LATEST IN SCIENCE

Abstract: Nanopore sequencers produce reads within minutes, but metagenomic analysis remains challenging where internet access is limited, data are sensitive or computing resources are modest. EMtrace is an offline pipeline that runs on a laptop during sequencing: it classifies MinKNOW reads, estimates taxon abundance by expectation-maximization (EM) and assembles reference-based consensus genomes. We benchmarked it against five virome and four 16S rRNA tools on simulated, mock-community and clinical data. On field isolates and 16 SARS-CoV-2 clinical samples, EMtrace had the highest species-level F1 score (1.00 and 0.97) and one false-positive species, against 3-7 for the comparators. It recovered divergent viruses as completely as ViMOP, and low-titre field libraries more completely, at [≥]99.7% identity. In replayed runs, it identified the target virus in the first 4,000-read batch (5-13 min). With a shared 16S reference, EMtrace reported two to three times fewer false-positive species than Emu.

Abstract: Accurate identification of HV-PPIs is essential for understanding viral infection mechanisms and for developing antiviral therapeutic discovery. However, the experimental identification process is often time-consuming and expensive, while many computational approaches depend on incomplete protein sequences, structural and interaction information. Here, we present HV-MultiScaleGNN, a topology-driven multi-scale GNN framework for HV-PPIs prediction that operates exclusively on the topology of a heterogeneous interaction network. The framework formulates HV-PPIs prediction as a heterogeneous link prediction problem by integrating the human interactome with experimentally validated HV-PPIs. HV-MultiScaleGNN systematically investigates three complementary graph representation strategies, viz. Local view, Global view and an integrated Local+Global view. The Local view employs one-hop relation-aware graph attention to model fine-grained neighbourhood interactions. The Global view captures higher-order structural dependencies through propagation-based neighbourhood aggregation. The integrated representation combines both views to jointly encode local connectivity and global topological context within a unified embedding space, enabling relation-specific link prediction without relying on sequence- or structure-derived features. Experiments on large-scale HV-PPIs datasets evaluate all three strategies using discrimination, calibration, and ranking metrics. The Local view achieved the best discriminative performance (AUROC = 0.9444, AUPRC = 0.9422), whereas the Global view exhibited the lowest ECE (0.0379). The integrated Local+Global view provided consistently strong predictive performance (AUROC = 0.9378, AUPRC = 0.9356). Comparative evaluation against representative graph representation learning methods showed competitive predictive performance. Furthermore, interfacial interactions were identified using protein-protein docking and atomistic-level understanding was explored. Moreover, GO and KEGG enrichment provided additional functional insights into the selected host proteins. The proposed framework enables systematic evaluation of complementary graph representation strategies for heterogeneous biological link prediction and helps experimental researchers to delineate novel HV-PPIs.

Abstract: Dissimilarity measures play a critical role in phylogenetics. They help quantify the accuracy of inferences when the ground truth is known, allow comparisons of inferences obtained using different methods, and provide a tool for summarizing sets of phylogenies. A wide array of dissimilarity measures has been introduced for phylogenetic trees over the past half-century, and they vary in their mathematical properties, empirical behavior, and computational complexity. When the evolutionary history involves reticulate processes, such as hybridization and horizontal gene transfer, it is represented by phylogenetic networks. Although the study of phylogenetic network dissimilarity is less established than its tree counterpart, the past two decades have seen an increasing number of measures developed for phylogenetic networks. However, while some have established mathematical properties---for example, metricity on certain classes of phylogenetic networks---none properly separates reticulation disagreement from tree-like disagreement. This is the rationale behind the introduction of the hybrid ancestry error (HAE) measure by Dai and Molloy. The HAE measure, however, is too simplistic to distinguish different reticulation scenarios. In this work, we propose a novel family of network dissimilarity measures that explicitly account for reticulation differences while being more informative than HAE. We prove that our measures are pseudometrics on rooted binary phylogenetic networks on a fixed taxon set, with two instantiations yielding metrics on binary level-1 networks. Empirically, the measures remain sensitive to reticulation errors as network size grows and respond to changes in hybrid identity and parental placement. Calibration experiments further characterize the measures' behavior under network edits of various magnitudes. Data Availability: The dissimilarities are implemented in the PhyNetPy library, available at https://github.com/NakhlehLab/PhyNetPy. All the data simulated for the study is available at https://doi.org/10.5281/zenodo.23933421. Funding: This work was supported in part by the NSF grants DMS/NIGMS-2153704 and DBI-2420499. Declaration of AI-assisted technologies: During the preparation of this work, the authors used OpenAI's Codex (GPT-6 Astra) and Anthropic's Claude (Fable 5) to assist with editing manuscript text, developing proofs and analysis code, and preparing figures. The authors reviewed and edited the resulting material, verified the analyses, and take full responsibility for the content of this article.

Abstract: Allosteric regulation offers multifunctional induction in therapeutic interventions, yet conformational plasticity hinders computational identification of allosteric sites. Machine learning models rely on static holo (ligand-bound) structures, generalizing poorly to apo or substrate-bound states where binding pockets are transient or cryptic. We present AsBench2, a cross-state site-recognition benchmark, and AllositeUltra, a deep learning framework integrating multimodal protein representations (combining structure-aware language models with physical dynamics) and residue-guided candidate supplementation, effectively decoupling site detection from static pocket geometry. AllositeUltra demonstrates superior generalization, achieving Top-3 site recovery rates of 65.2% and 70.6% on apo and substrate-bound structures, respectively, representing relative improvements of 36.4% and approximately 50% over the best-performing baselines. Prospective application to MAGI2 identifies a previously unrecognized interdomain allosteric pocket, validated at atomic resolution through small-molecule screening and X-ray crystallography. Proteome-wide application yields a high-confidence atlas of targetable allosteric sites, bridging the gap between structural representation learning and prospective AI-driven drug discovery.

Abstract: Accurate and efficient EC number prediction is essential for elucidating the catalytic roles of enzymes. Although similarity-based and deep learning-based approaches have advanced this task, they often fail to distinguish active enzymes from closely related decoys inactivated by local functional-site mutations, leading to high false-positive rates (FPRs). We propose DecoyOpt, a two-stage pipeline for decoy-aware enzyme discrimination and EC number prediction. Specifically, we generated putatively inactive functional-site decoys using ESM3 and evaluated 16 protein language models across three splits. An MLP trained with classification and ranking losses distinguishes enzymes from unannotated proteins and decoys before a contrastive learning model assigns EC numbers to predicted enzymes. On the most stringent GraphPart 40% split, DecoyOpt achieves an MCC of 0.899 and a D0 decoy FPR of 6.17%, outperforming retrained baselines. Across all splits, it outperforms these baselines in EC number prediction on the full test set while maintaining performance close to CLEAN on the enzyme subset. It also correctly annotates all four wild-type enzymes and rejects all eight inactive or nearly inactive variants. We believe that DecoyOpt has potential for annotating uncharacterized proteins and their variants, with applications in genomics, synthetic biology, and biocatalysis.

Abstract: Structural similarity searches reveal protein relationships that sequence alignments miss, but treating DALI results as isolated match lists misses deeper connections. We present DaliScope, an interactive Python framework that treats complete DALI target sets as structured populations for systematic analysis of protein structural neighbourhoods. DaliScope combines structural community detection with sequence profiling, Pfam annotation, residue mapping, and 3D visualization, enabling target sets to be subdivided and examined at the residue level. We demonstrate two applications. In a multidomain protein with sparse domain annotation, structural analysis uncovered previously unannotated domains. A structurally anchored HxD...H signature occurred in 697 of 1,093 Pfam-annotated CUPIN domains and in 15 domains lacking Pfam annotation; relaxing the central residue constraint recovered 96% of Pfam-annotated CUPIN domains among the DALI targets. In a second application, a DALI search with transitive closure returned >12,000 hits. Applying an 80% coverage criterion to a conserved structural core reduced this set to 6,708 domain instances, resolved into 23 structural communities. (80% was chosen instead of the default 50% because the GOLD domain has a compact core with short loops.) Many communities were strongly enriched for individual Pfam families while providing structural context for 1,097 targets lacking Pfam assignments. Three-dimensional superimposition revealed structural continuity among related {beta}-sandwich architectures through progressive changes in sheet geometry and auxiliary elements. Together, these examples show how a single DALI query provides a structural anchor for organizing large neighbourhoods, extending interpretation beyond sequence-based annotations, and deriving structurally anchored sequence signatures from defined populations.

Abstract: Predicting whether a chemical disturbs liver and mitochondrial function matters for drug safety and environmental risk assessment, but the data available for training such models are small and mostly curated by hand. We describe an automated pipeline that uses a large language model (LLM) to collect hepatotoxicity findings from the biomedical literature. The pipeline runs a PubMed/PMC query, downloads the matching full texts or abstracts, asks the LLM to record each reported effect in a fixed schema (tested chemical, biomarker, direction of change, and related fields) with controlled vocabularies, and writes de-duplicated records to a database that is updated as new papers appear. The current snapshot holds 189,212 findings on 17,002 chemicals from 53,575 articles (1995-2026), covering 235 biomarkers reported for at least 20 chemicals. In a manual check of 217 findings from 35 articles, 208 findings (from 33 articles) were fully correct; the other 9, all from 2 articles, were incorrect. As a use case, we train structure-based QSAR models, both graph neural networks and classic baselines, for each biomarker; at this data scale the classic models remain competitive. The pipeline and the knowledge base will be made publicly available.

Abstract: Clustering of single-cell RNA sequencing (scRNA-seq) data is widely used to identify cell populations. However, overclustering can partition a single cell population into multiple clusters, potentially leading to spurious identification of cell subpopulations. Assessing the statistical significance of clusters is therefore needed to support their biological interpretation. We propose ProClust, a statistical test for clustering significance in scRNA-seq data. ProClust defines a single cluster as a log-concave distribution and adopts a projection-based approach to quantify clustering uncertainty through p-values. These projections can also provide interpretable visual evidence for individual clusterings. ProClust merges clusters when there is insufficient evidence to reject the single-cluster null hypothesis, refining an initial overclustering without prespecifying the final number of clusters. Across parametric and PBMC-based semi-synthetic simulations, ProClust avoided false subdivisions under single-population null settings while consistently recovering the underlying population structure across different numbers of true populations. In real scRNA-seq datasets from the Human Lung Cell Atlas, Tabula Muris, and Hydra, ProClust produced clustering solutions that showed overall agreement with reference annotations while retaining supported substructure and merging groups with insufficient evidence for separation. Together, these results demonstrate that ProClust provides a statistically principled and interpretable framework for refining an initial overclustering and distinguishing supported cellular heterogeneity. The ProClust R package is available at https://github.com/Anonym027/ProClust.

Abstract: Looking up k-mers in an index is a core operation in genomics, underlying read mapping, metagenomic classification, and many other tasks. Queries are often streaming: the k-mers come from a longer sequence such as a sequencing read, and consecutive k-mers overlap. The fastest index for such queries is currently the hashing-based SSHash; indexes based on the spectral Burrows-Wheeler transform (SBWT) are competitive in space but slower. We show how to make SBWT indexes both faster and smaller. First, we process queries in small batches and prefetch the memory each will need next, hiding the latency of the dependent random accesses that dominate a lookup; together with a new rank structure that reads one cache line per query, this makes k-mer lookups about seven times faster than unbatched SBWT search. Second, we extend batching to streaming search and show that one failed lookup can rule out many k-mers of a read at once, which removes the need for the longest common suffix array. Third, we build an SBWT that stores only one orientation of each k-mer, thereby halving its size, with a modified search that exploits how rarely the orientation changes along a read. Finally, we adapt a compact subset rank structure to batched search. On the collections of the SSHash benchmark, and on two collections up to 15 times larger, our fastest index answers streaming queries for 31-mers faster than SSHash on every read set, by up to 4.9 times, using 12 bits per canonical k-mer against SSHash's 7.4-11.1. We also show that a 6.1-bit index is still faster than SSHash on most read sets, and that our smallest index needs about 4 bits per canonical k-mer. Unlike that of SSHash, the space of our indexes barely varies across collections. Our code is available at https://github.com/anadis504/sbwt-squeeze-rs.

Abstract: Molecular profiles and histopathology provide complementary information about cancer prognosis, but these data are often incomplete for individual patients. Here we present OmniT2I, a table-to-image framework for survival prediction and cross-modal inference from incomplete multimodal cancer data. Each molecular profile was converted into a three-channel image representation with correlation-based feature placement. OmniT2I combines available molecular and histopathology embeddings through availability-aware fusion and predicts missing modalities from those observed. Across 10,373 patients with 32 cancer types, the C-index was 0.631 within cancer types and 0.783 with all cancers pooled (0.660 and 0.787 in 5,605 patients with all five modalities). A single model handled all 31 modality combinations and, compared with nine survival methods trained separately for each, achieved higher C-indices in 240 of 279 pooled and 226 of 279 within-cancer comparisons. Adding predictions of withheld modalities to observed inputs improved pooled C-indices in all 180 pairings of observed and withheld modality sets and within-cancer C-indices in 164. Adding measured histopathology to RNA produced the largest within-cancer C-index gain among the four additional modalities evaluated, whereas adding copy number variation reduced it. In seven external cohorts, frozen RNA-only C-indices differed from test performance in the corresponding cancer type in TCGA by a median of -0.003 (range, -0.092 to +0.079). OmniT2I enables prognostic modeling with incomplete multimodal data.

Abstract: The emergence of drug resistance continues to limit the long term effectiveness of cancer therapies [1], as tumour cells adapt to therapeutic selective pressure through diverse genetic and non-genetic routes [2,3]. Despite advances in more potent and next generation inhibitors and mechanism based combination therapies [4,5], durable disease control remains difficult to achieve. Here we tested an alternative strategy: combining multiple inhibitors directed against the same oncogenic driver, each at ultra low doses with minimal individual activity, to delay resistance emergence. In epidermal growth factor receptor (EGFR) driven cancer models, such ultra low dose same target combinations (ULDCs) distributed individually weak perturbations across multiple inhibitors directed against the same target and achieved substantial growth inhibition in vitro and antitumour activity in vivo despite weak effects of the corresponding single agents. Under similar levels of growth inhibition (~50%), single-agent EGFR inhibitors generated resistant outgrowth, whereas a five drug ULDC delayed resistance emergence. Increasing the number of inhibitors further delayed resistance emergence, with no detectable resistant outgrowth observed for combinations containing 8 to 14 inhibitors within the experimental timeframe. The five drug ULDC also retained measurable growth inhibitory activity against gefitinib resistant PC-9/GR cells, and resistant cells arising under five drug ULDC treatment exhibited markedly greater reversibility following drug withdrawal than those arising under single agent treatment. Although pharmacological additivity or synergy may contribute to the collective growth inhibitory effects of ULDCs, the divergent long term resistance outcomes observed under similar levels of growth inhibition indicate that resistance behaviour is not determined by inhibitory strength alone. Instead, our findings support a model in which increasing same target inhibitor diversity progressively constrains the adaptive trajectories available during resistance evolution, while therapeutic perturbation architecture may also influence the stability of resistant states that emerge. More broadly, these findings support the concept of same target collective suppression, in which multiple individually weak perturbations directed against the same target can collectively produce substantial antitumour activity and constrain resistance evolution across multiple stages. We propose that increasing same target inhibitor diversity may restrict accessible adaptive trajectories, while therapeutic perturbation architecture may also influence the stability of resistant states.