多模态或多组学数据构建机器学习模型，用于消化疾病诊疗

Pancreatic ductal adenocarcinoma (PDAC) remains a highly lethal disease with limited tools for predicting treatment response or survival. Our prior work, which applied machine learning to the COMPASS trial (NCT02750657), demonstrated that multimodal integration can enhance performance. Here, we evaluate model generalizability on an independent cohort from the PASS-01 clinical trial (NCT04469556). We developed predictive models using data from the COMPASS trial, incorporating clinical variables, histopathology image features, radiology-derived imaging features, RNA sequencing (RNA-seq), and whole-genome sequencing (WGS). Our updated pipeline applied TabPFN, a transformer-based model, using repeated 5-fold cross-validation. Fusion approaches included both early and late modality integration. We focused on predicting disease control rate (DCR). We compared model performance to the PurIST RNA-seq classifier, a strong baseline. The area under the receiver operating characteristic curve (AUC) was the primary metric. We externally validated the performance of models trained on COMPASS in the PASS-01 trial dataset. Among unimodal models, RNA-seq-based predictors achieved the highest AUC at 0.709 (95% CI: 0.595-0.820), significantly outperforming PurIST (p = 0.01). The performance of other unimodal models varied (clinical: 0.680; DNA: 0.527), with no significant difference compared to PurIST. The late fusion model, “MULTIPL”, integrated clinical, RNA, and DNA modalities and achieved the best overall performance at 0.733 (95% CI: 0.613-0.832), significantly outperforming PurIST (p = 0.002). The top 25th percentile of patients based on MULTIPL predicted DCR had significantly better prognosis (median overall survival 13.9 versus 8.6 months, hazard ratio 0.47 (95% CI: 0.28-0.78). The probability of DCR predicted by MULTIPL was correlated with the PurIST predictions of basal and classical transcriptomic subtypes (r = 0.63, p < 0.001), indicating a shared biology, which was further evidenced with SHapley Additive exPlanation interpretability analyses. Nonetheless, MULTIPL captured additional prognostic information, since PurIST was only modestly associated with DCR (AUC 0.55) and not significantly prognostic. Furthermore, the multimodal model was significantly associated with survival within the classical transcriptomic subtype. In conclusion, multimodal models trained on COMPASS data generalized to the PASS-01 trial in external validation. Late fusion of clinical, RNA, and DNA features achieved the best predictive performance for DCR and was also associated with survival outcomes, including within each transcriptomic subtype. In contrast to other models, which typically identify poor prognostic subgroups such as basal-like cancers, our multimodal model for DCR identifies a subset of patients with a more favourable prognosis. Together, these results demonstrate the potential of multimodal machine learning to improve prognostic modeling in advanced pancreatic cancer and guide personalized treatment strategies. Wei Quan, David Henault, Amy Zhang, Gun Ho Jang, Nicholas Light, Zongliang Ji, Anna Dodd, Julie Wilson, Daniel Renouf, Daniel Laheru, Kenneth Yu, Kimberly Perez, Amber Habowski, Grainne M. O'Kane, Steven Gallinger, David Tuveson, Elizabeth Jaffee, Jennifer J. Knox, Rahul G. Krishnan, Sandra Fischer, Masoom A. Haider, Faiyaz Notta, Robert C. Grant. External validation of a multimodal machine learning system to predict outcomes in advanced pancreatic cancer in the PASS-01 trial [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Advances in Pancreatic Cancer Research—Emerging Science Driving Transformative Solutions; Boston, MA; 2025 Sep 28-Oct 1; Boston, MA. Philadelphia (PA): AACR; Cancer Res 2025;85(18_Suppl_3):Abstract nr B086.

Summary Despite the encouraging efficacy of anti-PD-1/PD-L1 immunotherapy in microsatellite-instability-high/deficient mismatch repair (MSI-H/dMMR) advanced gastrointestinal cancer, many patients exhibit primary or acquired resistance. Using multi-omics approaches, we interrogate gut microbiome, blood metabolome, and cytokines/chemokines of patients with MSI-H/dMMR gastrointestinal cancer (N = 77) at baseline and during the treatment. We identify a number of microbes (e.g., Porphyromonadaceae) and metabolites (e.g., arginine) highly associated with primary resistance to immunotherapy. An independent validation cohort (N = 39) and mouse model are used to further confirm our findings. A predictive machine learning model for primary resistance is also built and achieves an accuracy of 0.79 on the external validation set. Furthermore, several microbes are pinpointed that gradually changed during the process of acquired resistance. In summary, our study demonstrates the essential role of gut microbiome in drug resistance, and this can be utilized as a preventative diagnosis tool and therapeutic target in the future.

The gut microbiome is involved in human health and disease, and its comprehensive understanding is necessary to exploit it as a diagnostic or therapeutic tool. Multi-omics approaches, including metagenomics, metatranscriptomics, metabolomics, and metaproteomics, enable depicting the complexity of the gut microbial ecosystem. However, these tools generate a large data stream, which integration is needed to produce clinically useful readouts but, in turn, might be difficult to carry out with conventional statistical methods. Artificial intelligence and machine learning have been increasingly applied to multi-omics datasets in several conditions associated with microbiome disruption, from chronic disorders to cancer. Such tools show potential for clinical implementation, including the discovery of microbial biomarkers for disease classification or prediction, the prediction of response to specific treatments, the fine-tuning of microbiome-modulating therapies. Here we discuss the state of the art, potential, and limits, of artificial intelligence and machine learning in the multi-omics approach to gut microbiome.

Alpha‐fetoprotein‐producing gastric cancer (AFPGC) is a rare but highly aggressive subtype of gastric cancer. Patients with AFPGC are at high risk of liver metastasis, and the tumor microenvironment (TME) is complex. A multicenter retrospective study is conducted from January 2011 to December 2021 and included 317 AFPGC patients. Using a multivariable logistic regression model, a nomogram for predicting liver metastasis is built. By combining AFP and the neutrophil–lymphocyte ratio (NLR), we developed a novel and easily applicable predictive indicator, termed ANLiM score, for liver metastasis in AFPGC. An integrated multi‐omics analysis, including whole‐exome sequencing and proteomic analysis, is conducted and revealed an immunosuppressive TME in AFPGC with liver metastasis. Single‐cell RNA sequencing and multiplex immunofluorescence identified the potential roles of tumor‐associated neutrophils and tertiary lymphoid structures in shaping the immune microenvironment. These findings are validated in a real‐world cohort receiving anti‐programmed cell death 1 (anti‐PD‐1) therapy, which showed concordant effectiveness. In addition, the ANLiM score is also identified as a promising biomarker for predicting immunotherapy efficacy. Overall, a blood biomarker‐based predictive indicator is developed for liver metastasis and immunotherapy response in AFPGC. The findings on immune microenvironmental alterations for AFPGC with liver metastasis provide new insights for optimizing immunotherapy strategies.

The development of antibody-drug conjugates (ADCs) requires reliable tumor-associated antigens (TAAs) expression for efficacy, necessitating predictive preclinical models. Patient-derived xenografts (PDXs) preserve patient tumor characteristics, which serve as a mainstay in translational oncology research, while patient-derived organoids (PDOs) and PDX-derived organoids (PDXOs) have recently emerged as powerful in vitro 3D systems that offer enhanced scalability while retaining key biological features of the original tissue. However, their fidelity in maintaining TAA profiles requires multi-omics validation. This study evaluates TAA consistency across platforms and between PDXs and matched organoids. We analyzed 18 clinically relevant TAAs using IHC on ∼1000 PDX models across 18 cancer types. IHC quantification was analyzed using HALO AI platform to generate H-Score, this data was integrated with RNA-seq and MS-mass spectra-proteomics from Crown Bioscience’s database to determine the correlation coefficients. To assess model translatability, a focused panel of 10 key TAAs was selected for IHC assessment between a subset of ∼400 characterized PDXs and their paired PDOs/PDXOs, enabling a cross-model comparison. Our integrated multi-omics analysis within the extensive PDXs cohort demonstrated a high degree of concordance between protein expression (H-Score) and both transcriptomics and proteomics data for the 18 investigated TAAs, including HER2(RRNAseq=0.871,RProteomics=0.765), TROP2(RRNAseq=0.852, RProteomics=0.775), Nectin-4(RRNAseq=0.679, RProteomics=0.861), DLL3(RRNAseq=0.75, RProteomics=0.698), CEACAM5(RRNAseq=0.799, RProteomics=0.743). Critically, a remarkably high degree of concordance was observed in TAAs protein expression patterns between PDXs and their paired PDOs/PDXOs models, including TROP2(R=0.946, P<0.0001), Nectin-4(R=0.772, P<0.0001), DLL3(R=0.819, P<0.0001), HER3(R=0.659, P<0.0001), Claudin 18.2(R=0.775, P<0.0001). The consistency of IHC intensity and heterogeneity characteristics between organoids and their in vivo counterparts further supports this molecular fidelity. This study validated TAAs expression concordance across multi-omics platforms in a large PDX cohort. More significantly, we deliver compelling evidence that PDOs/PDXOs models exhibit exceptional fidelity in maintaining the TAA expression landscape of their corresponding PDX tumors, demonstrating PDOs/PDXOs as highly reliable and invaluable tools from initial target validation and lead antibody characterization to the formulation of biomarker-driven patient selection strategies in clinical trials. Xiaolong Tu, Likun Zhang, Jie Lin, Hengyuan Liu, Jun Zhou, Marrit Putker, Ludovic Bourre, Julie Myer. Integrated multi-omics analysis reveals conserved tumor-associated antigens (TAAs) profiles in PDX and organoid models for advancing ADC development [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 7540.

Modern multimodal oncology foundation models (OFMs) can predict patient trajectories and simulate treatment effects from clinical, genomic, transcriptomic, and histologic data, but they remain largely black boxes: they rarely explain which mechanisms drive risk or how to modulate those mechanisms with feasible interventions. We trained a multimodal OFM on over 1.2 million cancer patients with longitudinal clinical records, tumor DNA and RNA sequencing, and H&E imaging. The model learns a joint latent representation of each patient’s evolving disease state and can forecast outcomes and simulate counterfactual treatment trajectories. We then added a mechanism-aware reasoning layer that turns the OFM from a predictor into an engine for mechanism-based therapy design. Our key innovation is group-relative policy optimization (GRPO), a reinforcement-learning framework that links mechanism states inferred by the OFM (e.g., pathway activation patterns and resistance programs), a drug intervention space built from curated drug-target relationships and perturbation signatures, and an explicit clinical reward. Rather than optimizing an abstract objective for a single patient, we define the reward as improvement in predicted outcomes for cohorts of patients who share similar disease state and driver mechanisms. For each disease context, we identify mechanisms associated with poor outcome, link them to candidate drugs and observed outcomes in real-world data, and then use GRPO to evaluate policies (sets of mechanism-level interventions) by asking whether applying a policy to that cohort improves predicted survival or delays progression relative to matched standard-of-care controls. This group-relative reward stabilizes learning, avoids overfitting to idiosyncratic outliers, and aligns the learned policies with how clinicians naturally reason about “patients like these.” To make outputs biologically and clinically interpretable, we map proposed mechanism shifts to existing drugs and combinations, to mechanism-defined patient clusters whose outcomes are driven by similar latent programs, and to de novo mechanism opportunities where the optimal policy improves outcomes but no current drug fully explains the effect, flagging potential targets for discovery. Across multiple indications, GRPO recovers known mechanism-therapy relationships, identifies patient subgroups whose outcomes are improved when therapies they receive align with GRPO-suggested mechanism policies compared with matched patients receiving discordant therapies, and proposes novel mechanism-therapy hypotheses involving combined or sequential modulation of programs that are not jointly targeted today. This turns a large-scale OFM into a mechanism-aware decision engine that reasons about resistance and treatment opportunities in the same latent space used for outcome prediction. Eric E. Schadt, Jackson Stokes, Lihua Zhao, Jason Chin, Jonathan Tyler, Lijia Sun, Lauren Beck, Duo Xu, Aviva G. Beckmann, Iker Huerga, . Mechanism-aware cancer therapy planning with group-relative policy optimization on a multimodal oncology foundation model [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 1467.

Precision oncology seeks to tailor cancer treatments to individual patients. Patient-derived xenografts (PDXs) have emerged as a promising platform for selecting efficacious, personalized therapies and developing new oncology drugs. Here, we present a workflow for selecting and validating treatments for patients with colorectal cancer (CRC) that integrates orthotopic PDX (O-PDX) models, molecular profiling, machine learning, and in vivo pharmacological validation. Of the nine tumor biopsies collected, seven were successfully developed into O-PDX models (78%) with a median time to establishment of 119 days. Models were molecularly profiled for gene expression and were serially passaged to perform orthotopic pharmacology studies to validate test agents selected by oncologists as well those suggested by CertisAITM, a novel ensemble of machine learning models for the prediction of drug response in human cancers. The correlation between CertisAI's predicted therapy responses and the actual observed tumor growth inhibition (TGI) was r = 0.7, encompassing 37 distinct treatments across studies from six patients. This research collectively illustrates that the integration of artificial intelligence with functional in vivo assays offers a powerful platform for precision oncology, enabling the identification and validation of tailored treatments for individual cancer patients. Citation Format: Fernando Eguiarte-Solomon, Rowan Prendergast, Bianca Carapia, Javier Rodriguez, Kristen Buck, Derrick Gorospe, Elizabeth Valencia, Jose Lopez-Ramos, Itzel Gutierrez, Rafaella Pippa, Yuan-Hung Chien, Warren Andrews, Long Do, Jantzen Sperry, Jonathan K. Nakashima. Integrating artificial intelligence and functional precision oncology for individualized cancer therapies [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 4971.

External beam radiotherapy (EBRT)-based combination therapy yields heterogeneous survival outcomes in unresectable hepatocellular carcinoma (uHCC), underscoring the need for precise prognostic stratification. We conducted a multicenter retrospective study across six institutions, enrolling 875 uHCC patients treated with either EBRT combined with systemic therapy (ES cohort, n=383) or EBRT combined with transarterial chemoembolization (TACE) and systemic therapy (ETS cohort, n=492). After propensity score matching, median overall survival was significantly prolonged in the ETS cohort compared to the ES cohort (24.0 vs. 19.0 months; HR = 0.73, P < 0.0001). The multimodal deep learning model, TRIM-uHCC (transformer-based risk-stratification integrated multimodal model for uHCC), was developed to stratify patients into high-, intermediate-, and low-risk groups. Prognostic performance was compared with current guideline-based staging systems (BCLC/CNLC/AJCC-TNM) and deep learning models (Swin-Transformer/ViT/ResNet50/ResNeXt50) using the C-index and time-dependent AUC. TRIM-uHCC model showed significantly superior prognostic prediction performance compared to current guideline standards (C-indices: 0.71-0.79 vs. 0.51-0.61, all P < 0.0001) and deep learning models (C-indices: vs. 0.62-0.75, P < 0.0001-0.106) in the ETS and ES cohorts. Based on TRIM-uHCC, 8.8% (29/331) of patients in the ES cohort could potentially achieve improved survival by adjusting to ETS, whereas 7.9% (26/331) of patients in the ETS cohort were recommended to switch to ES treatment. Collectively, the TRIM-uHCC model offers more accurate individualized prognostic stratification than current guideline standards and other deep learning models, providing valuable decision-making support for EBRT-based combination therapies.

ADAPT is a nationwide initiative to transform cancer care by detecting and responding to tumor evolution in real time. Integrating multimodal data, interpretable AI, and an evolutionary clinical trial platform, ADAPT predicts emerging resistance traits and guides treatment adjustments as tumors change. A unified national infrastructure enables continuous learning across patients, linking discovery directly to care. By making therapy responsive to tumor changes, ADAPT delivers a scalable model designed to improve outcomes in precision oncology.

No clinically useful non-invasive biomarkers have been developed for diagnosis of chronic pancreatitis (CP), and molecular features of CP have not been characterised. Extracellular vesicles (EVs) consisted of abundant RNA species with specialised functions and clinical applications. Our study aimed to construct a diagnostic model for CP and depict molecular landscape of CP based on EV long RNA (ExLR). Candidate ExLRs were defined using prespecified expression-quality criteria and complementary discovery-stage screens, and a resampling-based consensus feature selection in the training cohort yielded a five-ExLRs panel for model construction. The ExLRs-based CP diagnostic model (ExLRCPdscore) was further confirmed in another two independent validation cohorts with different controls. To elucidate the biological architecture of CP through ExLR profiling, we integrated ExLR-seq, single-cell data and clinical information. ExLRCPdscore constructed by random forest demonstrated excellent performance for detecting CP. Importantly, ExLRCPdscore could effectively detect early-stage CP, CP without alarm symptoms, CP without significant imaging findings and CP without risk factors. Using ExLR profiling and phenotypic data, we pinpointed MUC5B + ductal cells exhibiting the strongest correlation with CP and derived an ExLR-based acinar-to-ductal metaplasia (ADM) score as a blood-based transcriptomic proxy of ADM-related programme. Integration of ExLR-seq and clinical information revealed significant associations between ADMscore and clinical characteristics, imaging findings and metabolic sequelae. Our study is the first to report an ExLRs-based diagnostic model that demonstrates exceptional robustness in differentiating CP from healthy controls and non-pancreatic disease controls. ExLRs offer a promising tool for CP molecular characterisation and pathophysiological quantification.

The malignant potential of pancreatic cystic lesions (PCLs) varies dramatically, leading to difficulties when making clinical decisions. This study aimed to develop noninvasive clinical‐radiomic models using preoperative CT images to predict the malignant potential of PCLs. It also investigates the biological mechanisms underlying these models. Patients from two retrospective and one prospective cohort, all undergoing surgical resection for PCLs, are divided into four datasets: training, internal test, external test, and prospective application sets. Eleven machine learning classifiers are employed to construct radiomic models based on selected features. Cyst fluid from the prospective cohort is collected for proteomic and lipidomic analysis. The radiomic models demonstrated high accuracy, with area under the receiver operating characteristic curves (AUCs) > 0.93 across the training (n = 262), internal test (n = 50), and external test (n = 50) sets. AUCs ranged from 0.92 to 0.96 for the prospective cohort (n = 34). Meanwhile, differentially‐expressed proteins and lipid molecules, along with their associated signaling pathways, are identified between high and low groups of clinical‐radiomic scores. This models can effectively and accurately predict the malignant potential of PCLs, with multi‐omics evidence suggesting the biological mechanisms involving secretion function and lipid metabolism underlying clinical‐radiomic models.

Current first-line treatments of paediatric ulcerative colitis (UC) maintain a 6-month remission in only half of the patients. Relapse prediction at diagnosis could enable earlier introduction of immunosuppressants. We collected intestinal biopsies from 56 treatment-naïve children, combining mucosal quantitative microbial profiling with host epigenomics, transcriptomics, genotyping, and in vitro and in vivo experiments on selected bacteria. Baseline bacterial diversity is lower in relapsing children, who have fewer butyrate producers but more oral-associated bacteria, whereof Veillonella parvula induces inflammation in epithelial cell lines and IL10−/− mice. Microbiota has the strongest association with future relapse, followed by host epigenome and transcriptome. Interferon gamma signalling is also linked to relapse-associated bacteria. Relapse-prediction using separate omics data is outperformed by a robust machine learning approach combining microbiomes and epigenomes. In summary, host-microbe data have prognostic potential in paediatric UC. Our translational findings also suggest that pro-inflammatory oral-associated colonizers can exploit the reduced colonic bacterial diversity of relapsing children. Current first-line treatments of pediatric UC maintain a 6-month remission in only half of the patients. Here, applying multi-omics on intestinal biopsies from treatment-naïve children, the authors show that relapse-prediction using separate omics data is outperformed by a robust machine learning approach combining microbiomes and epigenomes.

The sole use of single modality data often fails to capture the complex heterogeneity among patients, including the variability in resistance to anti-HER2 therapy and outcomes of combined treatment regimens, for the treatment of HER2-positive gastric cancer (GC). This modality deficit has not been fully considered in many studies. Furthermore, the application of artificial intelligence in predicting the treatment response, particularly in complex diseases such as GC, is still in its infancy. Therefore, this study aimed to use a comprehensive analytic approach to accurately predict treatment responses to anti-HER2 therapy or anti-HER2 combined immunotherapy in patients with HER2-positive GC. We collected multi-modal data, comprising radiology, pathology, and clinical information from a cohort of 429 patients: 310 treated with anti-HER2 therapy and 119 treated with a combination of anti-HER2 and anti-PD-1/PD-L1 inhibitors immunotherapy. We introduced a deep learning model, called the Multi-Modal model (MuMo), that integrates these data to make precise treatment response predictions. MuMo achieved an area under the curve score of 0.821 for anti-HER2 therapy and 0.914 for combined immunotherapy. Moreover, patients classified as low-risk by MuMo exhibited significantly prolonged progression-free survival and overall survival (log-rank test, P < 0.05). These findings not only highlight the significance of multi-modal data analysis in enhancing treatment evaluation and personalized medicine for HER2-positive gastric cancer, but also the potential and clinical value of our model.

Gallbladder cancer (GBC) frequently mimics gallbladder benign lesions (GBBLs) in radiological images, leading to preoperative misdiagnoses. To address this challenge, we initiated a prospective, multicenter clinical trial (ChicCTR2100049249) and proposed a multimodal, non-invasive diagnostic model to distinguish GBC from GBBLs. A total of 301 patients diagnosed with gallbladder-occupying lesions (GBOLs) from 11 medical centers across 7 provinces in China were enrolled and divided into a discovery cohort and an independent external validation cohort. An artificial intelligence (AI)-based integrated model, GBCseeker, is created using cell-free DNA (cfDNA) genetic signatures, radiomic features, and clinical information. It achieves high accuracy in distinguishing GBC from GBBL patients (93.33% in the discovery cohort and 87.76% in the external validation cohort), reduces surgeons' diagnostic errors by 56.24%, and reclassifies GBOL patients into three categories to guide surgical options. Overall, our study establishes a tool for the preoperative diagnosis of GBC, facilitating surgical decision-making.

Immunotherapy has become a cornerstone in the treatment of advanced gastric cancer (GC). However, identifying reliable predictive biomarkers remains a considerable challenge. This study demonstrates the potential of integrating multimodal baseline data, including computed tomography scan images and digital H&E-stained pathology images, with biological interpretation to predict the response to immunotherapy-based combination therapy using a multicenter cohort of 298 GC patients. By employing seven machine learning approaches, we developed a radiopathomics signature (RPS) to predict treatment response and stratify prognostic risk in GC. The RPS demonstrated area under the receiver-operating-characteristic curves (AUCs) of 0.978 (95 % CI, 0.950-1.000), 0.863 (95 % CI, 0.744-0.982), and 0.822 (95 % CI, 0.668-0.975) in the training, internal validation, and external validation cohorts, respectively, outperforming conventional biomarkers such as CPS, MSI-H, EBV, and HER-2. Kaplan-Meier analysis revealed significant differences of survival between high- and low-risk groups, especially in advanced-stage and non-surgical patients. Additionally, genetic analyses revealed that the RPS correlates with enhanced immune regulation pathways and increased infiltration of memory B cells. The interpretable RPS provides accurate predictions for treatment response and prognosis in GC and holds potential for guiding more precise, patient-specific treatment strategies while offering insights into immune-related mechanisms.

Tertiary lymphoid structures (TLS) are critical components of the tumor microenvironment in gastric cancer, but clinical assessment of TLSs is challenging. The development of automated annotation tools for histopathologic slide analysis could facilitate the identification of TLSs and enhance our understanding of the mechanisms driving TLS maturation. In this study, we generated a transformer-based deep learning model that enables quantitative characterization of TLS maturity from whole-slide images. Application of the model to a large gastric cancer cohort (n = 253) showed that higher TLS maturity correlated with improved patient survival. Integration of single-cell RNA sequencing data from 17 patients with gastric cancer combined with multiplex IHC, flow cytometry, and functional coculture assays identified a key immune circuit in mature TLSs involving CD8+ tissue-resident memory T cells, which recruit activated B cells via the CXCL13-CXCR5 axis to enhance tissue-resident memory T-cell cytotoxicity through granzyme B upregulation. Overall, this study established a clinically applicable artificial intelligence tool and uncovered key immune interactions that regulate TLS maturation and antitumor immunity in gastric cancer. A deep learning model demonstrates that higher tertiary lymphoid structure maturity predicts improved gastric cancer patient survival and identifies a key immune circuit, offering a clinically applicable tool that could guide treatment. This article is part of a special series: Driving Cancer Discoveries with Computational Research, Data Science, and Machine Learning/AI.

Background Deep learning (DL) models can potentially improve prognostication of rectal cancer but have not been systematically assessed. Purpose To develop and validate an MRI DL model for predicting survival in patients with rectal cancer based on segmented tumor volumes from pretreatment T2-weighted MRI scans. Materials and Methods DL models were trained and validated on retrospectively collected MRI scans of patients with rectal cancer diagnosed between August 2003 and April 2021 at two centers. Patients were excluded from the study if there were concurrent malignant neoplasms, prior anticancer treatment, incomplete course of neoadjuvant therapy, or no radical surgery performed. The Harrell C-index was used to determine the best model, which was applied to internal and external test sets. Patients were stratified into high- and low-risk groups based on a fixed cutoff calculated in the training set. A multimodal model was also assessed, which used DL model-computed risk score and pretreatment carcinoembryonic antigen level as input. Results The training set included 507 patients (median age, 56 years [IQR, 46-64 years]; 355 men). In the validation set (

Despite their promise, circulating tumor DNA (ctDNA)-based assays for multi-cancer early detection face challenges in test performance, due mostly to the limited abundance of ctDNA and its inherent variability. To address these challenges, published assays to date demanded a very high-depth sequencing, resulting in an elevated price of test. Herein, we developed a multimodal assay called SPOT-MAS (screening for the presence of tumor by methylation and size) to simultaneously profile methylomics, fragmentomics, copy number, and end motifs in a single workflow using targeted and shallow genome-wide sequencing (~0.55×) of cell-free DNA. We applied SPOT-MAS to 738 non-metastatic patients with breast, colorectal, gastric, lung, and liver cancer, and 1550 healthy controls. We then employed machine learning to extract multiple cancer and tissue-specific signatures for detecting and locating cancer. SPOT-MAS successfully detected the five cancer types with a sensitivity of 72.4% at 97.0% specificity. The sensitivities for detecting early-stage cancers were 73.9% and 62.3% for stages I and II, respectively, increasing to 88.3% for non-metastatic stage IIIA. For tumor-of-origin, our assay achieved an accuracy of 0.7. Our study demonstrates comparable performance to other ctDNA-based assays while requiring significantly lower sequencing depth, making it economically feasible for population-wide screening.

Despite the success of biological therapies in treating inflammatory bowel disease, managing patients remains challenging due to the absence of reliable predictors of therapy response. In this study, we prospectively sampled 2 cohorts of patients with inflammatory bowel disease receiving the anti-integrin α4β7 antibody vedolizumab. Samples were subjected to mass cytometry; single-cell RNA sequencing; single-cell B and T cell receptor sequencing (BCR/TCR-seq); serum proteomics; and multiparametric flow cytometry to comprehensively assess vedolizumab-induced immunologic changes in the peripheral blood and their potential associations with treatment response. Vedolizumab treatment led to substantial alterations in the abundance of circulating immune cell lineages and modified the T-cell receptor diversity of gut-homing CD4 These findings provide a reliable predictive classifier with significant implications for personalized inflammatory bowel disease management.

Accurate preoperative prediction of pathological complete response (pCR) following neoadjuvant chemoimmunotherapy (nCIT) could help individualize treatment for patients with esophageal squamous cell carcinoma (ESCC). This study aimed to develop and externally validate an interpretable multimodal machine learning framework that integrates CT radiomics and H&E-stained whole-slide images pathomics to predict pCR. In this multicenter, retrospective study, 335 patients with ESCC who received nCIT followed by esophagectomy were enrolled from three institutions. Patients from one center were divided into a training set (181 patients) and an internal test set (115 patients), while data from the other two centers comprised an external test set (39 patients). We developed unimodal radiomics and pathomics models, and two multimodal fusion models-an intermediate fusion model (MIFM) and a late fusion model (MLFM). Model performance was evaluated using the area under the curve (AUC), accuracy, sensitivity, specificity, and F1 score, with exploratory survival stratification by observed and model-predicted pCR status. Interpretability was treated as a design constraint and operationalized at both the feature and model levels. The MIFM outperformed unimodal models and the MLFM across all cohorts, achieving AUC/accuracy/sensitivity/specificity/F1 score of 0.97/0.93/0.84/0.96/0.86 (training set), 0.78/0.87/0.62/0.93/0.63 (internal test set), and 0.76/0.77/0.54/0.88/0.61 (external test set). Both observed and predicted pCR status showed exploratory prognostic stratification for overall survival. Feature definitions were mathematically or morphologically explicit, and case-level/cohort-level explanations together with decision-pathway views provided insights into model reasoning. We additionally provide a user-friendly Graphical User Interface to facilitate clinical practice. We developed and externally validated an interpretable radiopathomics fusion framework that predicts pCR after nCIT in ESCC using standard-of-care data. This model holds promise as an effective tool for guiding individualized decisions between surveillance and timely surgery.

Medical image classification requires labeled, task-specific datasets which are used to train deep learning networks de novo, or to fine-tune foundation models. However, this process is computationally and technically demanding. In language processing, in-context learning provides an alternative, where models learn from within prompts, bypassing the need for parameter updates. Yet, in-context learning remains underexplored in medical image analysis. Here, we systematically evaluate the model Generative Pretrained Transformer 4 with Vision capabilities (GPT-4V) on cancer image processing with in-context learning on three cancer histopathology tasks of high importance: Classification of tissue subtypes in colorectal cancer, colon polyp subtyping and breast tumor detection in lymph node sections. Our results show that in-context learning is sufficient to match or even outperform specialized neural networks trained for particular tasks, while only requiring a minimal number of samples. In summary, this study demonstrates that large vision language models trained on non-domain specific data can be applied out-of-the box to solve medical image-processing tasks in histopathology. This democratizes access of generalist AI models to medical experts without technical background especially for areas where annotated data is scarce.

Rome IV disorders of gut-brain interaction (DGBI) subtypes are known to be unstable and demonstrate high rates of non-treatment response, likely indicating patient heterogeneity. Cluster analysis, a type of unsupervised machine learning, can identify homogeneous sub-populations. Independent cluster analyses of symptom and biological data have highlighted its value in predicting patient outcomes. Integrated clustering of symptom and biological data may provide a unique multimodal perspective that better captures the complexity of DGBI. Here, integrated symptom and multi-omic cluster analysis was performed on a cohort of healthy controls and patients with lower-gastrointestinal tract DGBI. Cluster stability was assessed by considering how frequently pairs of participants appeared in the same cluster between different bootstrapped datasets. Functional enrichment analysis was performed on the biological signatures of stable DGBI-predominant clusters, implicating disrupted ammonia handling and metabolism as possible pathophysiologies present in a subset of patients with DGBI. Integrated clustering revealed subtypes that were not apparent using a singular modality, suggesting a symptom-only classification is prone to capturing heterogeneous sub-populations.

The potential of serum extracellular vesicles (EVs) as non-invasive biomarkers for diagnosing colorectal cancer (CRC) remains elusive. We employed an in-depth 4D-DIA proteomics and machine learning (ML) pipeline to identify key proteins, PF4 and AACT, for CRC diagnosis in serum EV samples from a discovery cohort of 37 cases. PF4 and AACT outperform traditional biomarkers, CEA and CA19-9, detected by ELISA in 912 individuals. Furthermore, we developed an EV-related random forest (RF) model with the highest diagnostic efficiency, achieving AUC values of 0.960 and 0.963 in the train and test sets, respectively. Notably, this model demonstrated reliable diagnostic performance for early-stage CRC and distinguishing CRC from benign colorectal diseases. Additionally, multi-omics approaches were employed to predict the functions and potential sources of serum EV-derived proteins. Collectively, our study identified the crucial proteomic signatures in serum EVs and established a promising EV-related RF model for CRC diagnosis in the clinic.

Gastric Cancer (GC) characteristically exhibits heterogeneous responses to treatment, particularly in relation to immuno plus chemo therapy, necessitating a precision medicine approach. This study is centered around delineating the cellular and molecular underpinnings of drug resistance in this context. We undertook a comprehensive multi-omics exploration of postoperative tissues from GC patients undergoing the chemo and immuno-treatment regimen. Concurrently, an image deep learning model was developed to predict treatment responsiveness. Our initial findings associate apical membrane cells with resistance to fluorouracil and oxaliplatin, critical constituents of the therapy. Further investigation into this cell population shed light on substantial interactions with resident macrophages, underscoring the role of intercellular communication in shaping treatment resistance. Subsequent ligand-receptor analysis unveiled specific molecular dialogues, most notably TGFB1-HSPB1 and LTF-S100A14, offering insights into potential signaling pathways implicated in resistance. Our SVM model, incorporating these multi-omics and spatial data, demonstrated significant predictive power, with AUC values of 0.93 and 0.84 in the exploration and validation cohorts respectively. Hence, our results underscore the utility of multi-omics and spatial data in modeling treatment response. Our integrative approach, amalgamating mIHC assays, feature extraction, and machine learning, successfully unraveled the complex cellular interplay underlying drug resistance. This robust predictive model may serve as a valuable tool for personalizing therapeutic strategies and enhancing treatment outcomes in gastric cancer.

Microsatellite instability (MSI) has been approved as a pan-cancer biomarker for immune checkpoint blockade (ICB) therapy. However, current MSI identification methods are not available for all patients. We proposed an ensemble multiple instance deep learning model to predict microsatellite status based on histopathology images, and interpreted the pathomics-based model with multi-omics correlation.

Immune checkpoint blockade (ICB) therapy has revolutionized cancer treatment but remains effective in only a subset of patients. Emerging evidence suggests that the gut microbiome and its metabolites critically influence ICB efficacy. In this study, we performed a multi-omics analysis of fecal microbiomes and metabolomes from 165 patients undergoing anti-programmed cell death protein 1 (PD-1)/programmed death ligand 1 (PD-L1) therapy, identifying microbial and metabolic entities associated with treatment response. Integration of data from four public metagenomic datasets (n = 568) uncovered cross-cohort microbial and metabolic signatures, validated in an independent cohort (n = 138). An integrated predictive model incorporating these features demonstrated robust performance. Notably, we characterized five response-associated enterotypes, each linked to specific bacterial taxa and metabolites. Among these, the metabolite phenylacetylglutamine (PAGln) was negatively correlated with response and shown to attenuate anti-PD-1 efficacy in vivo. This study sheds light on the interplay among the gut microbiome, the gut metabolome, and immunotherapy response, identifying potential biomarkers to improve treatment outcomes.

The complex etiological factors associated with metabolic dysfunction-associated fatty liver disease (MAFLD), including perturbed iron homeostasis, and the unclear nature by which they contribute to disease progression have resulted in a limited number of effective therapeutic interventions. Here, we report that patients with metabolic dysfunction-associated steatohepatitis (MASH), a pathological subtype of MAFLD, exhibit excess hepatic iron and that it has a strong positive correlation with disease progression. FerroTerminator1 (FOT1) effectively reverses liver injury across multiple MASH models without notable toxic side effects compared with clinically approved iron chelators. Mechanistically, our multi-omics analyses reveal that FOT1 concurrently inhibits hepatic iron accumulation and c-Myc-Acsl4-triggered ferroptosis in various MASH models. Furthermore, MAFLD cohort studies suggest that serum ferritin levels might serve as a predictive biomarker for FOT1-based therapy in MASH. These findings provide compelling evidence to support FOT1 as a promising novel therapeutic option for all stages of MAFLD and for future clinical trials.

Despite recent progress in our understanding of the association between the gut microbiome and colorectal cancer (CRC), multi-kingdom gut microbiome dysbiosis in CRC across cohorts is unexplored. We investigated four-kingdom microbiota alterations using CRC metagenomic datasets of 1,368 samples from 8 distinct geographical cohorts. Integrated analysis identified 20 archaeal, 27 bacterial, 20 fungal and 21 viral species for each single-kingdom diagnostic model. However, our data revealed superior diagnostic accuracy for models constructed with multi-kingdom markers, in particular the addition of fungal species. Specifically, 16 multi-kingdom markers including 11 bacterial, 4 fungal and 1 archaeal feature, achieved good performance in diagnosing patients with CRC (area under the receiver operating characteristic curve (AUROC) = 0.83) and maintained accuracy across 3 independent cohorts. Coabundance analysis of the ecological network revealed associations between bacterial and fungal species, such as Talaromyces islandicus and Clostridium saccharobutylicum. Using metagenome shotgun sequencing data, the predictive power of the microbial functional potential was explored and elevated D-amino acid metabolism and butanoate metabolism were observed in CRC. Interestingly, the diagnostic model based on functional EggNOG genes achieved high accuracy (AUROC = 0.86). Collectively, our findings uncovered CRC-associated microbiota common across cohorts and demonstrate the applicability of multi-kingdom and functional markers as CRC diagnostic tools and, potentially, as therapeutic targets for the treatment of CRC.

Irritable bowel syndrome (IBS) is a common gastrointestinal disorder that is thought to involve alterations in the gut microbiome, but robust microbial signatures have been challenging to identify. As prior studies have primarily focused on composition, we hypothesized that multi-omics assessment of microbial function incorporating both metatranscriptomics and metabolomics would further delineate microbial profiles of IBS and its subtypes. Fecal samples were collected from a racially/ethnically diverse cohort of 495 subjects, including 318 IBS patients and 177 healthy controls, for analysis by 16S rRNA gene sequencing (n = 486), metatranscriptomics (n = 327), and untargeted metabolomics (n = 368). Differentially abundant microbes, predicted genes, transcripts, and metabolites in IBS were identified by multivariate models incorporating age, sex, race/ethnicity, BMI, diet, and HAD-Anxiety. Inter-omic functional relationships were assessed by transcript/gene ratios and microbial metabolic modeling. Differential features were used to construct random forests classifiers. IBS was associated with global alterations in microbiome composition by 16S rRNA sequencing and metatranscriptomics, and in microbiome function by predicted metagenomics, metatranscriptomics, and metabolomics. After adjusting for age, sex, race/ethnicity, BMI, diet, and anxiety, IBS was associated with differential abundance of bacterial taxa such as Bacteroides dorei; metabolites including increased tyramine and decreased gentisate and hydrocinnamate; and transcripts related to fructooligosaccharide and polyol utilization. IBS further showed transcriptional upregulation of enzymes involved in fructose and glucan metabolism as well as the succinate pathway of carbohydrate fermentation. A multi-omics classifier for IBS had significantly higher accuracy (AUC 0.82) than classifiers using individual datasets. Diarrhea-predominant IBS (IBS-D) demonstrated shifts in the metatranscriptome and metabolome including increased bile acids, polyamines, succinate pathway intermediates (malate, fumarate), and transcripts involved in fructose, mannose, and polyol metabolism compared to constipation-predominant IBS (IBS-C). A classifier incorporating metabolites and gene-normalized transcripts differentiated IBS-D from IBS-C with high accuracy (AUC 0.86). IBS is characterized by a multi-omics microbial signature indicating increased capacity to utilize fermentable carbohydrates-consistent with the clinical benefit of diets restricting this energy source-that also includes multiple previously unrecognized metabolites and metabolic pathways. These findings support the need for integrative assessment of microbial function to investigate the microbiome in IBS and identify novel microbiome-related therapeutic targets. Video Abstract.

Chronic hepatic injury and inflammation from various causes can lead to fibrosis and cirrhosis, potentially predisposing to hepatocellular carcinoma. The molecular mechanisms underlying fibrosis and its progression remain incompletely understood. Using a proteo-transcriptomics approach, we analyze liver and plasma samples from 330 individuals, including 40 healthy individuals and 290 patients with histologically characterized fibrosis due to chronic viral infection, alcohol consumption, or metabolic dysfunction-associated steatotic liver disease. Our findings reveal dysregulated pathways related to extracellular matrix, immune response, inflammation, and metabolism in advanced fibrosis. We also identify 132 circulating proteins associated with advanced fibrosis, with neurofascin and growth differentiation factor 15 demonstrating superior predictive performance for advanced fibrosis(area under the receiver operating characteristic curve [AUROC] 0.89 [95% confidence interval (CI) 0.81-0.97]) compared to the fibrosis-4 model (AUROC 0.85 [95% CI 0.78-0.93]). These findings provide insights into fibrosis pathogenesis and highlight the potential for more accurate non-invasive diagnosis.

The progression from metabolic dysfunction-associated steatotic liver disease (MASLD) to metabolic dysfunction-associated steatohepatitis (MASH) is a critical link leading to cirrhosis and hepatocellular carcinoma. Yet the responsible cellular programs remain unclear. We integrated public single-cell, spatial, and bulk transcriptomic datasets to map microenvironmental remodeling and regulatory networks during MASLD-MASH progression. Among the seven major liver cell types identified, monocytes/macrophages and hepatic stellate cells (HSCs) were significantly enriched and demonstrated spatial co-localization within the context of MASH. We identified a DTNA+distinct macrophage subpopulation that was specifically enriched in MASH. This subpopulation exhibited characteristics consistent with M2 polarization, hypoxia, and enhanced inflammatory signaling. Pseudotime trajectory analysis revealed that this state represents a differentiation pathway originating from Kupffer cells to the DTNA+ state. RUNX2 emerged as the key transcriptional regulator. Cell communication analysis demonstrated that DTNA+ macrophages potentially interact with activated HSCs via the RUNX2-PLG-PARD3 axis, contributing to the exacerbation of liver fibrosis. Finally, ensemble machine learning models (mean AUC = 0.839), identified DTNA as the optimal predictive biomarker for distinguishing MASLD from MASH. This study highlight DTNA+ macrophages and the RUNX2-PLG-PARD3 axis as candidate mechanisms and targets for non-invasive diagnosis and therapy in MASH.

Despite the effectiveness of infliximab in treating Crohn's disease (CD), up to 40 % of patients fail to respond adequately. This study aimed to identify predictive biomarkers of primary non-response to infliximab in treatment-naïve CD patients by characterizing baseline gut microbiome-metabolome interactions and to validate their mechanistic role in driving therapeutic resistance. In a prospective cohort of 100 CD patients initiating infliximab therapy and 49 healthy controls, we performed longitudinal 16S rRNA sequencing and untargeted metabolomics on pre-/post-treatment fecal samples. Machine learning (twelve algorithms including K-Nearest Neighbors, Linear Discriminant Analysis, Naive Bayes, and LightGBM) identified predictive microbial and metabolic features, with findings experimentally validated through fecal microbiota transplantation (FMT) in a murine TNBS-induced colitis model. Non-responders at baseline demonstrated significant microbial dysbiosis marked by β-diversity variation, depletion of Bifidobacterium, Blautia, and Lachnospiraceae, and enrichment of Escherichia/Shigella. Metabolomic profiling identified 179 differentially abundant metabolites, including deficiencies in taurochenodeoxycholic acid (TCDCA) and perturbations in glycerophospholipid metabolism and primary bile acid biosynthesis pathways. Among single-omics models, the microbiome-based Linear Discriminant Analysis achieved optimal performance (test AUC = 0.805), surpassing metabolomics-only (best AUC = 0.634) and integrated multi-omics approaches (best AUC = 0.779). SHAP analysis revealed Bifidobacterium as the dominant protective predictor, with its depletion strongly associated with non-response. Mechanistically, MIMOSA2 analysis linked Bifidobacterium catenulatum to TCDCA production, while FMT from non-responders exacerbated murine colitis through Treg depletion and M1 macrophage polarization, confirming microbiome-driven immune dysregulation. These findings establish gut microbiome composition, particularly Bifidobacterium abundance, as a critical determinant of anti-TNF response in CD, mediated through bile acid-dependent regulation of Treg/M1 macrophage homeostasis. While multi-omics integration did not enhance predictive performance, microbiome-based machine learning models offer clinically actionable biomarkers for treatment stratification, providing a roadmap for precision therapy to overcome biological resistance in inflammatory bowel disease.

Gut microbiota and related metabolites are both crucial factors that significantly influence how individuals with Crohn's disease respond to immunotherapy. However, little is known about the interplay among gut microbiota, metabolites, Crohn's disease, and the response to anti-α4β7-integrin in current studies. Our research utilized 2,4,6-trinitrobenzene sulfonic acid to induce colitis based on the humanized immune system mouse model and employed a combination of whole-genome shotgun metagenomics and non-targeted metabolomics to investigate immunotherapy responses. Additionally, clinical cases with Crohn's disease initiating anti-α4β7-integrin therapy were evaluated comprehensively. Particularly, 16S-rDNA gene high-throughput sequencing and targeted bile acid metabolomics were conducted at weeks 0, 14, and 54. We found that anti-α4β7-integrin therapy has shown significant potential for mitigating disease phenotypes in remission-achieving colitis mice. Microbial profiles demonstrated that not only microbial composition but also microbially encoded metabolic pathways could predict immunotherapy responses. Metabonomic signatures revealed that bile acid metabolism alteration, especially elevated secondary bile acids, was a determinant of immunotherapy responses. Especially, the remission mice significantly enriched the proportion of the beneficial

DNA methylation accumulates in non-malignant gastric mucosa after exposure to pathogens. To elucidate how environmental, methylation, and lifestyle factors interplay to influence primary gastric neoplasia (GN) risk, we analyzed longitudinally monitored cohorts in Japan and Singapore. Asymptomatic subjects who underwent a gastric mucosal biopsy on the health check-up were enrolled. We analyzed the association between clinical factors and GN development using Cox hazard models. We further conducted comprehensive methylation analysis on selected tissues, including (i) mucosae from subjects developing GN later, (ii) mucosae from subjects not developing GN later, and (iii) GN tissues and surrounding mucosae. We also use the methylation data of mucosa collected in Singapore. The association between methylation and GN risk, as well as lifestyle and methylation, were analyzed. Among 4234 subjects, GN was developed in 77 subjects. GN incidence was correlated with age, drinking, smoking, and Helicobacter pylori (HP) status. Accumulation of methylation in biopsied gastric mucosae was predictive of higher future GN risk and shorter duration to GN incidence. Whereas methylation levels were associated with HP positivity, lifestyle, and morphological alterations, DNA methylation remained an independent GN risk factor through multivariable analyses. Pro-carcinogenic epigenetic alterations initiated by HP exposure were amplified by unfavorable but modifiable lifestyle choices. Adding DNA methylation to the model with clinical factors improved the predictive ability for the GN risk. The integration of environmental, lifestyle, and epigenetic information can provide increased resolution in the stratification of primary GN risk. The funds are listed in Acknowledgements section.

Metabolic dysfunction-associated steatotic liver disease (MASLD) is a critical health concern, with metabolic dysfunction-associated steatohepatitis (MASH) representing a severe subtype that poses significant risks. This study aims to develop a robust diagnostic model for high-risk MASH utilizing a multi-omics approach. We initiated proteomic analysis to select differential proteins, followed by liver transcriptional profiling to localize these proteins. An intersection of differential proteins and liver-expressed genes facilitated the identification of candidate biomarkers. Subsequently, scRNA-seq data helped ascertain the subcellular localization of these biomarkers in kupffer cells. We then established two MASLD models to investigate the co-localization of F4/80 and the target proteins in Kupffer cells using immunofluorescence dual-labeling. Correlation analyses were performed using blood samples from a discovery cohort of 144 individuals with liver pathology to validate the relationships between candidate biomarkers and MASLD phenotypes. Using LASSO regression, we established the ABD-LTyG predictive model for high-risk MASH (NAS ≥ 4 + F ≥ 2) and validated its efficacy in an independent cohort of 171 individuals. Finally, we compared this model against three classic non-invasive liver fibrosis diagnostic methods. A proteo-transcriptomic comparison identified 58 consistent biomarkers in plasma and liver, with 25 closely associated with MASLD phenotype. Utilizing single-cell data and the HPA database, we delineated the localization of these biomarkers in liver cells, identifying TREM2, IL18BP, and LGALS3BP predominantly in the Kupffer cell subpopulation. Validation in animal models confirmed elevated expression and cellular localization of TREM2, IL18BP, and LGALS3BP in MASLD. To enhance diagnostic capability, we integrated clinical characteristics using LASSO regression to develop the ABD-LTyG model, comprising AST, BMI, total bilirubin (TB), vitamin D, TyG, and the biomarkers LGALS3BP and TREM2. This model demonstrated an AUC of 0.832 (95% CI 0.753-0.911) in the discovery cohort and 0.807 (95% CI 0.742-0.872) in the validation cohort for diagnosing high-risk MASH, outperforming traditional assessments such as FIB-4, NFS, and APRI. The integration of circulating biomarkers and clinical parameters into the ABD-LTyG model offers a promising approach for diagnosing high-risk MASH. This study underscores the importance of multi-omics strategies in enhancing diagnostic accuracy and guiding clinical decision-making.

This study aimed to identify potential therapeutic targets for sepsis and elucidate the underlying molecular mechanisms, with a particular focus on the liver as a key target organ for experimental validation. Here, two sequential studies were conducted to uncover and characterize therapeutic targets involved in sepsis-induced liver injury. In Study 1, a proteome-wide Mendelian randomization (MR) analysis was performed using protein quantitative trait loci from deCODE (n = 35 559 Icelanders), UK Biobank (n = 54 219), ARIC (n = 7213 European Americans; n = 1871 African Americans), AGES-Reykjavik (n = 5368 Icelanders), and Fenland (n = 10 708 European ancestry). These datasets were integrated with a sepsis genome-wide association study (GWAS) (FinnGen R12; n = 17 133 cases and 439 048 controls) to identify causal protein candidates. Subsequently, summary-data-based MR was performed. The analysis was based on data from eQTLGen (n = 31 684), GTEx v8 (n = 838 donors), and sepsis GWAS (FinnGen R12; UK Biobank, n = 11 643 cases/474 841 controls). Its goal was to prioritize genes concordant with protein signals. Expression of the candidate was then validated in vivo and in vitro. In Study 2, a metabolite GWAS (n = 8299) was used to investigate metabolic mediation and perform pathway enrichment analysis. Finally, potential small-molecule therapeutics predicted to modulate the prioritized target were identified. Integrated multi-omics analyses identified tubulin-folding cofactor B (TBCB) as a promising therapeutic target for sepsis. Quantitative analyses in both in vivo and in vitro models consistently demonstrated significant upregulation of TBCB during sepsis-induced liver injury. Functional knockdown of TBCB significantly attenuated inflammatory signaling. Mechanistically, TBCB appears to contribute to sepsis progression by modulating intracellular lipid metabolism and metabolic homeostasis. Furthermore, molecular docking and dynamics simulations predicted several small molecules for TBCB, suggesting potential therapeutic value. This study identifies TBCB as a central regulator of sepsis pathogenesis, especially in sepsis-induced liver injury. These findings provide novel mechanistic insights and a promising therapeutic target for intervention.

Neoadjuvant chemotherapy assessment is imperative for prognostication and clinical management of locally advanced gastric cancer. We propose an incremental supervised contrastive learning model (iSCLM), an interpretable artificial intelligence framework integrating pretreatment CT scans and H&E-stained biopsy images, for improved decision-making regarding neoadjuvant chemotherapy. We have constructed and tested iSCLM using retrospective data from 2,387 patients across 10 medical centers and evaluated its discriminative ability in a prospective cohort (132 patients; ChiCTR2300068917). iSCLM achieves areas under receiver operating characteristic curves of 0.846-0.876 across different test cohorts. Computed tomography (CT) and pathological attention heatmaps from Shapley additive explanations and global sort pooling illustrate additional benefits for capturing morphological features through supervised contrastive learning. Specifically, pathological top-ranked tiles exhibit decreased distances to tumor-invasive borders and increased inflammatory cell infiltration in responders compared with non-responders. Moreover, CD11c expression is elevated in responders. The developed interpretable model at the molecular pathology level accurately predicts chemotherapy efficacy.

Accurate prediction of early postoperative recurrence in locally advanced gastric cancer (LAGC) remains challenging due to tumor heterogeneity and limitations of traditional clinicopathological factors. This study aims to develop and validate an interpretable multimodal model for precise recurrence prediction. 1580 LAGC patients are enrolled from six Chinese medical centers and a multimodal fusion Risk Stratification Assessment (RSA) model integrating clinical, radiomic, and pathomic data is developed. Model performance is evaluated using internal, external, prospective, and public dataset validations. Transcriptome sequencing is conducted to elucidate biological mechanisms underlying recurrence. The RSA model significantly outperforms clinical-only, radiomic-only, and pathomic-only models in predicting early recurrence, achieving area under the curve (AUC) values of 0.903 in the training cohort, 0.902 in internal validation, and ranging from 0.884 to 0.889 in external validations. Stratification by the RSA model consistently identifies high-risk patients with significantly poorer five-year survival across all cohorts (all P<0.001). Transcriptomic analysis reveals that high-risk patients exhibit significant immune cell infiltration, increased expression of immune checkpoint molecules, and activation of immune-related pathways, including interferon signaling and the IL-6/JAK/STAT3 pathway. The integrated multimodal RSA model effectively predicts recurrence risk and prognosis in LAGC, enabling precise patient stratification and individualized postoperative management.

High tumor recurrence after surgery remains a significant challenge in managing hepatocellular carcinoma (HCC). We aimed to construct a multimodal model to forecast the early recurrence of HCC after surgical resection and explore the associated biological mechanisms. Overall, 519 patients with HCC were included from three medical centers. 433 patients from Nanfang Hospital were used as the training cohort, and 86 patients from the other two hospitals comprised validation cohort. Radiomics and deep learning (DL) models were developed using contrast-enhanced computed tomography images. Radiomics feature visualization and gradient-weighted class activation mapping were applied to improve interpretability. A multimodal model (MM-RDLM) was constructed by integrating radiomics and DL models. Associations between MM-RDLM and recurrence-free survival (RFS) and overall survival were analyzed. Gene set enrichment analysis (GSEA) and multiplex immunohistochemistry (mIHC) were used to investigate the biological mechanisms. Models based on hepatic arterial phase images exhibited the best predictive performance, with radiomics and DL models achieving areas under the curve (AUCs) of 0.770 (95 % confidence interval [CI]: 0.725-0.815) and 0.846 (95 % CI: 0.807-0.886), respectively, in the training cohort. MM-RDLM achieved an AUC of 0.955 (95 % CI: 0.937-0.972) in the training cohort and 0.930 (95 % CI: 0.876-0.984) in the validation cohort. MM-RDLM (high vs. low) was notably linked to RFS in the training (hazard ratio [HR] = 7.80 [5.74 - 10.61], P < 0.001) and validation (HR = 10.46 [4.96 - 22.68], P < 0.001) cohorts. GSEA revealed enrichment of the natural killer cell-mediated cytotoxicity pathway in the MM-RDLM low cohort. mIHC showed significantly higher percentages of CD3-, CD56-, and CD8-positive cells in the MM-RDLM low group. The MM-RDLM model demonstrated strong predictive performance for early postoperative recurrence of HCC. These findings contribute to identifying patients at high risk for early recurrence and provide insights into the potential underlying biological mechanisms.

Given the extensive heterogeneity and variability, understanding cellular functions and regulatory mechanisms through the analysis of multi‐omics datasets becomes extremely challenging. Here, a comprehensive modeling framework of multi‐omics machine learning and metabolic network models are proposed that covers various cellular biological processes across multiple scales. This model on an extensive normalized compendium of Bacillus subtilis is validated, which encompasses gene expression data from environmental perturbations, transcriptional regulation, signal transduction, protein translation, and growth measurements. Comparison with high‐throughput experimental data shows that EM_iBsu1209‐ME, constructed on this basis, can accurately predict the expression of 605 genes and the synthesis of 23 metabolites under different conditions. This study paves the way for the construction of comprehensive biological databases and high‐performance multi‐omics metabolic models to achieve accurate predictive analysis in exploring complex mechanisms of cell genotypes and phenotypes.

While genome-wide association studies (GWAS) have implicated immune and metabolic pathways in Alzheimer’s disease (AD), their specific cellular impacts remain unclear. To address this, we employed bidirectional two-sample Mendelian randomization to identify single nucleotide polymorphism (SNP)-mediated, AD-associated immunometabolic signatures, which revealed both positively and negatively correlated immune cell types and metabolic pathways. Integrated single-cell omics analysis further delineated distinct astrocyte subpopulations in patient brains: one enriched for Glutamate-glutamine uptake and metabolism was positively associated with AD, while another characterized by Amino acid metabolism and transport was negatively associated. In peripheral blood, mononuclear cells (PBMCs) primarily displayed AD-negative metabolic signatures accompanied by downregulated immune responses. Leveraging these findings, we developed and optimized a blood transcriptome-based AD prediction model on a gene set derived from blood immune cells that is negatively associated with AD, using multiple machine learning approaches. This model is applicable to both European and Asian populations, enables pre-symptomatic detection for familial AD, effectively discriminates AD from other neurodegenerative disorders, and is readily accessible for clinical implementation. Our study provides novel evidence underscoring the critical role of immunometabolism in AD and delivers a practical predictive tool suitable for large-scale, routine population screening.

Mineralocorticoid receptor (MR) antagonists (MRA) are beneficial in cardiorenal outcomes in randomized controlled trials but the mechanisms are unclear. MAGMA was an NHLBI sponsored randomized, double-blind, placebo controlled, 12-month trial comparing Spironolactone (n = 37) vs. placebo (n = 42) in Type 2 diabetics with CKD stages 3-4 on maximal renin-angiotensin system (RAS) blockade and a prior atherosclerotic event and/or left ventricular (LV) hypertrophy. The primary outcome of percent (%) change in total aortic wall volume (TWV) at 12 months, measured by magnetic resonance imaging (MRI), was significantly reduced by Spironolactone. The purpose of this analysis was to understand mechanistic pathways of MRAs using a multi-omics approach that could help tease out molecular mechanisms of benefit. Revealing for the first time which of these integrated pathways are predictive of the primary outcome will help design treatments for targeted interventions. Plasma and peripheral blood mononuclear cells from patients randomized to Spironolactone or placebo at baseline and 3-months were measured for aptamer-based proteomic biomarkers (7,596 proteins) and 10-X platform-based single-cell RNA-sequencing (scRNAseq), respectively. Pre-selected candidate predictors were used as inputs of a predictive model of changes of TWV. We fit a Mixed-Multivariate Random Forest (RF) model, a variation of the RF supervised tree-based machine learning method to take the experimental design into account in the regression formulation. The two sources of "omics predictors" were integrated into a supervised multi-omics model to jointly explain the outcome using an extension of Sparse Generalized Canonical Correlation Analysis (SGCCA). Integrated multiome functional analyses with graphical visualizations were carried out by statistical enrichment analysis using Over Representation Analysis (ORA) and Gene Set Enrichment Analysis (GSEA) against databases of gene ontologies, biological pathways, putative regulatory motifs, proteins, or disease annotations. The plasma proteome in response to Spironolactone revealed downregulation of MR targets including fibrosis, immune activation/inflammation, leukocyte activation, proliferation and pathways involved in cytokine stimulation. scRNAseq pathways revealed negative regulation of cytokines production such as IL-2 and redistribution of multiple cell types. Predictors of plaque progression involved cytokine-receptor, complement-coagulation, cell adhesion and axonal guidance targets. Multiome integrated functional analyses results of predictive pathways will be presented. The changes in plasma proteomic profile with Spironolactone were consistent with the phenotype of reduced atherosclerosis and downregulation of multiple inflammatory, immune response and profibrotic pathways.

COVID-19 has had major global impacts, highlighting the importance of robust predictive surveillance and diagnostic systems to ensure effective public health responses. Traditional surveillance methods based on passive case counting and diagnostic testing of individual patients often suffer from delays and resource constraints, preventing timely responses. This study proposed an integrative framework integrating machine learning (ML)-derived predictive surveillance with wastewater-based diagnosis, aiming to predict temporal trends in Korea and identify disease-causing agents. The ML model utilized crowdsourced COVID-19-related keywords, climatic, and environmental data, optimized via model selection and feature selection. The integrated data-driven model predicted COVID-19 cases over three years more accurately than those using single source data (i.e., baseline model). The explainable AI technique (i.e., helping to inform how the model made those predictive decisions) identified six keywords (reducing phlegm, throat pain, long COVID-19, sore throat, COVID-19 self-kit, and COVID-19 kit) as robust predictors of disease trends. In a proof-of-concept experiment, wastewater-based genotyping of disease-causing agents and affected human communities in sewershed areas was conducted. Metatranscriptomics of municipal wastewater was conducted to identify COVID-19 viral variants, evolutionarily related to those clinically isolated strains, distinguishable from conventional diagnostic testing of individual patients. Wastewater-derived metagenomics was also performed to assess genomic variation in the affected human populations. The integrative framework proposed in this study offers a rapid, cost-effective approach for the surveillance and diagnosis of COVID-19 and other infectious diseases, thus strengthening or complementing existing health systems.

The globalization of food systems has heightened the risk of foodborne pathogens such as Salmonella, Listeria monocytogenes, and Campylobacter, exacerbated by rising antimicrobial resistance (AMR). Traditional pathogen identification and AMR risk surveillance methods are often labor-intensive and low-throughput, while single-omics approaches fail to capture microbial complexity. Moreover, reliance on individual machine learning (ML) models limits predictive robustness, posing challenges to food safety and public health. This systematic review evaluates ML-based predictive modeling integrated with omics techniques (genomics, metagenomics, and transcriptomics) for foodborne pathogen and AMR risk surveillance. Following PRISMA guidelines, 1245 articles from PubMed, Scopus, and other databases (2015-2025) were screened, selecting 13 relevant studies. These studies applied ML algorithms, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Support Vector Machines (SVM), to enhance predictive accuracy. The selected studies demonstrated predictive accuracies up to 99 % and AUROC scores above 0.90. Key discoveries include genetic markers for Salmonella virulence, Listeria attribution to fruits and dairy, and 145 mobile antimicrobial resistance genes (ARGs) in poultry. Despite these advancements, limitations such as small sample sizes, inconsistent metadata, overfitting, and computational scalability hinder real-world implementation. This review underscores the potential of ML-driven omics frameworks to revolutionize foodborne pathogen and AMR risk monitoring, paving the way for smarter, more resilient food safety systems. However, methodological inconsistencies necessitate standardized protocols, larger datasets, and explainable AI (XAI) to improve reliability and applicability in global food safety monitoring.

In non-small cell lung cancer (NSCLC), identifying circulating predictive biomarkers is critical to optimize patient selection and improve therapeutic outcomes. Plasma-based biomarkers offer a minimally invasive and serially accessible approach for monitoring treatment response; however, the low abundance of disease-relevant analytes within a background of highly abundant plasma proteins presents significant analytical challenges. Highly sensitive and reproducible multi-omics assays are therefore required to detect subtle but biologically relevant changes associated with therapy. To address this, we implemented a plasma multi-omics workflow integrating unbiased mass spectrometry (P2 DIA-MS), the NULISA inflammation panel, and targeted metabolomics using the Biocrates MxP Quant 1000 kit. This workflow was applied to longitudinal plasma samples from patients with NSCLC enrolled in the multicenter phase II clinical trial SAKK 17/18. Parallel profiling quantified approximately 6,000 plasma proteins, 250 inflammation-related markers, and more than 1,200 metabolites and lipids, enabling cross-platform integration of proteomic, cytokine, and metabolic signatures. For predictive biomarker discovery, we applied a machine learning framework to integrate proteomic, inflammatory, and metabolic data. Incorporation of multi-omics features improved the stratification of responders versus non-responders compared with single-omic models, underscoring the complementary nature of each dataset. To further investigate biological relationships among omics layers, we evaluated several data integration methods, including Multi-Omics Factor Analysis (MOFA), to identify shared sources of variation and linked biological pathways. This analysis revealed coordinated processes connecting plasma proteomic changes with inflammatory and metabolic networks. For example, immune activation signatures arising from the integration of unbiased proteomics with inflammation markers, and metabolic stress adaptation reflected in proteomic-metabolomic associations. In conclusion, this plasma multi-omics approach demonstrates the potential of integrated proteomic, inflammatory, and metabolic profiling to identify predictive circulating biomarkers and to elucidate biological mechanisms of treatment response in NSCLC. The workflow provides a scalable and clinically applicable strategy for biomarker discovery and response monitoring in oncology. Laura Heeb, Polina Shichkova, Sandra Schär, Luca Räss, Arthur Viodé, Martin Mehnert, Tobias Treiber, Esther Wortmann, Gordian Adam, Alice Limonciel, Markus Joerger, Jana Musilova, Stefanie Hayoz, Anurag Gupta, Yuehan Feng, Alessandra Curioni-Fontecedro. Integrated plasma multi-omics profiling identifies circulating predictive biomarkers and biological pathways associated with treatment response in NSCLC [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 106.

The aroma components and sensory characteristics of fresh cheese fermented with three novel probiotics (Lacticaseibacillus rhamnosus B6, Limosilactobacillus fermentum B44, and Lacticaseibacillus rhamnosus KF7) were investigated using an omics approach based on HS-SPME-GC-TOFMS. Multi-dimensional and single-dimensional predictive mathematical models were developed to analyze the relationship between sensory scores and characteristic compounds. The results demonstrated that the three probiotics significantly influenced the volatile metabolite composition and sensory properties of fresh cheese. Totally 16 key aroma compounds (OAV ≥ 1) were identified. Based on OAV and (O) PLS-DA, 4, 7, and 4 significantly upregulated key aroma compounds were detected in the B6, B44, and KF7 groups, respectively. The metabolic pathways of these key compounds were reconstructed, revealing their association with fatty acid β-oxidation, aromatic amino acid metabolism, glycolysis, and esterification. L. fermentum B44, L. rhamnosus KF7, and L. rhamnosus B6 promoted the production of favorable key volatiles, altering flavor profiles. The samples of B6, B44, and KF7 groups exhibited distinct flavor characteristics described as "milk odor", "cheese odor", and "lactic odor", respectively, with the B44 sample achieving the highest overall acceptance. A natural logarithm-based partial least squares regression model was optimized, and the suitability of a nonparametric Bayesian Gaussian regression model for fitting sensory scores was confirmed. Among the identified compounds, 1-pentanol emerged as the most likely sensory score marker. This study elucidated the mechanisms underlying the formation of characteristic flavors and metabolites in probiotic fresh cheese and provided a reliable correlation model to support flavor regulation and quality control.

Introduction:Artificial intelligence (AI) and machine learning (ML) are increasingly used in oncology to enable personalized treatment. In multiple myeloma (MM), novel drugs and combination regimens have improved patient outcomes, but treatment response and duration vary widely, due to clinical and molecular heterogeneity. To address this variability, we developed a multimodal machine learning framework that integrates gene expression and clinical data to predict individualized treatment duration, enabling patient stratification for therapy selection and helping identify regimens with the longest expected benefit or optimal choices when durations are similar, supporting precision oncology in MM. Methods:A total of 652 MM patients with paired clinical and genomic profiles from the MM Research Foundation (MMRF) CoMMpass IA22 dataset were included after quality control, which removed patients without treatment duration data, retained only MM-relevant clinical variables, and kept only the first treatment course and its duration for patients with multiple courses. We assembled a comprehensive feature set comprising 150 clinical variables (120 numerical and 30 encoded categorical) and 196,661 gene expression features derived from bone marrow aspirates. Treatment information included 17 therapeutic agents and 45 treatment regimens, covering doublet, triplet, and quadruplet combinations. All numerical features were normalized to zero mean and unit variance. To manage high dimensionality and multimodal data heterogeneity, we developed a structured machine learning analytical pipeline using gradient-boosted decision tree (GBDT) feature selection to identify clinical and molecular variables with the highest predictive value. We evaluated three machine learning approaches for estimating treatment duration: (1) Multimodal Neural Network (MMNN), a deep learning model that integrates clinical and gene expression data through separate neural pathways for prediction; (2) Extreme Gradient Boosting (XGBoost), a gradient boosting ensemble that builds sequential decision trees while maintaining interpretability through feature importance analysis; and (3) Feature Tokenizer Transformer (FTTransformer), a transformer-based architecture that uses self-attention mechanisms to capture complex patterns in biomedical data. These models predict the days a patient remains on their initial therapy until discontinuation or progression, using 5-fold cross-validation to ensure robustness. Model performance was assessed using accuracy, recall, F1-score, and Area Under the Curve (AUC). Results:The GBDT algorithm identified 146 top features, capturing 95% of cumulative importance. These included 5 clinical variables—percentage of plasma cells in bone marrow, serum Ig lambda light chains, glucose level, R-ISS stage, and sex. The remaining 141 features were gene expression markers, including SLAMF1, a surface antigen expressed on MM cells and a therapeutic target; CDK4 and MYB, key regulators of cell cycle and transcription implicated in MM pathogenesis; NUAK1 (ARK5), implicated in plasma cell survival; and FOXP3, whose overexpression in the bone marrow environment indicates MM suggests an accumulation of CD4 regulatory T cells that contribute to the immunosuppressive niche. We used these features to train and evaluate MMNN, XGBoost, and FTTransformer models. XGBoost achieved the best performance, with accuracy 91.2% (95% CI: 0.904–0.921), AUC 0.874 (95% CI: 0.863–0.884), F1-score 0.910 (95% CI: 0.902–0.918), and recall 0.912 (95% CI: 0.904–0.921). MMNN and FTTransformer yielded accuracies of 85.6% and 81.2%, and AUCs of 0.810 and 0.738, respectively. These results demonstrate the effectiveness of XGBoost, particularly when paired with high-dimensional feature selection. Conclusion:This study presents a multimodal machine learning-based framework that accurately predicts individualized treatment duration in MM patients by integrating clinical and gene expression data. The XGBoost model outperformed deep learning alternatives and demonstrated high predictive performance, suggesting immediate applicability in clinical decision-making. The proposed framework demonstrates the feasibility of integrating AI into MM treatment planning, offering physicians a data-driven tool to estimate therapy duration. Future work will focus on external validation and deployment in clinical workflows to support real-time, patient-specific decision-making.

Glioblastoma (GBM) is an aggressive tumor with a ~15-month median survival and high interpatient variability. Deep learning (DL) models using structural MRI have shown promise for individualized survival prediction, but most are developed and tested at a single institution, limiting generalizability. We externally validated a multimodal DL model trained on heterogeneous data from three sites using an independent GBM cohort from Moffitt. This study offers a rare and rigorous assessment of model performance on truly unseen data, highlighting its potential for real-world clinical deployment. We identified 149 patients with confirmed glioblastoma at Moffitt who had preoperative MRI scans acquired within four weeks of surgery. Each patient had four axial MRI sequences available: T1, T2, T1-postcontrast, and T2-FLAIR. The pretrained multimodal DL model was applied without additional fine-tuning. The model uses a ViT architecture and incorporates age as a clinical variable via a late fusion approach. Patients were stratified into high- and low-risk groups based on predicted survival scores. The concordance index (C-Index) and time-dependent area under the receiver operator curve (AUC) at one year were calculated (±95% CI). The model achieved a C-index of 0.680 ± 0.051 and an AUC of 0.773 ± 0.079. Kaplan–Meier analysis showed a significant difference in overall survival between model-predicted high- and low-risk groups (hazard ratio = 1.69, 95% CI: 1.22–2.36; p = 0.001). Age was significantly higher in the high-risk group (p < 0.05), consistent with its known prognostic relevance. Incorporating age via late fusion modestly improved performance, increasing the C-index to 0.686 ± 0.052. This externally validated multimodal deep learning model accurately stratified glioblastoma patients by survival risk using routine preoperative MRIs. The findings highlight the feasibility of deploying such models across institutions and support their potential to enable personalized prognosis and risk-adapted treatment planning in clinical neuro-oncology.