Sample editionVol. I · No. 2 · Wednesday, 4 November 2026No. 2 · Wed 4 Nov
Editor's noteToday leads with a few-shot ophthalmic AI benchmark, a multi-omics toxicity classifier for Evodia rutaecarpa, and an OIP5-AS1 gastric cancer study, with a BMP4 AI review also featured.
Strong day: 10 strong matches
This edition15 papers · tap a bar to filter the edition
By project
By study type
By match strength
No. 1 today·Match strength 79 of 100, Strong·Front Pharmacol·Published 19 Oct · new in PubMed 3 Nov
FIGURE 8. Screening of key metabolites and construction of predictive models (A) AUC-ROC comparison of models deployed across various omics; (B) pie chart of relative contribution based on AUC; (C) summary plot of salivary microbiome feature contributions; (D) summary plot of serum metabolome feature contributions (Beeswarm Plot); (E) summary plot of urine metabolome feature contributions; (F) individualized prediction waterfall plot for typical high-risk samples of salivary microbiota; (G) individualized prediction waterfall plot for typical high-risk samples of serum; (H) individualized prediction waterfall plot for typical high-risk samples of urine. From Tan X et al., Frontiers in Pharmacology (2026). Via PubMed Central (NLM). License: CC BY.
Multi-omics profiling of 86 gastric disease patients taking Evodia rutaecarpa-containing formulas identified ten discriminative features that distinguish those with adverse reactions from controls, combining serum and urine metabolomics with saliva 16S rRNA sequencing. Seven machine learning algorithms with LOOCV were used to build the models and derive an AUC-weighted joint prediction formula, with a score cutoff of 0.597. The authors report tentative trends rather than validated markers. The work matters for the reader's interest in multi-omics with machine learning, as an example of applying integrated omics classifiers to individual drug-safety risk.
Why it's here: "multi-omics + machine learning" in the title · 1 day ago
Key findings from the full text
In 86 patients with gastric disease taking EVR-containing formulas, multi-omics profiling and machine learning identified serum, urinary and salivary features associated with adverse reactions, combined into an AUC-weighted risk formula; the authors describe these as preliminary associations.
DesignObservational cohort of 86 patients with chronic gastritis-related diseases receiving EVR-containing formulas, split into adverse reaction (n = 36) and control (n = 50) groups, with serum, urine and saliva omics analyzed by seven machine learning classifiers under LOOCV.
AUC = 0.930 (95% CI 0.843–0.992) urinary metabolomics LightGBM model AUC
AUC = 0.842 (95% CI 0.724–0.935) salivary microbiome random forest model AUC
0.597 combined risk threshold of the AUC-weighted formula
P = 0.013 Simpson index difference in saliva between groups
specificity 1.000, sensitivity 0.886 urine LightGBM model performance
Methods and tools
Untargeted LC-MS metabolomics on Thermo Q Exactive HF-X with Compound Discoverer 3.3 processing
16S rRNA V3-V4 sequencing with DADA2 and QIIME 2 amplicon sequence variant processing
Seven classifiers evaluated with leave-one-out cross-validation after Welch's t-test and L1 regularized feature selection
PICRUSt2 functional prediction
OPLS-DA with permutation testing
Data and code
Raw 16S data not public; processed de-identified data available on request
Limitations
Small, partially overlapping subsets make models vulnerable to overfitting; estimates may be optimistic even with LOOCV.
Patients received compound formulas, so co-administered herbs and concomitant medications are confounders.
Observational design can only identify statistical associations; external validation is needed.
The model cannot be directly applied to routine clinical practice at present.
Why it mattersIt offers an exploratory multi-omics machine learning framework for individual risk prediction of adverse reactions to a low-toxicity herbal medicine, though it is not yet clinically validated.
FIGURE 2.Study design and participant flow diagram After screening, 86 participants with verified use of EVR–containing formula were enrolled and divided into adverse reaction group (n = 36) and control group (n = 50), followed by multi-omics analysis and machine learning to construct a metabolic biomarker panel.From Tan X et al., Front Pharmacol. License: CC BY. Via PubMed Central (NLM).
Checked against the full text: every quote, number and accession must appear in it.
DatesPublished in Frontiers in Pharmacology on 19 October 2026. Entered PubMed on 3 November 2026; Revue ranks by this date, the first day it could show the paper.
Why it may matter to youIt is a concrete clinical example of combining multi-omics data with machine learning classifiers to estimate individual risk, relevant to your multi-omics and machine learning interests.
OIP5-AS1 was prioritized as a candidate oncogenic regulator and putative ceRNA hub in gastric cancer, with significant grade-dependent overexpression in tumor tissue from a 30-patient clinical cohort. The study combined bioinformatics, single-cell RNA sequencing, machine learning and experimental assays, including siRNA silencing in AGS cells that induced apoptosis and reduced migration. The authors note that clinical utility and direct molecular binding still require validation.
Why it's here: "single-cell + machine learning" in the title · 1 day ago
DatesPublished in Computational Biology and Chemistry on 27 October 2026. Entered PubMed on 3 November 2026; Revue ranks by this date, the first day it could show the paper.
Why it may matter to youIt applies single-cell transcriptomics and machine learning to multi-omics data to nominate a candidate biomarker and therapeutic target, relevant to your interest in machine learning on biomedical data.
FIGURE 2. Context-dependent duality of BMP4 signaling across disease states. From Sun W et al., Frontiers in Molecular Biosciences (2026). Via Frontiers. License: CC BY.
This narrative review finds AI has accelerated BMP4 research across drug discovery, structure prediction and clinical prognostication, though predicted models and prognostic models still await experimental and prospective validation. It surveys English-language publications from 2020 to 2026 across PubMed, Scopus, Web of Science, bioRxiv, arXiv and Semantic Scholar, categorizing included studies by BMP4-specific, BMP-pathway-targeted, or transferable AI methods. Key gaps remain, including no BMP4-specific de novo protein design, few large training datasets, and limited multimodal AI in BMP4 clinical research.
Why it's here: "single-cell + machine learning" in the abstract · new today
Key findings from the full text
The review finds that AI has accelerated BMP4-related research along several axes, including AlphaFold3-predicted BMP4-receptor complex models that remain unvalidated, ML-guided virtual screening that yielded BMP pathway (receptor) modulators for FOP but no BMP4-specific small molecules, and single-cell and prognostic ML analyses of BMP4 in glioblastoma, pancreatic and gastric cancer that await prospective validation.
DesignNarrative review with systematic search elements of PubMed, Scopus, Web of Science, bioRxiv, and arXiv publications from 2020 to 2026, with studies labeled as BMP4-specific, BMP receptor/pathway-targeted, or transferable AI methodology.
Key numbers
AUC: 0.889; accuracy: 84.21% Random Forest PDAC diagnostic performance (Xiao et al., 2025)
C-index values of 0.738 (training) and 0.724 (validation) Nomogram with BMP4 for pancreatic adenocarcinoma prognosis
97.25% Accuracy of neuro-symbolic DPP-4 inhibitor classification
FDR: 2.74 × 10 −2 Enrichment of BMP signaling in IPF disease-associated co-expression module
approximately 140,000 compounds Size of computational screen that led to RK783 for FOP
approximately 70% Fraction of BMP/SMAD parameter combinations on the Pareto front
Methods and tools
AlphaFold3 structure prediction of BMP4 prodomain phosphorylation and complexes
RNA-velocity trajectory analysis of BMP4-treated glioblastoma cultures
LASSO-Cox regression for hypoxia-immune prognostic panel in gastric cancer
STAG mathematical model fitted to barcoded lineage-tracing data
Naive Bayes classifiers with ECFP6 fingerprints for ULK1 screening
Limitations
AlphaFold-type models are predictive, not experimental, and high global confidence does not guarantee accuracy for flexible loops or post-translational states.
The gastric cancer findings derive from retrospective cohorts and have not been prospectively validated.
The pancreatic nomogram remains externally unvalidated.
Deep learning gene perturbation prediction does not yet outperform simple linear baselines on standard benchmarks.
Why it mattersIt maps which AI tools (AlphaFold3, LASSO-Cox, single-cell ML, Boolean and trajectory models) have and have not been applied to BMP4 biology, showing where multi-omics and protein structure methods could be applied next.
FIGURE 4.RNA-velocity trajectory analysis of BMP4-treated patient-derived glioblastoma (GBM) cultures reveals context-dependent cell state transitions, including depletion of pro-inflammatory and neuronal subpopulations and induction of a quiescent, BMP/Hippo pathway-active cluster in BMP4-sensitive cultures.From Sun W et al., Front Mol Biosci. License: CC BY. Via Frontiers.
Checked against the full text: every quote, number and accession must appear in it.
DatesPublished in Frontiers in Molecular Biosciences on 20 October 2026. Entered PubMed on 4 November 2026; Revue ranks by this date, the first day it could show the paper.
Why it may matter to youIt maps how AI tools such as AlphaFold3 and single-cell machine learning are being applied to BMP4 biology, which bears directly on your interests in protein structure prediction, single-cell machine learning and multi-omics AI.
Seventy-two studies were included in this scoping review of LLM-generated case-based materials in health professions education, with 67/72 (93%) published in 2024-2026. Most focused on medical education (59/72) and generated case-based MCQs (28/72) or clinical cases/vignettes (22/72), with GPT-4 family models most common (40/72). Materials were often judged clinically plausible and well aligned with objectives after refinement, but recurring limitations included factual inaccuracies, superficial realism, psychometric weakness and bias, so expert oversight is still needed.
Why it's here: "large language models + clinical" in the abstract · 1 day ago
DatesPublished in JMIR medical education on 3 November 2026. Entered PubMed on 3 November 2026; Revue ranks by this date, the first day it could show the paper.
Why it may matter to youFor a reader working on large language models in medicine, this maps how LLMs are being used to create clinical cases and assessment materials for medical training and where their reliability falls short.
Variational autoencoders trained on ancestral and near-ancestral sequence datasets generated ethylene-forming enzyme variants with improved predicted and experimentally measured thermostability compared with variants from modern-sequence training data. All tested ancestral and ML-generated enzymes produced detectable ethylene, though ML-generated variants had reduced activity relative to wild-type. The authors also fine-tuned ESM2 on ancestral-enriched sequence sets for endolysin and lysozyme C stability classification, where ancestral representations were competitive in selected settings.
Why it's here: "protein structure prediction" in the abstract · 1 day ago
DatesPublished in Protein Engineering, Design & Selection on 3 November 2026. Entered PubMed on 3 November 2026; Revue ranks by this date, the first day it could show the paper.
Why it may matter to youIt offers a data-centric route to improving protein sequence generation and representation learning for stability, a direction relevant to protein structure and ML-based protein engineering work.
Figure 5. Study-standardized index performance. From Pan C et al., Frontiers in Cardiovascular Medicine (2026). Via PubMed Central (NLM). License: CC BY.
An eight-gene blood transcriptome index (BAX, EDA2R, MCEMP1, SOCS3, ASGR2, RNASE1, ADM, KPNA5) discriminated acute myocardial infarction with an AUC of 0.984 in the discovery cohort and 0.863, 0.655 and 0.876 in three validation cohorts. The authors pooled four GEO blood transcriptome datasets (348 participants), applied nested machine learning and WGCNA, and checked selected inflammatory genes by RT-qPCR in THP-1-derived macrophages and EA.hy926 cells. The abstract also reports exploratory outcome analyses (AUCs 0.819 and 0.569) and RT-qPCR fold changes for CXCL10, IL6, NLRP3 and CASP1.
Why it's here: "multi-omics + machine learning" in the abstract · 1 day ago
Key findings from the full text
An eight-gene blood transcriptomic index (BAX, EDA2R, MCEMP1, SOCS3, ASGR2, RNASE1, ADM, KPNA5) derived with nested machine learning discriminated acute myocardial infarction from controls, with repeated out-of-fold AUC 0.984 in discovery and independent validation AUCs of 0.863, 0.655, and 0.876.
DesignRetrospective multi-cohort analysis of public GEO peripheral blood transcriptomes: discovery in GSE59867 (111 STEMI, 46 stable CAD) and validation in GSE48060, GSE60993 and GSE66360 (348 participants total across four datasets).
Key numbers
AUC 0.984 repeated out-of-fold AUC in discovery cohort GSE59867
0.863 (95% CI 0.739–0.960) AUC in validation cohort GSE48060
0.655 (0.412–0.874) AUC in validation cohort GSE60993
0.876 (0.807–0.934) AUC in validation cohort GSE66360
r = 0.78, p < 0.001 phenotypic correlation of WGCNA turquoise module
4.15 ± 0.44-fold IL6 expression in LPS/ATP-stimulated THP-1 macrophages by RT-qPCR
Methods and tools
Welch t-tests with Benjamini–Hochberg FDR correction for study-wise differential expression
Weighted gene co-expression network analysis (WGCNA) for module identification
Nested repeated stratified cross-validation with L2-regularized logistic regression, linear SVM, random forest and gradient boosting
Marker-based immune program scoring with Mann–Whitney U tests and Cliff's delta
RT-qPCR using 2^-ΔΔCt normalized to GAPDH in ATCC-authenticated THP-1 and EA.hy926 cells
Data and code
Data are stated to be in online repositories; accession numbers given in the article are the GEO series used
Discovery and validation analyses used GEO series GSE59867, GSE48060, GSE60993, GSE66360, and single-cell GSE121893
Limitations
Peripheral blood rather than infarct tissue was used, requiring caution in extrapolation
The two binary-outcome analyses were severely underpowered and lacked patient-level censoring data
The single-cell dataset contained no AMI specimens
WGCNA module assignments used the full-discovery dataset, a residual source of potential enrichment
Figure 4.Leakage-controlled machine learning development. (A) Gene-selection frequencies across 25 outer folds; green bars indicate final model genes. (B) Inner cross-validation (CV) AUC across the logistic regularization strengths. (C) Final standardized logistic coefficients. (D) Random forest permutation importance for the same eight genes. (E) Repeated outer-fold AUC comparison among the four classifiers. (F) ROC curves from the averaged outer-fold predictions.From Pan C et al., Front Cardiovasc Med. License: CC BY. Via PubMed Central (NLM).
Checked against the full text: every quote, number and accession must appear in it.
DatesPublished in Frontiers in Cardiovascular Medicine on 19 October 2026. Entered PubMed on 3 November 2026; Revue ranks by this date, the first day it could show the paper.
DOI10.3389/fcvm.2026.1904246
Data and codeGEO GSE59867 · GEO GSE48060 · GEO GSE60993 · GEO GSE66360 · GEO GSE121893
Why it may matter to youMatches your topics: multi-omics + (machine learning / deep learning / artificial intelligence), single-cell + machine learning.
Figure 1. Hepatobiliary cancers digital twin system architecture. From Liu K et al., Journal of Translational Internal Medicine (2026). Via PubMed Central (NLM). License: CC BY.
A review proposes a three-layer digital twin architecture for hepatobiliary cancers that combines AI-driven multi-omics with patient-derived organoid models to support individualized, adaptive treatment. It outlines a Data Layer digitizing static atlases, a Physical Layer of engineered organoid avatars using tunable biomaterials, and a Computational Layer integrating both with patient data. It also discusses clinical applications such as therapy selection and drug sequencing, and barriers including limited paired datasets and variable organoid feasibility.
Why it's here: "multi-omics + artificial intelligence" in the title · 1 day ago
Key findings from the full text
The review proposes a three-layer digital twin framework (Data, Physical, Computational) for hepatobiliary cancers that links static multi-omics and pathology atlases to patient-derived organoid functional readouts and longitudinal clinical data, as an emerging translational strategy rather than a mature clinical tool.
DesignNarrative review of the literature on digital twins for hepatobiliary cancers (HCC, CCA, GBC) integrating digital pathology, spatial omics, patient-derived organoids and clinical data; no sample size is reported.
Key numbers
92.3% (12/13 evaluable) concordance of PDO ex vivo screening with clinical response in one BTC vignette
Methods and tools
Deep-learning pipeline Organoid for semantic segmentation and tracking of organoids
Trajectory inference with CellRank for mapping cell states
Barcode-assisted CRISPR screens in three-dimensional organoids
High-content imaging with AI-based phenotyping of organoids
Cell-graph construction from whole-slide images for spatial relationships between tumor, immune and stromal cells
Data and code
No additional data; data availability statement reports none beyond the review.
Limitations
Limited paired datasets linking atlas and PDO data, which are fragmented and lack harmonized metadata.
Most PDO systems lack perfusable vasculature and neural inputs and incompletely reproduce biliary flow and immune-stromal dynamics.
Current cohorts of PDO-guided decisions remain modest and require further confirmation.
Temporal mismatch between cellular signaling (seconds), organoid responses (weeks) and clinical progression (months to years) is often ignored by current models.
Figure 2.Engineering patient-matched physical avatars in hepatobiliary cancers. This schematic depicts the Physical Layer of a hepatobiliary cancer digital twin. A parameter-translation pipeline converts quantitative digital pathology features (e.g., collagen signatures, cell-cell interaction networks) into engineering constraints for patient‑matched hydrogel synthesis and multi‑compartment ecosystem construction. For HCC, vascular-rich lesions with low stiffness and immune‑inflamed architectures are modeled using organoid-on-a-chip systems with physiological flow. For BTC, densely fibrotic stroma with high stiffness and immune‑excluded barriers are replicated. HCC, hepatocellular carcinoma; BTC, biliary tract cancer. Created in BioRender (2026). (https://BioRender.com/n7juqfy).From Liu K et al., J Transl Int Med. License: CC BY. Via PubMed Central (NLM).
Checked against the full text: every quote, number and accession must appear in it.
DatesPublished in Journal of Translational Internal Medicine on 23 October 2026. Entered PubMed on 3 November 2026; Revue ranks by this date, the first day it could show the paper.
Figure 4. Patient and clinician ratings of postpartum consultation responses. From Zhao X et al., Frontiers in Public Health (2026). Via Frontiers. License: CC BY.
ChatGPT 5.5 outperformed DeepSeek V3 on most clinically relevant dimensions in real-world postpartum consultations and nursing care plan generation, while DeepSeek V3 scored higher on completeness and safety for standardized questions. The study compared the two models across three stages: repeated querying of 30 guideline-based questions, 120 postpartum consultations, and 120 de-identified retrospective cases, with blinded clinical expert and patient ratings. Its authors conclude that LLMs may support health education and drafting initial nursing care plans, but outputs require clinical review.
Why it's here: "large language models + clinical" in the abstract · new today
DatesPublished in Frontiers in Public Health on 20 October 2026. Entered PubMed on 4 November 2026; Revue ranks by this date, the first day it could show the paper.
Why it may matter to youIt offers a task-by-task evaluation of LLMs in clinical decision support, relevant to applying large language models in medicine and clinical care.