IEEE Transactions on Computational Biology and Bioinformatics

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.

Drug-target interaction prediction and binding affinity prediction are two key tasks in drug discovery and drug repurposing. Although deep learning methods have made significant progress, existing models typically rely on global representations of drugs and proteins, making it difficult to adequately model fine-grained interactions between their local units. Fixed multimodal fusion strategies also struggle to dynamically adjust the contributions of different modalities for different drug-target combinations. To address these issues, we propose DQHTFI, a fine-grained interaction prediction framework for drug-target interaction classification and binding affinity regression. DQHTFI employs BRICS fragments and Pfam functional domains as the basic interaction units and jointly learns semantic and structural representations. We design a dynamic-query hypergraph Transformer framework in which hyperedges are constructed among the multimodal features of fragment-domain pairs. Dynamic queries are generated from the cross-conditioned features of fragment-domain pairs to adaptively adjust the contribution of each modality, thereby modeling higher-order interactions between local units. Our proposed model achieves competitive results on multiple benchmark datasets.

BackgroundPhylogenetic tree reconstruction from single cell data based on copy-number alterations (CNAs) is an important problem in cancer genomics. Methods have been developed to address this problem by computing pairwise distances between copy-number profiles and employing a tree reconstruction algorithm. Despite the tight interplay between distance estimation and tree reconstruction, these two steps are often treated as separate problems, with the choice of the reconstruction algorithm receiving little attention. Most methods rely on classical Neighbor Joining (NJ), an algorithm designed for unrooted phylogenies that does not account for the fixed diploid root inherent to copy-number evolution. ResultsWe identify the Deepest Least Common Ancestor NJ (DLCA-NJ), not previously applied in this context, as the appropriate algorithm for phylogenies from copy-number data. By leveraging the known diploid root, it consistently outperforms standard NJ on simulated benchmarks across all evaluated metrics, with the most pronounced improvement in root placement accuracy. Building on these findings, we introduce Sparse Rooted Neighbor Joining (SRNJ), a scalable adaptation of DLCA-NJ. SRNJ significantly reduces running time while trading off only a minor loss in accuracy. We provide theoretical and empirical evidence of robustness to mutation rate using both synthetic and real biological datasets. ConclusionsRooted NJ variants offer a principled way to exploit the known diploid root when reconstructing phylogenies from copy-number data, and SRNJ extends this advantage to datasets whose size places the full distance matrix out of reach. The gains are clearest where distances are reliable, as on simulated data, while on real data accuracy appears to be constrained by distance estimation rather than by the reconstruction algorithm, leaving room for improvement as callers advance.

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

Large-scale prediction and assessment of clinical patient responses (i.e., RECIST class) to drug combinations remains challenging due to scarce patient-derived data. The existing prediction methods mainly rely on cancer cell line models. However, substantial biological heterogeneity between cancer cell lines and cancer patients within same tissues, as well as the heterogeneity between one tissue and another, often limit the generalizability of these methods in clinical patients. To overcome these limitations, here we present CaMeRe, a Causally-inspired Meta-representation learning framework designed to predict patient-specific clinical Response to drug combinations. In situations where stable causal factors and domain-specific response-modulating factors are unobservable, explicit discrete domain labels are unavailable, and data is scarce, CaMeRe designed a domain-invariant causal representation learning (DICRL) model guided by the invariant information bottleneck theory and causal intervention invariance principle, and also built a meta-learning framework with bi-level domain generalization to optimize DICRL model for achieving multi-domain generalization within and across tissues. By integrating the causal representation learning and meta learning framework, CaMeRe not only exhibited robust multi-domain generalization performance across multiple clinical drug combination response datasets and PDXs drug combination response datasets and generalization scenarios, but also had better interpretability. We applied CaMeRe to predict drug-combination response scores for 3,423 patients across 542,080 drug combinations. The predicted scores were significantly associated with biomarkers of known drug combinations and enabled the prioritization of candidate drug combinations across 11 cancer types, with stronger support from literature and clinical trial evidences than random baselines. We believe that CaMeRe can be a useful tool for predicting large-scale clinical individual drug combination responses and it has broad clinical applications.

Protein-protein interactions shape cellular organization and disease processes, but their prediction remains difficult because known interaction networks are incomplete, noisy and highly uneven across proteins. Hyperbolic graph learning is well suited to the hierarchical structure of such networks, yet most methods emphasize learning hyperbolic embeddings and pay less attention to how these embeddings are propagated and decoded for type prediction. Here we introduce the Lorentz Tangent Diffusion Adapter (LTDA), a post-embedding module for PPI type prediction, that is, for predicting which of several interaction types a candidate protein pair undergoes. LTDA learns relation-specific protein representations on the Lorentz manifold, projects them to tangent space, applies lightweight graph diffusion, and decodes the diffused representations directly in tangent space. This design uses the tangent space as a prediction-oriented interface between hierarchical representation learning and interaction classification. On SHS27K and SHS148K under BFS and DFS splits, LTDA consistently improves over representative PPI methods and a reproduced HI-PPI baseline, with average gains of 1.41 percentage points in Micro-F1, 3.21 in AUPR, 1.31 in AUC and 1.25 in accuracy. Ablation, degree-bin and biological-evidence analyses indicate that the gains are not explained by simple smoothing or hub correction. LTDA therefore shows that post-embedding adaptation can make Lorentz representations more effective for PPI type prediction.

Drug repurposing and target discovery offer critical strategies for advancing therapeutic development by uncovering the potential biological pathways and novel associations among drugs, genes, and diseases. However, experimental discovery remains expensive and time-consuming, which limits the scalability of large-scale studies. In addition, existing computational approaches often struggle to effectively integrate heterogeneous biomedical data, capture the complex higher-order topological signatures of biological interactomes, and generalize to unseen entities. Here, we present HANAMI (Heterogeneous grAph coNtrastive leArning for drug-gene-disease Motif predIction), a multi-view deep graph learning framework designed to model complex interactions among drugs, genes, and diseases. HANAMI integrates diverse heterogeneous biomedical knowledge, including chemical structures, genomic sequences, and clinical phenotypes, and leverages relation-aware topology encoding, structure-aware aggregation, and contrastive learning to enable accurate motif prediction with biological context from the network. Systematic evaluation on benchmark datasets shows that HANAMI achieves up to 6% improvements over existing state-of-the-art methods in predicting drug-gene-disease motifs. The framework further demonstrates strong inductive generalization, maintaining an [~]18% performance advantage in zero-shot settings involving previously unseen entities. Beyond predictive performance, HANAMI effectively prioritizes drug-disease relationships investigated in Phase II or III trials while identifying candidate genes that suggest plausible mechanistic links. Together, HANAMI provides a computational framework for interpreting complex biomedical interactions, offering a scalable foundation to accelerate drug repurposing and therapeutic innovation.

1Drug--drug interactions between biotech and small-molecule drugs play a critical role in medication safety and therapeutic efficacy. However, most existing computational DDI prediction methods focus primarily on interactions between small-molecule drugs, leaving biotech-small-molecule interactions comparatively underexplored. In this study, we propose B-SMART-Former, an explainable deep learning framework for predicting interaction types between biotech and small-molecule drugs. The proposed framework integrates ChemBERTa embeddings and Morgan molecular fingerprints for small molecules with ProtBERT embeddings for biotech drugs, eliminating the need for similarity-based features while leveraging complementary molecular representations. These multimodal features are processed by a hybrid architecture that combines Transformer-based self-attention, residual convolutional learning, and a multi-layer perceptron classifier to capture both global contextual dependencies and local discriminative patterns. The model is formulated as a multi-class classification task and evaluated using stratified 10-fold cross-validation. To improve model transparency, Integrated Gradients is employed as a post-hoc explainability method to identify the molecular features that contribute most strongly to each prediction. Experimental results demonstrate that B-SMART-Former achieves a micro-averaged AUROC of 0.9978 and an AUPR of 0.9682 while relying solely on intrinsic molecular representations, remaining competitive with similarity-based approaches. The proposed framework offers an effective and explainable solution for biotech-small-molecule DDI prediction and provides a practical foundation for future computational drug interaction studies.

Genotype-by-environment interactions can complicate plant breeding because genetic effects may vary across weather regimes, locations, and years. Although genomic prediction models can use dense marker data to predict complex crop traits, they often provide limited information about how marker relevance changes across environmental conditions. This study introduces a graph-based framework for estimating environment-associated marker rankings from multi-environment crop trial data. We introduce an attention-based graph neural network that jointly represents SNP markers and weather-based environment vectors as interacting nodes and is designed to learn environment-associated marker rankings from supervised yield data. The model is trained using yield prediction as the supervised objective, and the learned marker-level attention signals are aggregated to estimate marker rankings overall and within groups of similar environments. We evaluate the framework on two genotype-by-environment datasets. In both datasets, we examine whether learned marker rankings are stable across independent training runs and determine the overlap between high-ranked markers and external GWAS references. In a maize dataset with data from 212 different site-years, we rank markers separately within each weather-defined environment cluster and define a global marker set as the markers shared across all clusters. This provides a structured way to examine how marker rankings vary across weather-defined groups and which markers remain important across environments. We further encode each weather variable with a separate LSTM combined by attention, which makes the contribution of individual weather variables readable, and compare the resulting variable ranking against an independent SHAP-based attribution. Genes located near the top-ranked markers are also examined for functional enrichment. Overall, the results suggest that graph-based, environment-conditioned marker attribution may complement genomic prediction as an exploratory tool for studying context-dependent genomic signals in multi-environment trials.

Accurate molecular property prediction requires effective molecular representations that can describe a molecule from multiple complementary perspectives. Existing deep learning approaches typically use SMILES strings, two-dimensional molecular graphs, or three-dimensional conformations as inputs. These representations capture different aspects of molecular information: SMILES encodes a sequential description, molecular graphs describe atom-bond connectivity, and 3D conformations provide spatial information from atomic coordinates. However, these modalities are often learned separately or merged with low effective fusion operations, which may not sufficiently capture interactions among sequential, topological, and geometric features. In this work, we propose Tri-Modality Cross-Attention (TMCA), a multimodal framework that integrates 1D SMILES, 2D molecular graphs, and 3D molecular conformations for molecular property prediction. TMCA uses pretrained encoders for the 1D and 3D branches, with a SMILES-based Transformer for the 1D branch and a recent conformation-aware pretrained model for the 3D branch. At the same time, a trainable 2D graph encoder is designed to support modality fusion. All the parameters from the pretrained models are frozen to make overall training more efficient. We evaluate TMCA on four datasets (i.e., BBBP, BACE, ClinTox, and HIV) from the MoleculeNet database for classification tasks and compare its performance with state-of-the-art competing approaches. Experimental results show that TMCA achieves the best overall performance, demonstrating the value of multi-modality integration for drug property prediction and the effectiveness of our cross-attention-based fusion strategy. The source code of TMCA is available at: https://github.com/Ay-Zhao/TMCA

Deep learning methods, such as AlphaFold and RosettaFold, achieve high accuracy in protein structure prediction. However, predicting the structure of large protein complexes remains challenging due to their large size and intricate multi-chain interactions. Docking-based methods can handle large proteins, but are limited by the huge combinatorial binding space of multichains. Assembly-based approaches offer an alternative, but their accuracy critically relies on the precision of predicted subcomponents. Addressing the challenges, we propose HDOCK-Multimer (HDM), a structure prediction framework of large protein complexes by integrating ab initio docking and combinatorial assembly. HDM can efficiently reduce reliance on subcom-ponent accuracy through docking process, while leveraging the pairwise interactions of subcom-ponents through assembly strategy. HDM is extensively validated on three benchmarks of 35 large heteromeric complexes, 172 large protein complexes, and 7 CASP15 targets, and compared with state-of-the-art methods including MoLPC, CombFold, AlphaFold-Multimer (AFM), and AlphaFold3 (AF3). It is shown that HDOCK-Multimer substantially outperforms the other methods. In addition, HDM also shows ability to predict the stoichiometry and model the complex without stoichiometry input. It is anticipated that HDM will serve as a powerful tool for studying large protein complexes or molecular machines. The HDM package is freely available at https://github.com/huang-laboratory/HDOCK-Multimer/.

In the recent growing interest of AI research toward biology, genetic association studies of AI- derived phenotypes from high-content modalities such as images emerges as a powerful means for biological discovery. However, such AI-phenotyping methods still lacks a good optimization target and an efficient evaluation framework. The number of discovered loci was the major criterion for evaluating the quality of AI-derived phenotypes. However, the time and computational resources required for running the GWAS and subsequent loci-clumping are substantial, limiting the rapid development and iteration of deep learning representation algorithms. Here we present a 1000x faster and lightweight framework, sum-h2, than traditional GWAS framework for evaluating genetic discovery through total heritability. We revisit tr(P^(-1) G), previously proposed in the context of evolutionary studies, as a measure of multi-phenotype heritability. We showed that sum-h2, the sum of heritability over phenotypic Principal Components (PCs), is equivalent to the linear transformation-invariant tr(P^(-1) G), through both theoretical proof and simulation studies. Moreover, sum-h2 can be estimated rapidly with minimal information loss over a relatedness-enriched sample, while preserving the relative ranking of endophenotypes by GWAS loci counts. Based on selected UKB data and sum-h2, we set up a Genetic Discovery Arena, enabling rapid and fair comparisons for the development and optimization of deep learning-derived phenotyping methods.

Accurately predicting the effects of mutations on protein-RNA binding is crucial for elucidating disease mechanisms. Yet, exhaustively exploring the space of all possible variants is prohibitively expensive, motivating computational methods that can quantify mutation-induced changes in binding affinity (aka {Delta}{Delta}G) accurately and efficiently. We present iSCALE, an interpretable and generalizable deep learning method that adopts an implicit Spatial Coupling-Aware Ligand Encoding strategy to predict mutation-induced binding affinity changes. By injecting this implicit multiscale encoding scheme into a bidirectional state space modeling architecture, iSCALE learns a generalizable multiscale coupling pattern that achieves superior performances on not only the protein-RNA binding {Delta}{Delta}G, but also the protein stability {Delta}{Delta}G and protein-protein binding {Delta}{Delta}G predictions. Detailed analyses demonstrate that the model attention scores align well with structural characteristics. In addition, iSCALE shows good discriminative ability when predicting close samples such as complexes of same mutation but with different ligands or the same complex but with different mutation sites. In summary, iSCALE serves as an effective in silico tool for large-scale protein-RNA binding {Delta}{Delta}G prediction, which pushes the border of understanding in mutation-induced pathological outcomes.

Biobank-scale genomic analyses are increasingly constrained by computational costs, as hundreds of thousands to millions of samples and variants must be analyzed together. The genotype representation graph (GRG) compactly encodes population genetic variation to accelerate computation, but the current approach does not exploit modern accelerator architectures. This work introduces Mikado, a new methodology for expressing GRG-based computation using sparse linear algebra primitives. Under a reverse topological ordering of the graph nodes, the GRG adjacency matrix is strictly block-lower-triangular, and the genotype matrix-vector product becomes a sparse triangular solve that can be further decomposed into a pipelined sequence of blocked sparse matrix-vector multiplies. By decoupling computation from graph representation, our approach exposes fine-grained parallelism and enables hardware-optimized sparse primitives on GPUs. Mikado achieves an order-of-magnitude speedup and cost savings for PCA and BOLT-LMM compared with the original GRG traversal approach, including on All of Us cohorts. It provides a scalable, hardware-portable, researcher-friendly tool for population genetics at biobank scale.

Current studies in human genomics typically rely on the standard genome reference GRCh38 which is known to be biased toward populations of European ancestry and therefore has limitations when applied to other populations. Although various graph-based pangenome references were constructed for several populations to deal with this bias, their usage in practice is currently still limited compared to linear genome references. Here we present a framework for constructing a population-specific genome reference using GRCh38 as backbone with alternate-contig awareness to enhance genomic data analysis in the target population. We demonstrated the advantages of our framework using both public and in-house Vietnamese whole-genome sequencing (WGS) datasets. Genomic variants derived from high-coverage WGS data of the 1000 Vietnamese Genomes Project (VN1K) were imported into our framework to build a Vietnamese-specific Genome Reference (VGR). VGR was then compared to GRCh38 in read alignment and variant calling using high-coverage WGS data of 99 Vietnamese individuals (KHV) from the 1000 Genomes Project (1kGP). Using Omni array genotyping data from 99 KHV samples as an independent benchmark, we found that VGR improved variant-calling precision and reduced false-positive calls compared to GRCh38. Our framework could be easily used for other populations as long as they have a variant database similar to VN1K. Our code is publicly available at github.com/VinGenome/VGR

MotivationIn this paper, we demonstrate that it is feasible to train a deep generative model for dimensionality reduction with millions of features using few samples, which makes this type of generative model a more versatile alternative to standard methods for dimensionality reduction. Specifically, we hypothesize that for a decoder-only model, the number of training samples required is almost independent of the feature dimensionality in most network architectures. ResultsThrough an extensive set of experiments on synthetic non-linear data, we validate this hypothesis. We also train the model on a downsampled version of the 1000 Genomes Project (1KGP) dataset to further assess its behavior under controlled reductions in sample size. Furthermore, we train a deep generative decoder (DGD) on a curated dataset from the International Cancer Genome Consortium (ICGC), which contains 4.4 million features. It is trained on approximately 4,000 samples and tested on 1,000 samples. The resulting latent representation exhibits clear clustering, and when methods are reduced to the same number of dimensions, it outperforms PCA and VAE for tumor type classification. Additionally, the DGD is computationally efficient and can be trained on a 16GB GPU. Availability and implementationCode is available at https://github.com/cpancott/ReceptiveDGD. Contactcorrado.pancotti@helmholtz-munich.de; akrogh@di.ku.dk Supplementary informationSupplementary data are available with this preprint.

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.

Conventional short read DNA alignment algorithms, such as widely used BWA-MEM2, devote a substantial fraction of their execution time to seed generation. We present RosaSeed, a short read alignment algorithm, based on a configurable seeding framework, that significantly accelerates alignment while maintaining accuracy comparable to that of BWA-MEM2. RosaSeed replaces the BWA-MEM2 seeding kernel and directly supplies candidate seeds to the existing BWA-MEM2 chaining and alignment extension pipeline, producing standard SAM output suitable for downstream variant analysis. RosaSeed was evaluated on both simulated and real short read datasets, and its seeding time, fastq-to-SAM time, alignment accuracy, and memory footprint were compared against BWA-MEM2, Minimap2, Bowtie2, and the Enumerated Radix Tree (ERT) algorithm. We also compared against ERT2, a proposed simplified variant of ERT that sacrifices a small amount of alignment accuracy to significantly increase the alignment throughput. With single-threaded software execution, our recommended RosaSeed configuration achieves 13.06x and 3.85x speedups in seed processing and fastq-to-SAM time, respectively, over BWA-MEM2. Corresponding seed processing (and fastq-to-SAM) speedups are 3.96x (2.48x) over ERT, 1.2x (1.44x) over ERT2, 4.82x (3.31x) over Minimap2, and 33.24x (10.83x) over Bowtie2. RosaSeed requires only 49.61 GB of peak memory, approximately 25% less than ERT and ERT2, while maintaining alignment accuracy comparable to that of BWA-MEM2. Following the recent release of minibwa, a prefetch-optimized FM index aligner, we additionally benchmarked RosaSeed against it on an upgraded 24-core, 48-thread AMD Zen 3 workstation. Our speed-optimized miniRosaSeed configuration achieves a 2.12x fastq-to-SAM speedup over minibwa for single-threaded execution while producing 2.58% higher standard accuracy. Against Strobealign, miniRosaSeed reported a 2.18x fastq-to-SAM speedup and 3.1% higher standard accuracy. For multi-threaded execution, miniRosaSeed was both faster and more accurate than minibwa for up to 26 software threads on the Zen 3 workstation. RosaSeed provides multiple runtime-memory configurations, spanning memory-efficient and high-performance strategies, enabling users to select operating points appropriate for diverse computational environments while maintaining alignment quality comparable to that of BWA-MEM2.

Clinical interpretation of genetic variation depends on the diploid genotype, including zygosity, allele dosage and whether multiple variants occur in cis on the same homologue or in trans on different homologues. Most genomic language models process haploid sequences or combine independently encoded haplotypes downstream, so they do not directly represent the paired genotype in a single sequence. We introduce a reference-aligned diploid encoding for single-nucleotide variants (SNVs) and short insertions and deletions (indels), together with unphased and phase-retaining tokenizers that accept phased genotypes and convert them to single-sequence diploid representation. Using Nucleotide Transformer v3 backbones, we continue training 8-million- and 100-million-parameter models and evaluate an auxiliary Contrastive Phase Loss (CPL) designed to retain the phasing information of the variants in contextual representations. We evaluate on a novel compound-heterozygous benchmark containing 9,460 examples. Models whose inputs did not distinguish relative phase remained near chance, whereas our diploidic models improved discrimination with AUROC 0.649, compared to 0.506 for the vocabulary-adapted control. These findings establish a method for making diploid genotype information accessible to genomic language models, rather than a universal improvement in variant prediction; validation in naturally observed, accurately phased clinical cohorts remains necessary.

Accurate, reliable, large-scale disease predictions are essential to ensure rice production. Existing disease prediction models often face a trade-off between interpretability and predictive capability, necessitating the integration of mechanistic knowledge and data-driven learning within a modelling framework. Accordingly, we propose a physics-informed hybrid gated recurrent unit (PI-HGRU) model for spatiotemporal dynamic prediction of rice sheath blight disease, caused by a fungus. Our model embeds differential equations describing disease transmission dynamics into a hybrid gated recurrent unit (HGRU) framework as mechanistic constraints, thereby enabling collaborative modelling between epidemiological processes and data-driven learning. We conducted model training and evaluation using a long-term, multi-source dataset spanning 17 years (2000-2016) and covering 16 major rice-producing provinces in southern China. These rich data include spatiotemporally aligned field disease observations, remote sensing data, meteorological data, and soil property data. In addition, to address the challenges of irregular sampling intervals and inconsistent sequence lengths in disease survey data, we adopted a sliding time-window-based prediction framework. We further conducted a time-window sensitivity analysis to determine appropriate configurations of the input time window and lag time, enabling the model to represent the cumulative and delayed effects of environmental factors. Our PI-HGRU framework substantially outperforms the purely data-driven HGRU baseline model, improving the squared Pearson correlation coefficient (r2) by 22.8% while reducing the root mean square error (RMSE) and mean absolute error (MAE) by 10.2% and 18.0%, respectively. Furthermore, analysis of the models intermediate variables showed that the transmission rate {beta}(t) exhibited interpretable relationships with environmental conditions within the input time window, providing a process-related link between environmental drivers and modeled disease transmission dynamics. Overall, our work demonstrates that integrating epidemiological mechanisms into deep learning models in a physics-informed manner can improve predictive accuracy and stability while enhancing model interpretability, highlighting its potential for large-scale disease forecasting and precision disease management.