Biology subjects

Liu, E.

Publications and source records attributed to Liu, E..

At least 19 recordsLinked to original sources

bioRxiv2026

CAI-Vax: Computational Design of Personalized mRNA Cancer Vaccine Constructs After Antigen Selection

CAI-Vax (Cancer Antigen Individualized Vaccine Design) arranges selected antigens and generates synonymous coding sequences for personalized mRNA vaccine constructs. We evaluated 357 vaccine-candidate peptides from 31 patients across four published cohorts. At the primary cumulative MHCflurry percentile-rank threshold of [≤]2%, HLA-aware search reduced predicted junction-spanning peptide-allele binder occurrences by 3.49 per junction relative to amino-acid-pattern optimization (paired Hodges-Lehmann estimate; nominal 95% interval 2.75-4.48; Holm-adjusted P = 7.45 x 10-). All 31 primary paired differences and the estimated effects in all 16 threshold-by-normalization settings favored HLA-aware search within the MHCflurry model used for optimization. Among 25 patients with an exact solution for the additive search objective, single-start greedy search found the minimum in 5 patients and the best individual heuristic in 16. Selecting each patients best result across all tested heuristics increased this to 19. MHCflurry and NetMHCpan EL identified identical tied best-plan sets in 6 of 31 patients after rescoring six frozen plans per patient. In a retrospective internal analysis, repeat-aware synonymous encoding reduced maximum 10-mer multiplicity in every patient, with mean codon-use-preference losses of 0.0041-0.0055 across cohorts; however, it performed worse than matched random encoding for 6-mers. These findings demonstrate that HLA-guided construct optimization can reduce predicted junctional binding within the optimization model, although the preferred antigen arrangement depends on the prediction model used. Experimental studies are needed to determine the biological and manufacturing relevance of these computational differences.

bioinformatics↗

bioRxiv2026

Genomic characterization of patient derived xenograft models of multiple myeloma reveals selection of multiple pre-existing high risk clones

Multiple myeloma (MM) is a genomically complex malignancy of terminally differentiated plasma cells. Here, we have utilized the SCID-rab patient derived xenograft (PDX) model as a platform to expand scarce primary MM samples across disease stages and major cytogenetic subgroups for comprehensive molecular analysis. The model generated abundant CD138+ cells while retaining key genomic, transcriptomic and proteomic features of corresponding patient tumors. Compared with human myeloma cell lines, PDX cells more closely resembled primary MM in proliferative state and expression of therapeutically relevant targets. WGS and single cell WGS showed that PDXs arose from the selection of pre-existing subclones, including clones harboring high-risk alterations and convergent mechanisms of tumor suppressor inactivation. These findings establish SCID-rab PDX models as a platform for multiomic profiling, studies of clonal evolution, and determining therapeutic response in MM.

cancer biology↗

bioRxiv2026

Extended amygdala function is governed by threat probability, not uncertainty

Anxiety is a universal experience that arises as we anticipate negative events that may or may not happen. When extreme, anxiety is a major cause of distress, suffering, and disability-adjusted life years, with nearly one in four individuals experiencing clinically relevant anxiety in their lifetime1-3. Elucidating the biological basis of anxiety has the potential to help people understand, cope with, and ultimately receive improved treatments for their anxiety. The canonical neuroscientific model of anxiety posits that distinct regions of the extended amygdala (EA) elicit distress that is conditional on threat uncertainty4. Here, we refute this model using a precise computational framework to disambiguate specific dimensional facets of threat uncertainty in a racially diverse sample of 104 individuals. We show that self-report, psychophysiological, and neural correlates of threat anticipation are much more strongly associated with threat probability than uncertainty per se. Importantly, EA regions, including the central nucleus of the amygdala (Ce) and bed nucleus of the stria terminalis (BST), were most reliably associated with threat probability--contrary to the canonical model. Re-examination of the pioneering experiments that motivated the canonical model reveal a previously unappreciated confound between probability and uncertainty, suggesting the novel hypothesis that EA regions differ in their sensitivity to low-probability threats. Our fMRI data support this, showing that BST is more sensitive to low-probability threats than Ce. Moreover, lesion experiments in rhesus monkeys (Macaca mulatta) demonstrate that BST causally contributes to defensive behavior in low-probability threat contexts. Together, these data reconcile previous empirical inconsistencies, provide a fresh theoretical perspective on the neurobiology of anxiety, and underscore the theoretical importance of a formal lingua franca for coordinated cross-species emotion research.

neuroscience↗

bioRxiv2026

Architectural trade-offs between environmental stability and genomic redundancy reveal divergent pneumoviral entry strategies

Human respiratory syncytial virus (RSV) and human metapneumovirus (hMPV) exhibit distinct seasonal epidemiology, with RSV circulating in early autumn and hMPV peaking in midwinter, yet the structural basis for this niche partitioning remains undefined. Here, we integrate in situ cryo-electron tomography and functional virology to decode the architectural logic governing their entry dynamics. RSV employs a matrix (M)-regulated prefusion F (pre-F) organization, partitioning trimers into stabilizing hexagonal superlattices and fusion-competent pools to maintain superior thermotolerance. By contrast, hMPV compensates for its intrinsically unstable, monomeric pre-F with extreme ribonucleoprotein polyploidy, packaging [~]4-fold more genome equivalents to ensure productive infection. Fusion events localize exclusively to M-depleted, non-arrayed membrane regions, establishing a spatial checkpoint for activation. These findings reveal a conserved trade-off between environmental resilience and genomic redundancy that dictates divergent pneumoviral entry strategies, explaining their distinct seasonal ecological niches.

microbiology↗

bioRxiv2026

The physiological dynamic clamp allows insect flight muscle to transition between two actuation modes in virtual reality

In most muscles, contraction is initiated by neural activation. Some groups of insects break this rule, flapping at frequencies far exceeding the neural drive to their flight muscles. These insects muscles (termed asynchronous) produce force in response to stretch, enabling flight at faster frequencies than would be possible through the slow calcium-dependent processes associated with neural activation. The first flapping insects lacked stretch-activated physiology, which then evolved on top of neural activation dynamics before likely being reduced again in some groups including moths. Stretch and neural activation can co-exist, but it remains unclear if stretch-activation alone is sufficient to generate asynchronous flapping in insect flight muscle. Building on prior closed-loop muscle physiology platforms, we develop a new way to perform a gain-of-function muscle physiology experiment called the physiological dynamic clamp. Inspired by dynamic clamp experiments in neuroscience, we couple isolated intact flight muscle from a hawkmoth, Manduca sexta, to simulated stretch-activation in virtual reality. Tuning virtual reality parameters allows us to manipulate the degree of stretch-activation in silico while retaining all other physiological properties of the muscle. With artificially enhanced stretch activation, we find that hawkmoth muscle can support stretch-activated work at typical wingbeat frequencies. When simultaneously stimulated at wingbeat frequency, interference between stretch and neural activation results in variable work production. However, this interference disappears when the two activation timescales are close to each other resulting in entrainment to the neural drive. Matching time scales suggests an evolutionary path for smoothly transitioning to stretch-activated, asynchronous flight and back again.

physiology↗

bioRxiv2026

A hierarchical activation threshold shapes RSV F antibody repertoires from toddler vulnerability to adult protection

Respiratory syncytial virus (RSV) causes severe infant morbidity, yet the mechanisms underlying their suboptimal immunity and the failure of prefusion F-stabilized adult vaccines in infants remain unclear. Here we integrate deep mutational scanning of 761 preF-binding antibodies, repertoire profiling of 102 repeatedly exposed pediatricians and 61 RSV-experienced toddlers, structural analysis, and >40,000 viral genomes to decode RSV F immunity. We resolved 12 immunologically distinct functional antibody subclasses and uncovered a hierarchical activation threshold: apical epitope-targeting antibodies with superior neutralizing potency demand extensive somatic hypermutation and cooperative CDR networks, whereas central-to-basal epitope-directed, marginal or non-neutralizing antibodies engage germline-encoded antibodies through minimal mutations (e.g., S31G). Toddler repertoires are confined to low-threshold, non-productive sites; adult repertoires enrich for apical elite neutralizers whose immune pressure drives contemporary RSV-B evolution. These reframe pediatric RSV vulnerability as a threshold-gated repertoire deficit and prescribe vaccine strategies that actively redirect immunodominance from permissive epitopes toward high-barrier apical targets.

immunology↗

bioRxiv2026

OCellus: A Language-Model Framework for Single-Cell, Spatial, and Perturbation Biology with Natural-Language Reasoning

Computational modeling of cellular behavior--the virtual cell--has emerged as a stated grand challenge at the intersection of artificial intelligence and biology, yet existing foundation models remain specialized: single-cell models process dissociated transcriptomes only, spatial models require dedicated spatial-aware architectures, and perturbation predictors depend on manually curated knowledge bases that cap generalization. Here we introduce OCellus, a single nine-billion-parameter language model (Qwen3.5-9B) fine-tuned on twenty-two biological tasks that simultaneously addresses all three limitations through three coordinated technical contributions on a shared backbone. First, EvenClock encodes two-dimensional spatial coordinates as eighteen clockface sectors of text, enabling spatial reasoning on a vanilla language model without architectural modification; on ten spatial transcriptomics tasks OCellus attains 77 percent spatial-neighborhood accuracy, 96 percent spatial-cellchat accuracy, and 0.70 proportion-cosine similarity on spatial deconvolution, all without any spatial-aware architectural components. Second, per-gene language-model embeddings replace the Gene Ontology annotations that GEARS depends on, achieving Pearson correlation 0.945 on the Replogle 2022 perturbation benchmark versus 0.84 for GEARS across 457 completely unseen knockout genes. Third, OCellus-Agent provides a Planner-Router-Verifier natural-language interface that achieves 75 percent pipeline accuracy on eighty multi-task queries. Removing language-model embeddings collapses perturbation Pearson to 0.06, confirming that learned functional representations--not graph topology--drive the gain. As a cell-type encoder, OCellus ranks first among fourteen foundation models in linear-probe accuracy at 95.1 percent across four benchmark datasets, and reaches 72.6 percent average across twenty-two evaluated biological tasks--a 57-percentage-point absolute gain over the strongest baseline configuration. As a language model, OCellus uniquely generates natural-language explanations of its predictions, a capability absent from all competing methods. Code, pre-trained model weights, the graph-neural-network module, and the agent system will be made available upon publication.

bioinformatics↗

bioRxiv2026

Bulk and single-cell transcriptomics reveal elevated endogenous retrovirus expression linked to immunopathology in severe COVID-19.

SARS-CoV-2 infection triggers expression of endogenous retroviruses (ERVs), but whether this persists during COVID-19 and contributes to disease severity has not been well characterized. In this study, we analyzed bulk and single-cell RNA-sequencing datasets from multiple cohorts of severe COVID-19 patients to investigate the role of ERVs in COVID-19 immunopathology. We identified 33 differentially expressed proviral ERV loci in severe COVID-19 compared with healthy individuals (severe COVID-19 signature ERVs). ERV activation was associated with dysregulation of ERV epigenetic regulators, and their expression strongly correlated with key inflammatory pathways implicated in severe COVID-19, including neutrophil degranulation, interleukin signaling, and inflammasome activation. Anakinra treatment reversed activation of signature ERVs. Six signature ERVs were specific to intensive care unit (ICU) admission and significantly correlated with hospital-free days at day 45 (HFD-45). At the single-cell level, signature ERVs were significantly upregulated in erythroid-like and erythroid precursor cells and macrophages of patients with severe disease. Within these cells, we found evidence of ERV-associated differences in inflammatory gene expression, whereby cells expressing signature ERVs showed heightened expression of innate immune genes compared with cells not expressing these ERVs. Together, our study unmasked specific ERV loci activated in severe COVID-19 that are linked to immunopathology.

immunology↗

bioRxiv2026

mirCCC: Repression-aware graph learning for miRNA-mediated cell-cell communication inference

Cell-cell communication analyses usually focus on protein ligands and receptors and therefore miss extracellular vesicle transfer of microRNAs, an important route of signalling in cancer. Here we show that microRNA-mediated communication can be inferred from standard single-cell RNA sequencing by detecting coordinated decreases in the expression of validated miRNA target genes. We developed mirCCC, a computational framework that estimates cell-specific microRNA activity, models cellular sending and receiving capacity for extracellular vesicle transfer, and learns microRNA-resolved communication graphs from transcriptomic data. In synthetic benchmarks with strong confounding signals, mirCCC improved whereas all comparison methods declined. Applied to a human colorectal cancer atlas, mirCCC recovered known colorectal cancer-associated microRNAs and identified stromal- and myeloid-to-epithelial communication converging on a plasticity program linked to TGF-{beta} and Wnt/{beta}-catenin signalling. These results provide a practical route for studying extracellular vesicle-mediated communication in existing single-cell atlases. Author summaryCell-cell communication plays important roles across a wide range of biological processes. Most studies of cell-cell communication focus on interactions between ligands and receptors. However, cells can also use extracellular vesicles to deliver microRNAs (miRNAs), which regulate gene activity in receiving cells and contribute to cancer progression, immune regulation, and changes in the tissue environment. Since standard single-cell sequencing does not directly capture miRNAs, this mode of communication remains largely unexplored. Here we developed mirCCC, a computational method for inferring miRNA-mediated communication from standard single-cell transcriptomic data. Instead of detecting miRNAs directly, mirCCC examines the effects they leave in receiving cells. Active miRNAs often cause coordinated suppression of their target genes. mirCCC uses these coordinated reductions in the expression of known target genes to estimate miRNA activity and combines them with the ability of sending cells to package and release miRNAs and of receiving cells to process them. It can therefore reconstruct directional communication networks for individual miRNAs. Our evaluations show that mirCCC remains reliable in the presence of misleading signals and reveals biologically meaningful regulatory patterns in human colorectal cancer data. Overall, mirCCC provides a practical way to study miRNA-mediated cell-cell communication using existing single-cell transcriptomics datasets.

bioinformatics↗

bioRxiv2026

A Spatial Proteomic Atlas of Tertiary Lymphoid Structures in Non-Small Cell Lung Cancer Identifies a Novel Predictive Class of Lymphoid Aggregates

Tertiary lymphoid structures (TLS) predict benefit from immune checkpoint inhibitors (CPIs), yet mature, germinal-center-rich TLS are infrequent in solid tumors by histological review. Here, using 38-plex MIBI spatial proteomics across 165 lymphoid structures from 14 NSCLC resections, we establish a continuum of TLS maturity using high dimensional compositional, spatial, and molecular features. We demonstrate that histologically-defined lymphoid aggregates (LA) comprise a heterogeneous class of structures, which span this continuum of maturity. We identify a subset of lymphoid aggregates that harbor follicular dendritic cell networks, T follicular helper cells, and activated B cell states characteristic of mature TLS, yet are not readily distinguished from other LA structures in our histological review. We developed a novel digital pathology classifier to identify mature LAs in CPI trials, and demonstrate in a retrospective analysis of Atezolizumab in advanced NSCLC that the inclusion of mature LAs greatly expands the biomarker-eligible population while maintaining strong predicted benefit. Together, these data redefine the biological spectrum of tumor-associated lymphoid aggregates and provide a framework for implementing maturity-informed TLS biomarker strategies.

cancer biology↗

bioRxiv2026

An apical junction protein antagonizes mechanosensitive calcium signaling to establish stochastic choices of olfactory neuron subtypes

Mechanical forces regulate brain development and left-right body patterning. However, the role of mechanical signaling in brain lateralization remains unclear. In Caenorhabditis elegans, the left and right AWC olfactory neurons communicate via a gap junction network to stochastically differentiate into the default AWCOFF and induced AWCON subtypes. SLO BK potassium channels, SLO-1 and SLO-2, act redundantly to inhibit a calcium-regulated protein kinase pathway in the specification of the AWCON subtype. Here, we identified a role for AJM-1 (apical junction molecule 1) in promoting AWCON from an unbiased forward genetic screen for mutants that suppress the slo-1(gf) 2AWCON phenotype. AJM-1 is located at three distinct tight junctions between amphid neurons (including AWC) and sheath glia, sheath and socket glia, and socket glia and hypodermal cells (also known as epidermal cells) at the anterior tip of the animal. In addition to its cell-autonomous function, the non-cell-autonomous function of AJM-1 in glial and hypodermal cells is required for the specification of the AWCON subtype. Furthermore, we identified a role for the DEL-1 mechanosensitive DEG/ENaC channel in the calcium signaling pathway, mediated by UNC-2 and EGL-19 voltage-activated calcium channels, that specifies AWCOFF. Together, our results suggest a mechanism in which AJM-1 promotes SLO-1 expression and antagonizes mechanosensitive calcium signaling, thereby promoting the AWCON subtype. This study provides insight into the role of mechanical force in the stochastic lateralization of olfactory neurons.

developmental biology↗

bioRxiv2026

Defective BRCA1-mediated DNA end resection drives tandem duplication formation and FANCM synthetic lethality

AbstractBRCA1-linked cancer genomes contain abundant genome-wide [~]10 kb Group 1 tandem duplications (TDs) that are drivers of tumorigenesis. Group 1 TD formation is recapitulated at a chromosomal Tus/Ter site-specific replication fork barrier in DNA end resection-defective mouse embryonic stem (mES) cells lacking Brca1 exon 11. To explore relationships between DNA end resection and Group 1 TD formation, we analyzed Brca1 coiled coil (CC) domain mutants--separation-of-function alleles that are impaired for homologous recombination but competent for DNA end resection. Notably, Brca1 CC mutants retain the ability to suppress Group 1 TDs in the Tus/Ter system and in a mouse model of Brca1-linked tumorigenesis. These data show that Brca1 CC domain mutant cancers follow a path of tumorigenesis distinct from that of other pathogenic Brca1 alleles. FANCM is a TD co-suppressor, the loss of which is synthetic lethal/sick in combination with Brca1 exon 11 mutation. In contrast, Fancm deletion is well-tolerated by Brca1 CC mutant mES cells. Thus, Group 1 TD formation and Fancm synthetic lethality are linked phenotypes related to defective BRCA1-mediated DNA end resection.

molecular biology↗

bioRxiv2025

APOL1 G1 and G2 risk alleles modulate severity of diet-induced obesity in a transgenic mouse model

APOL1 G1 and G2 risk alleles are associated with an increased risk of chronic kidney disease. However, a causal relationship between these alleles and cardiometabolic traits has not been experimentally validated. To address this gap, we placed transgenic APOL1 G0, G1, and G2 FVB/NJ mice on a high-fat diet and analyzed them for weight gain as well as obesity-related cardiometabolic phenotypes. To test whether APOL1 risk alleles modulate the inflammatory basis of obesity, we also exposed bone marrow derived macrophages from these mice to pro-inflammatory, pro-hypertensive, and dyslipidemic conditions. APOL1 high-risk allele female mice gained fat mass more readily than their low-risk female counterparts, while APOL1 high-risk male mice gained fat mass less readily than low-risk males. A parallel sex difference was seen in expression of higher levels of Abca1, Hmox1, and Srebf1 in lipid-loaded female bone-marrow derived macrophages expressing G1 and G2 APOL1, along with minor differences in cardiac function. However, this finding occurred independently of hypertension and insulin resistance, and with only minor albuminuria. Thus, our results highlight the importance of sex as a biological variable in future APOL1 experiments.

genomics↗

bioRxiv2025

Multi-Dimensional Spatiotemporal Attention Neural Network for Next Generation Sequencing Basecalling

Next-generation sequencing (NGS) remains the most used sequencing technique in the field of genomics. Traditional basecall methods face significant challenges in decoding high density sequencing data due to inherent noise in biochemical reactions and limitations of instruments. Here, we present a multi-dimensional deep learning neural network based on spatiotemporal attention mechanism named AICall. The network skips computationally heavy but less effective steps of peak finding and brightness extraction/correction, and directly basecalls from the time sequence of multi-dimensional image stacks obtained in real time. By introducing attention mechanism, it effectively extracts spatial and time-related key information including spatial crosstalk, spectral crosstalk, phasing, base-quenching, and intensity decay, and significantly improves basecall accuracy. We demonstrate that AICall achieves an average error rate less than 0.01% and provides more reliable sequencing results for downstream analysis.

bioinformatics↗

bioRxiv2025

Integrating Genomics into Multimodal EHR Foundation Models

AO_SCPLOWBSTRACTC_SCPLOWThis paper introduces an innovative Electronic Health Record (EHR) foundation model that integrates Polygenic Risk Scores (PRS) as a foundational data modality, moving beyond traditional EHR-only approaches to build more holistic health profiles. Leveraging the extensive and diverse data from the All of Us (AoU) Research Program, this multimodal framework aims to learn complex relationships between clinical data and genetic predispositions. The methodology extends advancements in generative AI to the EHR foundation model space, enhancing predictive capabilities and interpretability. Evaluation on AoU data demonstrates the models predictive value for the onset of various conditions, particularly Type 2 Diabetes (T2D), and illustrates the interplay between PRS and EHR data. The work also explores transfer learning for custom classification tasks, showcasing the architectures versatility and efficiency. This approach is pivotal for unlocking new insights into disease prediction, proactive health management, risk stratification, and personalized treatment strategies, laying the groundwork for more personalized, equitable, and actionable real-world evidence generation in healthcare.

bioinformatics↗

bioRxiv2025

Localization of Fascin to Dendritic Protrusions Regulates Postsynaptic Plasticity

The Fascin family of actin-bundling proteins is best known for organizing actin filaments (F-actin) into tightly packed bundles that drive dynamic membrane protrusions such as filopodia. In neurons, fascin has been thought to primarily function in axons, as previous studies reported its absence from dendritic filopodia and spines. Here, we demonstrate that fascin is both present and functionally important in dendritic compartments. Using optimized immunocytochemistry and CRISPR-based endogenous tagging of fascin1 in cultured hippocampal neurons, we show that fascin is concentrated in developing dendritic filopodia and enriched in mature dendritic spines. Super- resolution imaging further reveals that fascin is organized into discrete nanoscale foci within spine heads, but not the spine neck. Finally, we show that CRISPR-mediated knockout of fascin1 in mature hippocampal neurons impairs synaptic potentiation, without affecting baseline excitatory synaptic transmission. Together, our findings uncover a previously overlooked aspect of actin organization in dendritic spines and establish fascin as a critical regulator of postsynaptic plasticity. Summary StatementThe actin bundling protein fascin localizes to dendritic filopodia and spines, where it regulates activity-dependent synaptic plasticity.

neuroscience↗

bioRxiv2025

AIChatBio: An Artificial Intelligence Chatbot Model for Biological Knowledge Retrieval and Biomacromolecule Design

Conversational agents for bioinformatics data analysis and interpretation remain largely inaccessible to the broader biological research community. This gap is especially pronounced in the current Generative AI era, which demands a paradigm shift in how researchers interact with computational tools. There is a pressing need to bridge well-established biological infrastructures and databases with the capabilities of Generative AI to democratize access to bioinformatics insights. In this study, we present an integrated framework that connects the robust bioinformatics resources of the National Center for Biotechnology Information (NCBI) with Generative AI through a novel Artificial Intelligence Chatbot Model for Biological Knowledge Retrieval and Biomacromolecule Design, AIChatBio. This operational model positions Generative AI as an intelligent information hub. User interactions with the chatbot are enriched by real-time data retrieval from web portal of the biological databases hosted at NCBI, which are then translated into structured inquiries toward the web applications of NCBI and bioinformatics analysis tools. These inquiries are directed toward bioinformatics analysis tools to perform tasks such as sequence alignment and primer design. Additionally, the outputs generated by these tools are interpreted by the chatbot, allowing users to gain meaningful insights without requiring deep technical expertise in bioinformatics. To demonstrate the feasibility of this approach, we developed a prototype implementation that integrates PCR primer design using Primer-BLAST [1], literature interpretation via PubMed for general topics, and the LitVar2 for SNPs associated topics [23]. This system was built using TypeScript and the ChatGPT API combining the bioinformatics web applications from NCBI, and its source code is publicly available via GitHub and the Chrome extension is available at Chrome Web Store. Our work highlights the potential of Generative AI to transform biological data analysis workflows, making them more intuitive, accessible, and scalable for researchers across disciplines.

bioinformatics↗

bioRxiv2025

HPInet: Interpretable prediction of Host-Pathogen protein-protein Interactions using a transformer-based neural network

Gram-negative bacteria utilize a series of secretion systems (T1SS-T10SS) to deliver secreted effector proteins (T1SE-T10SE) into host cells, leading to infections and diseases. Understanding the interactions between these effector proteins and host proteins is crucial for unraveling the pathogenic mechanisms of bacterial pathogens. Despite advancements in sequencing technologies that have significantly enhanced our knowledge of effector protein diversity and structure, the molecular mechanisms underlying their interactions with human host proteins remain unclear. Experimental approaches to validate these interactions are labor-intensive, time-consuming, and insufficient to comprehensively map the extensive effector-host protein interaction (PPI) network. Additionally, specialized computational tools for predicting these interactions remain scarce. To address this gap, we developed HPInet, a transformer-based deep learning model for predicting interactions between bacterial effector proteins and human host proteins. HPInet introduces convolutional layers and global response normalization (GRN) into the feed-forward network (FFN) of the transformer architecture, enhancing the extraction of local sequence features. Compared to existing state-of-the-art methods, HPInet significantly improves prediction performance, increasing accuracy from 0.502 to 0.891 and boosting sensitivity by 35.3%. Furthermore, by leveraging the models attention mechanism, HPInet identifies critical residues at interaction sites, demonstrating its capability to capture local structural features of PPI sites solely from sequence information. To facilitate its application in research, HPInet has been implemented as a freely accessible web server available at https://bis.zju.edu.cn/hpinet. This platform provides a powerful and interpretable tool for studying effector-host protein interactions, offering valuable insights into bacterial pathogenesis and potential therapeutic targets.

bioinformatics↗