Bioinformatics Advances

Biomedical knowledge graphs can connect information across genes, phenotypes, tissues, pathways, experiments, and clinical resources, but they are difficult to query correctly without detailed knowledge of the graph. Large language models can help write Cypher code for knowledge graphs, yet a query focused on a bioinformatics task that looks reasonable may still use the wrong identifier, relationship direction, source, intermediate node, or output unit if the user is not completely trained on the schema. We developed ddkg.skill, an Agent Skill that can be loaded into a compatible LLM session to support querying of the NIH Common Fund Data Ecosystem Data Distillery Knowledge Graph (DDKG), a biomedical property graph integrating more than 180 ontologies, over 40 genomics datasets, and dozens of additional cross-domain biomedical datasets. Rather than utilizing a simple markdown prompt, ddkg.skill contains a controller, a compositional library of DDKG-specific references and structured tables, primary documentation, validated query patterns, a routing table, and a Python script that checks the internal links among these materials. The skill build for the December 2025 DDKG release contains 38 bundled files and 239 routing relationships and is identified by checksum so users can state the exact build used to compose a query. The evaluation reported here was performed separately on an earlier build with 38 files and 227 routing entries. In that orthogonal nine-test evaluation against a live December 2025 DDKG instance, five of seven tests targeting sources absent from the worked examples produced correct executed results. The evaluation also exposed a failed query, a cross-source comparison that was not biologically well posed, and errors in the skills own reference material that informed later revisions. The skill guides an AI through entity resolution, graph inspection, query construction, and validation for biomedical and bioinformatics queries, while keeping the DDKG itself as the source of returned results. This design provides a portable and versioned method for giving general-purpose AI systems practical knowledge of a complex biomedical knowledge graph to empower complex bioinformatics data integration. Three executed biomedical use cases further demonstrate cross-species phenotype-to-expression querying, genomic overlap between 4DN chromatin-loop anchors and GTEx eQTLs, and integration of congenital heart phenotype associations with ClinVar and GTEx heart evidence.

Small RNA (sRNA) sequencing is widely used to study siRNA, miRNA and piRNA pathways and to profile sRNAs derived from viruses, transposons and host transcripts. However, many sRNA workflows rely on command-line mapping and separate software for downstream analysis and visualisation. We present sRNAMap, an HTML5/JavaScript application for mapping and analysing sRNA sequencing data. sRNAMap accepts either raw FASTA/FASTQ reads or pre-aligned BAM files. For raw reads, sRNAMap uses a k-mer-seeded alignment algorithm that runs directly in the browser using parallel Web Workers where supported. It generates strand-specific per-base coverage, 5'- and 3';-end profiles, read-length distributions and nucleotide-bias summaries, with optional library-size normalisation. It also quantifies characteristic read-overlap signatures, including 5';-to-5'; overlaps associated with ping-pong-like processing. Additional modules quantify phased sRNA production, non-templated 3'- and 5'-additions, and sequence diversity. Analyses can be saved and restored as JSON or exported as self-contained HTML reports containing figures and analysis metadata. Benchmarking against Bowtie 2 under approximately matched alignment settings using a representative virus-infected arthropod small-RNA dataset showed highly concordant mapping profiles. Per-position depth showed an R^2 of approximately 0.998, while 5';- and 3';-end counts showed R^2 values of approximately 0.97. sRNAMap is available at https://github.com/rhparry/sRNAMap under the MIT licence.

Motivation: LLM agents increasingly draft functional interpretations of gene lists, but can cite Gene Ontology (GO) terms that no current enrichment backend returned for that list, and can pair real GO accessions with fabricated labels. Results: We present GroundAnnot, a client for PANTHER, Enrichr, and g:Profiler that returns a closed vocabulary of GO term IDs and their backend labels, and enforces two contracts: Contract A (no ID or label outside the backend payload) and Contract B (no enrichment claim outside the backend's significant set). Across three open-weight models (Qwen2.5-7B local; Qwen3.8-27B and GPT-OSS-120B via Groq) on six curated disease gene lists, mean valid-enriched rates were 15.3%, 20.1%, and 35.1%. All unsupported IDs were real GO terms classified by QuickGO as wrong-biology, obsolete, or wrong-branch (zero fabricated accessions). A distinct failure mode emerged: for the Alzheimer's gene list, the local 7B model paired every one of its ten stable picks with a label that does not match the current GO term for that accession (10/10 mismatches), producing a coherent synaptic-signalling narrative that the IDs do not support. Mean overlap with the backends' FDR top-10 union was 0.33-0.67/10. On 50 MSigDB Hallmark gene sets, the strongest model named the defining pathway in 21/44 (47.7%) directly named cases. When an enriched shortlist was placed in the prompt, all three models complied fully (60/60 for each model; 180/180 overall), showing that a bounded vocabulary removes these output classes when the downstream agent is required to use it.

Agentic artificial intelligence (AAI) is increasingly used by biomedical scientists, where it has substantially lowered the difficulty of integrating computational, statistical and data-science approaches into day-to-day research activities, assisting with project design, analysis and manuscript preparation. Significant barriers remain, however, which are unlikely to be removed solely by releasing more capable AI models. Meaningful use still demands a technical fluency that many biomedical researchers lack: a PI and their lab members must learn how to use the AAI system, establish standard operating procedures (SOPs) for managing, sharing and curating data, and enforce rules that address data privacy and security. Existing commercial AAI systems, which were designed primarily for software engineering purposes, need to be significantly re-purposed for biomedical-related data science. We present Murmurent, shared software that sits beneath the agentic AI and provides the skills, procedures and protections to address the concerns above, removing the need for a lab to develop them independently. This includes (1) infrastructure to support projects and so-called choreographies} involving multiple group members; (2) a set of specialized agents each dedicated to typical biomedical data science tasks including software and statistical development, data visualization, literature search, equity, diversity, inclusion, and decolonization (EDID) review, and others; (3) a tiered memory which can retain important research decisions, intermediate, derived outputs and sensitive data, and which can exploit this information to automatically build better contexts in AAI sessions; (4) traceability records which are used by the AAI session to better plan and execute data analyses and software builds by, for example, helping to avoid repeating decisions that lead to dead-ends; (5) enforcement of SOPs for data maintenance and data governance across all lab members; and (6) multi-user capacity to allow groups to interact and collaborate. We use the system to identify putative inhibitors of Peptidyl-prolyl cis-trans Isomerase NIMA-interacting 1 (Pin1), a protein with a shallow catalytic site making it difficult to target. We describe the construction of several approaches to identify Pin1 inhibitors and the results they yield. Murmurent is open source.

Graph embedding transforms complex networks into low-dimensional representations for analysis and downstream machine learning. Existing methods often rely on eigenvalue ordering, random-walk sampling, or neural optimization. We introduce MOSAIC, Multistep Operator Spectral Accumulation with Isolated Components, an unsupervised framework that constructs an accumulated diffusion operator from successive powers of a graph transition matrix. MOSAIC uses Krylov subspace analysis to identify dominant spectral components according to their convergence behavior. This approach emphasizes persistent and spectrally separated diffusion modes rather than selecting embedding coordinates solely by eigenvalue magnitude. We evaluated MOSAIC using a biomedical graph projection derived from the Data Distillery Knowledge Graph that contained 5,171 human genes and 10,147 phenotypes. Across embedding dimensions from 16 to 128, MOSAIC achieved node-type classification AUROCs from 0.9863 to 0.9926 and exceeded FastRP, Node2Vec, HashGNN, and GraphSAGE. For gene-phenotype link prediction, validation and test associations were withheld before embedding generation. Under progressive removal of all edges or specific edge classes, MOSAIC retained more node-label information than FastRP, including after extreme graph sparsification. In a focused DICER1 analysis, 106 of 137 documented phenotype associations received scores above 0.90, demonstrating concordance with known pleiotropic biology and prioritizing additional relationships for independent evaluation. These results support MOSAIC as a fully unsupervised framework for learning informative graph representations. Although demonstrated using a biomedical graph projection, MOSAIC is domain independent and can be applied to other complex networks

Recent advances in artificial intelligence have enabled models to capture protein sequence and structural features. We present PeTriPPI2, a hybrid framework for protein--protein interaction (PPI) prediction that combines sequence embeddings from ESM-2 with structural representations from PeTriBERT, an encoder originally designed for inverse folding. We evaluate this sequence--structure approach on the Pinder dataset. On a test set of 2,342 protein pairs, the selected checkpoint achieves 0.898 accuracy, 0.917 precision, 0.876 recall, and 0.896 F1 score for the interacting class (AUROC 0.963). Compared with SpatialPPIv2 on the same test data, PeTriPPI2 has higher precision (0.917 versus 0.886) but lower recall (0.876 versus 0.942) and F1 score (0.896 versus 0.913). Component ablations indicate that both the sequence and structural pathways contribute to this configuration.

Genotype imputation offers a biological application for optimization methods that may eventually run on larger quantum computers. We developed QuImputer to encode allele-frequency priors and signed linkage disequilibrium (LD) in a reduced Ising/QUBO objective, with only missing genotypes represented as optimization variables. The same objective can be passed to exhaustive classical search, quantum approximate optimization algorithm simulation, or quantum hardware. We evaluated the formulation on 100 rice accessions across 12 chromosomes, estimating LD and allele frequency from the same post-masking observed matrix supplied to Beagle 5.4. Evaluation was restricted to known truth at sites with exact dosage coding and an alternate allele that was not the major allele, with MAF below 0.50, yielding 1.88 billion scored genotypes. Beagle achieved higher rare-ALT sensitivity (2.52% versus 0.54%), F1 (0.047 versus 0.011), Matthews correlation coefficient (0.088 versus 0.044), and pooled hard-call dosage R-squared (0.231 versus 0.014). QuImputer produced fewer false ALT calls, with Type I error of 0.038% versus 0.223%. Analyses of a chromosome-1 subset showed limited local LD support for many masked rare-ALT carriers. In five small IBM Fez validation problems, the selected minimum-energy outputs matched unique all-reference exact optima; the pooled frequency of those optima was 1.21-2.63%. Further IBM Fez executions covered eleven sparse instances with 12-97 variables and a separate 120-variable instance, retaining all measurements. Selected outputs recovered the exact optima in eight instances through 49 variables. Larger instances remained below the all-reference baseline; noiseless comparisons distinguished finite-shot sampling from deviations in hardware output. QuImputer provides an LD-informed formulation and a route to quantum-hardware execution, together with an empirical assessment of the limitations of the evaluated configuration. These results establish a basis for studying richer information models and alternative classical or quantum solvers.

Motivation: Enrichment analysis is widely used to extract biological meanings from gene lists. However, the large number of available tools and reference libraries often make the interpretation of results challenging. Inexperienced users lack clear guidance on which datasets and terms are most relevant to their data, whereas experienced researchers must navigate large result sets and consult multiple resources to identify biologically relevant findings. Integrating AI into this process can help to bridge these gaps. Results: We present Enrich.AI, a conversational biology assistant that integrates statistical enrichment analysis with automated biological interpretation. The agent operates on an LLM-based ReAct architecture and dynamically selects among the tools for enrichment (220+ gene set libraries), literature search (40+ million publications via Europe PMC), gene annotation, and database retrieval. Given a gene list and a short description of the biological context, Enrich.AI produces a written interpretation grounded in the retrieved literature and six interactive, configurable visualizations. A conversational interface enables human-in-the-loop refinement, letting researchers without programming experience explore results iteratively. The reliability of these interpretations was validated through a systematic evaluation on 50 gene sets, demonstrating high faithfulness to retrieved evidence and a low hallucination rate.

Protein inverse folding, which infers compatible amino acid sequences from a given backbone structure, is central to rational protein design. Current methods rely largely on sequence recovery rate as the evaluation metric, emphasizing structural compatibility but seldom addressing function. Yet functional activity depends on the proper network of non-covalent interactions-protein-ligand contacts being a prime example. We argue that function-oriented protein design is governed by three complementary constraints: structural, interaction, and evolutionary, none of which are handled in a unified manner by existing pretrained models. To bridge this gap, we introduce zPolaron, a graph neural network built on protein-ligand complexes for functional protein design. zPolaron represents residues and ligand atoms as a heterogeneous graph and jointly encodes structural, interaction, and evolutionary information by fusing physical-interaction-enhanced graph representations with ESM-2 embeddings. Across multiple fitness prediction benchmarks, zPolaron outperforms all baselines, with gains evident across structural, interaction, and evolutionary dimensions. Wet-lab experiments further confirm its utility in functional enzyme screening and design: the optimized CalA6-D205A-S160G mutant achieves a 17.84-fold improvement in activity, while the CalB-I121M mutant yields a 1.91-fold increase. Together, these results show that integrating interaction and evolutionary information fills the missing puzzle of inverse folding, enhancing protein fitness prediction beyond structural compatibility. zPolaron is available at https://github.com/zelixirSH/zPolaron.

Background and Objective: Prescribing multiple drugs simultaneously, known as polypharmacy, is increasingly common in managing chronic diseases such as diabetes and cardiovascular conditions. Studies show that roughly 40% of elderly patients are on five or more medications, putting them at significant risk of adverse drug-drug interactions (DDIs) [1]. While computational methods based on Graph Neural Networks (GNNs) have shown strong potential for predicting these interactions from a drugs molecular structure, most existing models are limited in two important ways: they rely on attention mechanisms that cannot adapt to the unique context of each atom in a molecule, and they only predict whether an interaction exists rather than how dangerous it might be. Methods: We propose HMVA-DDI, a two-stage deep learning framework that addresses both shortcomings. In the first stage, our model predicts whether two drugs will interact. If it detects an interaction, the second stage classifies it into one of three clinical severity levels: minor, moderate, or major. To better capture molecular structure, we replace the standard attention mechanism with a more powerful version called Graph Attention Network v2 (GATv2), which considers each atoms specific context when learning its representation. We also introduce a technique to prevent information loss in deeper network layers, and we tailor the input features separately for each prediction task. Results: Evaluated on two standard benchmarks, DrugBank 5.0.3 [2] and DDInter [3], HMVA-DDI achieves an AUROC of 0.9819 and AUPRC of 0.9759 on binary interaction detection, a marked improvement over the baseline. For severity classification, our best model variant (qcSvr4) reaches an overall AUROC of 0.9444, with a particularly large improvement in detecting low-severity (Minor) interactions, a historically difficult class. Conclusion: HMVA-DDI bridges the gap between computational drug research and clinical usefulness by not only predicting interactions accurately but also indicating interaction severity. This two-level system can help healthcare professionals prioritize warnings, reducing the alert fatigue problem where doctors are overwhelmed by too many non-critical notifications.

Abstract Accurate taxonomic classification in metagenomics is often compromised by low-complexity sequences, which lead to chance matches that in turn cause sequences to be misclassified. Here we present Kmask, an entropy-based masking tool implemented for use either standalone or as part of Kraken [1,2] database construction, which replaces low-entropy regions with Ns. Using a sliding window size aligned with Kraken's default k-mer length and parameters optimized across 12 control bacterial genomes spanning a broad range of GC content, Kmask efficiently removes low-complexity sequences while retaining high-complexity regions. To benchmark performance, we applied Kmask to a newly constructed database, Microbial2025, that contains over 71,000 bacterial, archaeal, viral, and fungal genomes, and we then classified human reads against both masked and unmasked versions of the database using KrakenUniq [2]. We found that Kmask substantially reduced misclassifications, driving down the false positive rate to 5.78% from 7.52%. Notably, Kmask performed comparably to an SDUST-masked [3] database, achieving a similar false positive rate (5.78% vs. 5.17%) while masking out fewer bases (1.33% vs. 1.85%). We also tested Kmask on a database of human cancer sequences, where we found that it eliminated many false positives caused by low-complexity matches between bacterial genomes and human DNA. These results demonstrate that Kmask is an effective method for masking low-complexity sequences in large microbial databases, thus improving the accuracy of metagenomic classification.

In network medicine, the protein-protein interaction network, or interactome, is an essential resource for identifying candidate proteins underlying diseases and other phenotypic traits. Indeed, it can be leveraged via the guilt-by-association paradigm, which asserts that interacting proteins are likely to participate in the same molecular processes. Network propagation that combines the topology of the interactome with prior knowledge about disease genes is a promising strategy to identify new candidate genes contributing to diseases. However, existing network propagation algorithms are often biased toward highly connected proteins, called "hubs". We present BFWalk, a novel network propagation algorithm designed to avoid inflated scores for hubs. We tested BFWalk across four human phenotypes: multiple morphological abnormalities of the sperm flagella, dyschromatopsia, hypertrophic cardiomyopathy, and chronic kidney disease. For each phenotype, we assessed whether BFWalk could recover known phenotype-associated genes and whether new candidate genes were enriched in relevant tissues. Depending on the phenotype, BFWalk outperformed or matched the state-of-the-art network propagation methods. Furthermore, the results confirmed that BFWalk is free of bias toward high-degree nodes. Therefore, BFWalk is a powerful method for identifying new genes across diverse phenotypes and constitutes a strong alternative to existing network propagation algorithms. Availability: BFWalk is implemented in Python and C. The code is available under the GNU GPL (v3.0) on: https://github.com/jedrzejkubica/BFWalk

Motivation: Precursor microRNA (pre-miRNA) prediction is used to prioritize genomic hairpins for miRNA annotation. Published predictors are trained and evaluated on different species collections, negative sets and data partitions, making reported gains difficult to compare. Results: We introduce pre-miRBench, a benchmark of 7,056 200-nt windows centered on curated pre-miRNA loci and 70,560 genomic negatives with predicted hairpin structures from 71 animal species, with a fixed 1:10 positive-to-negative ratio in every partition. Four test sets cross whether the species and miRNA family are represented in training: known species/known family, known species/held-out family, held-out species/known family, and held-out species/held-out family. We retrained six published pre-miRNA predictors (DeepMir, deepMiRGene, dnnPreMiR, mirDNN, miRe2e and MuStARD) under the same benchmark splits and evaluated them alongside the pre-miRBench model, an ensemble of three neural networks developed with Agentomics, an autonomous LLM-agent system for biomedical machine learning. The pre-miRBench model achieved the highest mean average precision (AP) across the four test sets (0.9809; dnnPreMiR, 0.9719) and led three tests. The two held-out species were Gallus gallus and Drosophila melanogaster, and the joint species-and-family holdout contained 69 positives and 690 negatives. dnnPreMiR had the highest AP on the joint species-and-family holdout (0.9677 versus 0.9659). The AP difference (pre-miRBench minus dnnPreMiR) was -0.0018, with a paired bootstrap 95% confidence interval of -0.0445 to 0.0341. pre-miRBench establishes a common evaluation for species and miRNA-family holdouts and provides a reproducible reference model for future comparisons. Availability and implementation: Reproducible source code is available at https://github.com/dimostzim/pre-miRBench. Datasets, trained model weights, record-level predictions, evaluation metrics and species-panel metadata for pre-miRBench and the six retrained predictors are available at https://zenodo.org/records/22813044.

Motivation: Traditional methods primarily rely on statistical features such as k-mers and GC content, making it difficult to capture complex internal relationships within sequences. Deep learning models typically rely on discrete encodings, leading to issues such as information sparsity, dimensional redun-dancy, and disruption of sequence continuity. Our previous UniWave-1 framework transforms se-quences into waveform signals; however, it lacks comprehensive physicochemical integration and com-putational efficiency. Results: We propose the UniWave-2 feature extraction framework. UniWave-2 incorporates three key advancements: (i) integration of hydrophobicity into the encoding scheme; (ii) enhancement of wave-form resolution through Fourier transform, coupled with wavelet transform to precisely capture both local details and global periodic patterns; and (iii) development of a lightweight multi-scale GRU model that facilitates cross-dimensional feature interactions, captures bidirectional temporal dependencies, and integrates long- and short-range patterns with spatial positional awareness, thereby yielding rich feature representations. UniWave-2 achieved competitive performance across multiple genomic tasks, while ISM analysis demonstrated that waveform encoding enables interpretable attribution without re-lying on gradient information. These results establish UniWave-2 as a lightweight and interpretable framework for nucleic acid sequence analysis.

The advent of high-throughput sequencing technologies has made it essential to employ advanced tools for data integration and interpretation. In this work, we propose to build upon two existing efforts by expanding their scope and applicability to multi-omic data integration through the development of a pipeline that generates a topologically informed ranking of genes to assess gene relevance to disease. Specifically, we exploit topological data analysis, capitalizing on: its graph-based framework, to refine complex interactions characteristic of biological systems; its intrinsic explainability, to enable interpretation of model outputs in terms of biologically meaningful features; its robustness to small input perturbations, to ensure stable results in the presence of noisy data, as omics are. In particular, we employ Harmonic Persistent Homology (HPH), and propose two relevant advances. First, we adopt an integration and clustering tool (i.e. iNETgrate) as an informed means to reduce data size and integrate multiple omic layers by aggregating methylation beta values from loci to the gene level, to be combined with gene expression values. This enables the application of HPH, otherwise limited by the computational burden. Second, as a non-trivial consequence, we apply HPH to multi-omic molecular data, and not to patients as done so far, enlarging HPH scope. Finally, we follow up on recent efforts toward performance standardization, and validate our results against biomarkers from a well-known TCGA breast cancer benchmark dataset.

Learning from imbalanced data remains a major challenge for graph neural networks (GNNs), as minority nodes are not only rare but also structurally marginalized within the graph. We address this issue with CLARA, a hierarchical learning framework that decomposes node classification into two stages: a coarse subgraph-level classifier that selects regions likely to contain minority instances, followed by a fine-grained node-level predictor within these regions. This design improves sensitivity while maintaining scalability, filtering out irrelevant areas and focusing learning on topologically meaningful neighborhoods. Experiments on benchmark graph datasets demonstrate substantial gains over established imbalance-handling methods, with CLARA reaching an F1-score of 88.3\%. The same strategy achieves significant improvements in protein-ligand binding site prediction, underscoring its broad and consistent effectiveness across both biological and general graph learning tasks.

Phenotype prediction from public gene-expression data is straightforward to implement but reliable prediction is harder. Leakage during feature selection can inflate performance estimates. Class imbalance can hide poor performance in minority classes. Confounds like tissue type, repeated sampling, and demographic differences can also create patterns that resemble phenotype-associated biology. Here, I present LOL (Latent Omics Learning), a Python package for phenotype prediction from bulk GEO RNA-seq and microarray data. LOL uses leakage-safe preprocessing, imbalance-aware evaluation, and group-aware validation where repeated samples are present. Each analysis produces two audit diagnostics. The leakage inflation index ({Delta}LII) measures the change in performance between naive and leakage-safe evaluation. PVCA-lite examines associations between the latent representation and available metadata covariates, including phenotype. I evaluated the workflow across six independent GEO cohorts. The datasets covered microarray and RNA-seq platforms, binary and multi-class phenotypes, and sample sizes ranging from 18 to 566. The analyses produced a varied set of outcomes. Across diverse GEO datasets, LOL can reveal phenotype-associated expression patterns while identifying factors distort their interpretation. These include information leakage, class imbalance, repeated sampling, tissue differences, and demographic or other metadata-associated confounding. The resulting diagnostics help distinguish robust predictions from results that require further scrutiny. LOL does not introduce a new representation-learning algorithm. Its contribution is a reproducible workflow that places validation and diagnostic checks alongside prediction. The resulting analysis can expose weaknesses in an apparently convincing result and make those limitations visible. LOL is freely available under the MIT license from https://github.com/ngangao/lol-omics.

Accurate modelling of biomolecular interactions is fundamental to drug discovery, yet current artificial intelligence (AI) workflows remain fragmented across structure prediction, affinity estimation, molecular design, and experimental decision-making. We introduce AnewDDE, an agentic Drug Discovery Engine that connects these capabilities into a closed-loop system for biomolecular interaction modelling and design. We demonstrate several components of this system: AnewFold delivers superior performance on challenging targets, including antibody-antigen complexes and molecular glues, providing more reliable structural hypotheses together with pocket identification and conformational analysis; AnewDesign achieves a 10.7% success rate in identifying binders with single-digit-nanomolar affinities measured by surface plasmon resonance (SPR) in a representative nanobody discovery campaign, substantially reducing the experimental search burden; AnewAffinity combines accuracy approaching that of free-energy perturbation (FEP) methods with the speed required for high-throughput screening and prioritizes promising transformations; and AnewMind, a large language model (LLM) at the hundred-billion-parameter scale that underwent in-house full-parameter post-training, provides scientific reasoning and absorption, distribution, metabolism, excretion, and toxicity (ADMET) prediction for pharmaceutical research and development (R&D). To systematically evaluate AnewMind, we introduce PharmBench, an internal benchmark designed by pharmaceutical R&D experts to evaluate long-horizon decision-making in drug discovery. Evaluations on ADMET tasks and PharmBench show that AnewMind performs competitively with leading frontier models. Together, these capabilities establish AnewDDE as an integrated engine for addressing challenging therapeutic targets and translating molecular insights into experimentally validated candidates at scale.

This paper predicts antimicrobial resistance (AMR) from matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectra using DRIAMS, evaluated across 13 species-antibiotic datasets spanning sensitive/resistant imbalance ratios of 2.95:1 to 141.21:1 and raw spectral similarities of 0.9166-0.9946. We propose HetSAGE, a heterogeneous graph neural network (GNN) that combines full-spectrum "raw" edges with consensus "biomarker" edges derived from a 4-method feature-selection vote and applies per-edge-type dropout to regularize the two views separately. HetSAGE outperforms a plain multilayer perceptron (MLP) only when spectral similarity is high, consistent with oversmoothing effects reported in the broader GNN literature; the biomarker view converges on features close to independently established clinical markers: within {+/-}1 Da in Staphylococcus aureus, Oxacillin, and within range in Klebsiella pneumoniae, Meropenem, confirming it is not a black box. At the same time, per-edge-type dropout protects this smaller signal from dilution by regularization tuned for the noisier raw spectrum. Together, these results indicate that heterogeneous graph structure is not a universal win for MALDI-TOF AMR prediction but a conditional one, tied to a measurable property of the data, spectral similarity, rather than to architecture alone. This caveat is largely unaddressed in current DRIAMS-based GNN work. It provides a concrete signal for practitioners deciding when a graph model is worth the added complexity over a plain MLP.

An Ischemic stroke is one of the main causes of death and long-term disability around the world. There are very few treatment options available, especially for the cases where the short time period to give clot dissolving medicine has passed (Powers et al., 2019). While single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool for understanding cell-to-cell differences within at issue (cellular heterogeneity) (Tang et al., 2009). However translating these high dimensional datasets into clinically useful biomarkers remains challenging (Rajko-mar et al., 2019). A major problem with current biomarker studies is their reliance on a single machine learning model, where results may reflect model bias rather than true biological signals (Lopez-Rincon et al., 2020). Methods: We developed a multi-model consensus framework integrating eight different machine learning methods from three groups: traditional models (logistic regression, random forest, XGBoost), a hybrid deep learning model (CNN-XGBoost), and transformer-inspired models (scGPT, scBERT, sc-Former, Geneformer). The proposed framework was applied to scRNA-seq data from a mouse middle cerebral artery occlusion (MCAO) model (GSE174574; n=6 mice, 54,599 cells). To prevent data leakage, we implemented rigorous subject-wise train/test splitting, ensuring all cells from individual animals remained either in the training sets or in the testsets. Results: All eight models achieved excellent performance on the independent test data, with AUC scores between 0.9778 and 0.9975. From six models that allowed gene level importance extraction, we identified 63 consensus genes (threshold: [≥] 4/6 models), with 17 genes selected by all 6 models. Pathway enrichment analysis showed significant enrichment of interferon signaling, with a novel dual-interferon pattern. Twelve of the consensus genes respond to interferon, and five of them (Ifitm6, Ifitm1, Gbp2, Igtp, Tgtp1) respond to both Type I (IFN-/{beta}) and Type II (IFN-{gamma}) interferons. Notably the consensus included Il1rn (interleukin-1 receptor antagonist), the target of Anakinra -- an FDA-approved drug that has completed Phase II stroke trials with promising safety profiles (Smith et al., 2018; Emsley et al., 2005). Conclusions: The Multi-model consensus framework, proposed in this paper, provides a robust and reliable approach to biomarker discovery that goes beyond the limitations of any single method and reduces individual algorithm bias. The convergence of classical, ensemble, and deep learning methods on interferon and IL-1 pathway genes supports their biological validity as stroke biomarkers. Our findings provide a prioritized gene list for further clinical validation and highlight interferon signaling as a possible treatment strategy warranting further investigation. Keywords: ischemic stroke; machine learning; single-cell RNA sequencing; biomarker discovery; consensus analysis; interferon signaling; neuroinflammation