and low-pass sequencing, enabling the most comprehensive atlas to date of telomere dynamics across cancers in tissues and plasma. The analysis reveals accelerated, sex-stratified, and cancer-type-specific telomere erosion, substantial inter-tumor heterogeneity, and distinct maintenance mechanisms. cfDNA TL patterns closely mirror tissue-derived trends and highlight the potential of telomere metrics as minimally invasive biomarkers of telomere dysfunction, tumor biology, and treatment-relevant telomere states.
Somatic mutation Prediction via Artificial intelligence for Rapid discovery and Knowledge translation. SPARK imputes genome-wide somatic mutations that targeted sequencing panels do not cover, by learning from multimodal data such as mutations and gene expression across cancers.
AI framework for imputing genome-wide somatic mutations in pan-cancer targeted sequencing data for biomarker discovery
Cancer Research 85 (8 Suppl. 1), AACR Annual Meeting 2025, Abstract 1116; poster award, Stanford Bio-X Symposium 2026
Background. Tumor genotyping via high-throughput sequencing has revolutionized our understanding of various cancer types, and genotyping of tumor or liquid biopsies is now part of clinical routine. The assays used, whether cancer-specific, pan-cancer solid tumor, or heme panels, often target less than 0.5% of the genome, focusing on known and recurrent genes. This enrichment allows mutations to be found at lower allele fractions, but it misses hundreds or thousands of potential mutations, some of which could be biomarkers for prognostic or predictive models or targets for future drug development. We hypothesize that an artificial intelligence (AI) framework can unravel the complex interactions between mutations and thereby impute the off-target space.
Methods. We introduce SPARK (Somatic mutation Prediction via Artificial intelligence for Rapid discovery and Knowledge translation), an algorithm for imputing missing genomic loci in existing genomic panels. SPARK relies on the assumption that there is a latent space in which mutations co-occur or are mutually exclusive across multiple malignancies. The framework uses other data modalities, such as gene expression, to train a unified model for solid tumors. During training we use somatic mutations and RNA expression data, then employ a second network to transfer knowledge, so that a small number of mutations from any sequencing panel suffice to impute the rest.
Results. We trained the model on genotyping and RNA expression data from TCGA across multiple carcinomas (n > 2,000 samples), then fine-tuned it for each panel. To evaluate performance we focused on genes present on the MSK-IMPACT panels from the GENIE Project v16. A number of genes mutated at over 1% frequency across carcinomas in TCGA were not covered by MSK-IMPACT; imputing them achieved precision from 60% to 100% for a minority of genes (for example ETV4 and ETV5), and much lower precision (under 10%) for many others, suggesting those mutations are likely not functional and lack strong associations with panel genes. We plan to apply the model to the full GENIE Project cohort to impute the genes with acceptable precision, enabling biomarker discovery from a much larger set of samples.
An unsupervised, annotation-free framework for single-cell segmentation and morphological profiling in multiplex tissue imaging, producing soft whole-cell masks that respect the crowded, overlapping cells of lymphoid tissue.
Distinct morphological features of malignant cells in large B-cell lymphoma identified by soft-segmentation in imaging-based spatial proteomics
Blood 146 (Suppl. 1), ASH Annual Meeting 2025
Background. Mature B-cell neoplasms display densely packed and morphologically heterogeneous cells, often complicating tissue-level analysis. Spatial profiling platforms such as imaging mass cytometry (IMC) enable single-cell, in situ analysis using multiplexed antibody panels, but accurate and interpretable segmentation remains a challenge in lymphoid tissues where cells frequently overlap, deform, and exhibit non-canonical marker profiles. Common tools, including DeepCell Mesmer and Cellpose, assume round, non-overlapping cells with clear boundaries, a simplification that often fails in DLBCL, FL, and cHL. These limitations can yield biologically misleading segmentations and restrict precise characterization of the tumor microenvironment (TME).
Methods. We developed SegMO (Segmentation and Morphological feature Optimization), an unsupervised, annotation-free segmentation framework compatible with multiple spatial proteomics platforms. SegMO integrates nucleus-anchored seeding, probabilistic modeling of marker expression, cluster-guided refinement, and a reinforcement learning (RL) module to produce biologically realistic, soft whole-cell masks that allow partial overlaps and irregular shapes. Protein expression is modeled with flexible statistical distributions to accommodate partial or co-expressed markers, and cells are clustered by expression features before iterative mask refinement to improve marker purity and spatial coherence. The RL model optimizes boundaries through pixel-level adjustments to maximize cell fidelity, recover low-signal regions, and capture realistic morphology.
Results. We applied SegMO to 24 IMC samples (tonsil, spleen, brain, and lymphoma) stained with 43 protein markers, from a larger cohort of 76 individuals (DLBCL, FL, cHL, and healthy donors). SegMO segmented ~74,000 cells and grouped them into biologically distinct clusters. Compared to Mesmer, SegMO improved nuclear signal integrity (0.92 vs. 0.90), detected fewer weak-signal nuclei (<0.1% vs. 2.5%), and generated more accurate boundaries in high-density and overlapping regions. Soft segmentation preserved subtle morphological features such as protrusions, compression, elongation, and partial overlaps that rigid models did not capture. Morphological profiling showed that the cluster enriched in tumor tissues, likely neoplastic B cells, had among the largest cell sizes and highest nuclear ratios, and a CD4⁺CD7⁺ lymphocyte cluster enriched in the GCB subtype showed irregular shapes reflecting activation, mobility, and spatial pressure in the TME. SegMO also uncovered architectural features including immune–tumor colocalization, irregular clustering, and localized crowding in aggressive lymphomas.
Conclusions. SegMO is a robust, versatile, annotation-free AI framework for single-cell segmentation in multiplex tissue imaging of hematologic samples. By integrating probabilistic modeling, cluster-guided refinement, and reinforcement learning, SegMO produces biologically realistic boundaries in complex TMEs, enhancing both morphological fidelity and marker signal accuracy, and enabling precise characterization of tumor-associated features, immune interactions, and proliferative states in lymphomas and other hematologic malignancies.
A mechanistic, generative model of cell-free DNA fragmentation that simulates nucleosome positioning and nuclease cutting, now informed by cell-type-specific nucleosome patterns to move toward tissue-of-origin deconvolution.
Using cell-type-specific nucleosome patterns to improve cell-free DNA fragmentation models in lung cancer
Cancer Research 86 (7 Suppl.), AACR Annual Meeting 2026, Abstract 6845
Background. Cell-free DNA (cfDNA) profiling has emerged as a promising tool in cancer detection. Methods include the identification of circulating tumor DNA (ctDNA), the profiling of cfDNA methylation, and the analysis of cfDNA fragments, including fragment size and position. We previously showed that fragment size diversity is strongly correlated with gene expression (Esfahani et al., 2022) and subsequently developed a mechanistic model for the cfDNA fragmentation process to enhance gene expression inference from cfDNA (Liu et al., 2025). However, analysis of cfDNA fragment size distribution alone leaves out useful information about fragment position. Since fragment position arises from nucleosome position, we incorporate prior knowledge about nucleosome positioning to improve our cfDNA fragmentation models. We hypothesize that incorporating cell-type-specific nucleosome patterns can progress cfDNA models towards cell-type deconvolution.
Methods. To incorporate prior knowledge about cell-type-specific nucleosome organization into our cfDNA fragmentation model, we use Micrococcal Nuclease sequencing (MNase-seq) data and MNase-Transcription Start Site Sequence Capture (mTSS-seq) data. For each transcription start site (TSS), we leverage this prior knowledge about nucleosome organization to simulate well-positioned nucleosomes. We then generate additional nucleosomes given the gene's strand information and cell type, generate cuts, and produce a mixture of simulated fragments from various cell types.
Results. To evaluate the updated cfDNA fragmentation model, we calculate the coverage over 10 bp intervals within a window of the TSS and compare it with the coverage for cfDNA control samples, alongside the same comparison for fragments simulated with our previous model. The updated model incorporating prior knowledge about nucleosome organization outperformed the previous model in simulating cfDNA fragmentation.
Conclusions. We show that integrating cell-type-specific nucleosome patterns improves cfDNA fragmentation models. Further improvements can be made given nucleosome organization information from additional cell types. We envision that this model-based approach will enhance tissue-of-origin classification and cancer detection.
Recovers B- and T-cell receptor clonality from Heme-STAMP comprehensive genomic profiling, using IGH and TCR J-segment coverage together with off-target and unmapped reads, so that dedicated V(D)J assays are not required.
Accurate T- & B-cell clonality inference from comprehensive genomic profiling by targeted sequencing
ASH Annual Meeting 2026, poster presentation; ASH Abstract Achievement Award
Background. Immune-receptor clonality testing is often used for diagnosis and monitoring of lymphoid malignancies but generally requires dedicated V(D)J assays to interrogate B- and T-cell receptor (BCR/TCR) recombinations. We developed J-TRACE to recover clonality information from Heme-STAMP, Stanford's institutional comprehensive tumor-genotyping panel for hematolymphoid malignancies, by exploiting its IGH and TCR J-segment coverage together with off-target/unmapped reads, and now share this resource.
Methods. We analyzed 1,328 tissue specimens from 977 Heme-STAMP profiled patients at ~2,000× median depth. T-cell lymphomas, DLBCL, FL, and myeloma comprised the most prevalent lymphoid tumor types, accounting for ~30% of cases. Of these 1,328 cases, 942 (71%) had clonal status available by an orthogonal assay (Invivoscribe LymphoTrack) for training/validation, with the remaining 386 (29%) used for additional validation by malignancy type. We developed J-TRACE to retrieve reads overlapping J segments, their mates, unmapped reads, and off-target reads at V loci, and to refine mappings to 1,169 IMGT V- and J-region alleles by BLAST, assembling fixed VJ-pairings with per-chain V×J fragment-count matrices for TRB/G and IGH/K. These count matrices were further summarized into 76 to 392 features to train probabilistic BCR/TCR clonality models through an ensemble machine learning framework. The performance of J-TRACE was then evaluated by cross-validation. As a comparison, MiXCR was benchmarked on identical reads from 400 balanced specimens.
Results. For TCR clonality analysis (TRB/G), we studied 842 specimens (337 nonclonal, 128 oligoclonal, 377 clonal), and for BCR clonality analysis (IGH/K), we studied 358 specimens (213 nonclonal, 4 oligoclonal, 141 clonal). Among specimens with TRB scores, maximum clone abundance increased monotonically from 7.8% in nonclonal to 15% in oligoclonal and 26.8% in clonal specimens (Kruskal-Wallis P << 0.001), supporting oligoclonality as an intermediate state; similar trends were seen for BCR clonality. Among the features used by J-TRACE, maximal V×J clone fraction for TRB and V×J Gini for TRG, IGH, and IGK showed the strongest discriminatory power for clonality, with corresponding AUCs of 0.801, 0.736, 0.607, and 0.683. Performance increased with junctional depth. J-TRACE-TCR achieved an AUC of approximately 0.86 while the J-TRACE-BCR model achieved an AUC of 0.76. Notably, the J-TRACE-TCR model gained 0.03 in AUC over the TRB maximal clone fraction, and achieved 68% sensitivity at 90% specificity. Gains were greatest in bone marrow and FFPE specimens, whereas high-depth peripheral-blood specimens derived little incremental benefit. Against MiXCR on identical reads, repertoire burden correlated strongly (Spearman ρ = 0.74; P < 0.001).
An improved EPIC-Seq framework. iEPIC is a deep generative model that learns a domain-transfer function from plasma promoter fragmentation entropy to RNA-like expression profiles, trained on paired and unpaired tumor and plasma samples, and underpins noninvasive classification and molecular subtyping of lymphoid neoplasms.
Non-invasive mantle cell lymphoma risk stratification by inference of tumor proliferation from cfDNA
Blood 146 (Suppl. 1), ASH Annual Meeting 2025
Background. Variability in tumor proliferation is a key prognostic factor in mantle cell lymphoma (MCL), but its measurement by gene expression profiling of frozen tissue or by Ki-67 index of fixed tissue has faced limitations. The MCL35 gene expression signature is a validated independent prognostic biomarker that quantifies proliferation on routinely fixed tissues, yet its clinical use is limited by the need for invasive biopsies and by technical challenges in bone marrow and leukemic samples. No liquid biopsy approach has yet bridged the gap between blood- and tissue-derived prognostic signatures for measuring MCL proliferation-associated risk noninvasively.
Methods. We profiled 161 samples from 117 treatment-naïve MCL patients in the Phase 3 randomized LYMA trial, including 88 matched tumor–plasma specimens from 44 patients. Tumor specimens (n = 100) were profiled by RNA-Seq; plasma specimens were profiled by EPIC-Seq using a customized lymphoma-specific panel that includes key proliferation-associated MCL genes. We developed a novel deep generative machine-learning model, termed iEPIC, to better infer transcriptomic expression from cfDNA fragmentomic features. Trained on both paired and unpaired tumor and plasma samples (60% training, 40% test), iEPIC learns a domain-transfer function that reconstructs RNA-like profiles from raw plasma EPIC-Seq promoter fragmentation entropy (PFE) measurements. MCL35 scores computed from the predicted profiles were benchmarked against tumor-derived scores and evaluated for survival stratification and prediction of early progression (POD24).
Results. In tumor RNA-Seq profiles, higher MCL35 scores were associated with blastoid morphology and Ki-67 index, and with inferior PFS and OS, and remained prognostic in multivariate models adjusting for MIPI and Ki-67. In plasma, iEPIC significantly improved single-gene correlations between noninvasive cfDNA measurements and tumor RNA, including key mitotic genes MKI67, TOP2A, and FOXM1. Across the full MCL35 signature, iEPIC reached ~0.7 Pearson correlation against tumor RNA-Seq, improving precision, recall, and F1 of noninvasive measurement by ~23%. The noninvasive iEPIC model stratified patients across Ki-67 groups and predicted POD24 with an AUC of 0.73, compared with 0.83 for invasive MCL35 from tumor RNA-Seq.
Conclusions. We describe a noninvasive approach for computing MCL tumor proliferation scores directly from blood cfDNA, toward risk stratification without tissue biopsies. The model maintains gene-level biological fidelity and could be deployed where tissue is unavailable or serial monitoring is required. In a companion study, EPIC-Seq profiling of 748 subjects enabled noninvasive classification of ten lymphoid neoplasms (top-2 accuracy 94%), cell-of-origin and genetic subtyping of DLBCL, H1/H2 subtyping of Hodgkin lymphoma, and detection of high-grade B-cell lymphoma with a dark-zone signature.
Software accompanying our publications is released on the lab's GitHub.