Generalizable cancer detection from ultra-low-pass WGS via deep contextual modeling of cfDNA sequences

Integrative cfDNA profiling from low-pass whole-genome sequencing enables tissue-of-origin prediction in cancer

Cancer type classification is challenging due to tumor heterogeneity and undefined tissue of origin (TOO), particularly in cancers of unknown primary (CUP) and multiple primary cancers (MPC). Accurate TOO identification is critical for guiding treatment and prognosis. We developed a stacked ensemble machine learning classifier that integrates 11 multidimensional cfDNA features spanning genomic, fragmentomic, methylation/repeat, and microbial signals. Base models were constructed using five algorithms, including Deep Learning, Distributed Random Forest, Gradient Boosting Machine, Generalized Linear Model, and XGBoost, within a five-fold cross-validation framework, and their predictions were aggregated into a final ensemble optimized for top-1 accuracy. The classifier achieved robust performance across 17 cancer types, with top-1 and top-2 accuracies of 78% and 89% in the training cohort (n = 1,814), and 80% and 90% in an independent validation cohort (n = 1,221). Notably, predictive performance was retained in samples with low tumor fraction (71% top-1, 85% top-2). Sensitivity varied across tumor types, with the highest performance observed in head and neck and colorectal cancers. Among CUP cases, 11 of 15 (73.3%) predictions matched clinically inferred primary sites based on multimodal diagnostics. Feature importance analysis identified nucleosome positioning, fragment size distribution, and repeat elements as key contributors to model performance. Collectively, this cfDNA-based classifier provides a robust and non-invasive approach for accurate cancer type identification and has the potential to support clinical decision-making.

An Evaluation of DMR Informed Fine-Tuning of Tissue Array Pretrained CpGPT for Gastrointestinal Cancer Classification Using cfDNA Targeted Methylation

MoESurv: a zero-sample and transferable survival prediction framework for rare cancers using mixture of experts

Abstract Motivation Accurate survival prediction is crucial for personalized cancer treatment but remains challenging for rare cancers due to limited data. Most deep learning models require large training datasets, which are unavailable for rare cancer types, creating a significant clini-cal bottleneck. Results We propose MoESurv, a zero-sample survival prediction framework that leverages a mix-ture-of-experts architecture to extract generalizable prognostic patterns from pan-cancer data. The model integrates shared experts, cancer-specific experts, and routing experts within an autoencoder to disentangle common and type-specific survival features. Evaluated on seven rare TCGA cancer types, MoESurv achieved state-of-the-art performance, improving the average C-index by 4 percentage points over the best baseline. Further-more, external validation across diverse populations and independent cohorts—including a Chinese glioma cohort (CGGA mRNAseq_693, C-index = 0.7433), a rare GBM IDH-mutant subtype (C-index = 0.8064), and the pan-cancer PCAWG cohort (C-index = 0.7090)—demonstrated that MoESurv possesses the most robust predictive per-formance, highlighting its generalizability. MoESurv also effectively stratified high- and low-risk patient groups and identified potential survival-associated genes, demonstrating both clinical utility and biological interpretability. Availability The code is freely available at https://github.com/HuaYC666/MoESurv and https://zenodo.org/records/20785500.

When Does Deep Representation Learning Help Single-Cell Clustering? A Sensitivity-Aware Diagnostic Benchmark for Biomedical AI Pipelines

Per-dataset analysis reveals three reproducible regimes: probabilistic variational autoencoder variants help on the smallest datasets, deep autoencoders win on mid-scale data with multi-batch or many-type structure, and classical PCA pipelines remain competitive when linear projection already captures the dominant variation.

A cfDNA fragmentomics-based liquid biopsy assay for early detection of oral cancer

Early detection of oral cancer improves outcomes, but remains limited by the invasiveness and low sensitivity of current screening methods. This study included 342 participants (200 in the training and 142 in the validation cohorts). Plasma cfDNA underwent low-depth whole-genome sequencing, with data normalized to a standardized 5× coverage via down-sampling to ensure consistent analysis. Three cfDNA-derived features, including fragment size ratio (FSR), copy number variation (CNV), and repetitive element profiles (REP), were extracted and integrated using a stacked machine learning ensemble to generate prediction scores. Model performance was evaluated by area under the receiver operating characteristic curve (AUC), sensitivity, and specificity with 95% confidence intervals. In the training cohort, oral cancer patients had a median age of 53 years (91% male), while healthy individuals had a median age of 57 years (37% male). In the validation cohort, the median ages were 53 and 57 years, with male proportions of 83.1% and 38.0% in the cancer and healthy groups, respectively. Cancer samples exhibited shorter cfDNA fragments, recurrent CNV gains at 3q and 8q, and losses at 3p and 5q. The individual AUCs for FSR, CNV, and REP were 0.973, 0.983, and 0.979 in the training cohort, and 0.960, 0.982, and 0.972 in the validation cohort, respectively. The integrated stacked model achieved AUCs of 0.996 (95% CI: 0.991-1.000) and 0.988 (95% CI: 0.975–0.988) in the training and validation cohorts, respectively. To prioritize high specificity for large-scale screening, a decision threshold was established in the training cohort to target 98.0% specificity, which yielded a corresponding sensitivity of 98.0%. When this pre-fixed threshold was blindly applied to the independent validation cohort, a sensitivity of 94.4% and a specificity of 95.8% were achieved. Prediction scores were not significantly associated with age, sex, or tumor site. The cfDNA fragmentomics-based stacked model distinguished oral cancer from healthy individuals, including early stage disease, supporting its potential for noninvasive approach for oral cancer detection.

Explainable AI for analyzing cancer outcomes using large-scale genome sequencing data

A multi-tier, explainable AI framework designed to risk-stratify patients and predict overall survival using clinical and genomic covariates is developed and demonstrates that explainable machine learning models can robustly predict survivability and highlight actionable features for oncology dashboards.