Patterns

Scientific imaging is undergoing a fundamental transformation. Advances in multiplex immunofluorescence, computational pathology, and electron microscopy now enable observation of biological and physical systems at unprecedented spatial resolution. Rather than simply classifying known objects, these data offer the opportunity to discover previously unknown organizational principles. However, scientific image datasets are typically highly curated, information-rich, and limited in size, making reliable representation learning particularly challenging. We show that standard continuous self-supervised adaptation can become unstable in this regime, allowing teacher representations to drift and degrading downstream transfer. We introduce Kodiak, a discrete representation learning architecture that discovers latent scientific concepts through Sinkhorn-balanced codebook assignments. Instead of learning unconstrained continuous embeddings, Kodiak constructs a reusable vocabulary of discrete concepts that provides stable supervision for both masked-patch prediction and cross-view representation learning within a single-stage framework. Across five scientific imaging domains, including multiplex immunofluorescence for pancreatic cancer pathology and scanning electron microscopy, Kodiak consistently improves representation quality under both continued foundation-model adaptation and training from scratch. Beyond improved classification and transfer,the learned concepts are interpretable, reusable, and produce spatially coherent concept maps that support applications such as cell phenotyping, segmentation-free patch phenotyping, spatial niche discovery, tissue organization, and materials characterization. We view Kodiak as the representation-learning engine of our foundation model for spatial biology. More broadly, our results suggest that discrete representation learning provides a principled foundation for discovering reusable latent scientific concepts that can accelerate scientific interpretation across imaging domains.

Unveiling directional causality in complex, high-dimensional dynamical systems remains a fundamental challenge. Conventional nonparametric approaches, such as Convergent Cross Mapping, require computationally intensive surrogate tests and frequently yield false-positive detections under strong coupling due to generalized synchronization. Here, we present F-CODEC (Forest-weighted Conditional Dependence Coefficient), a forward-looking causal discovery framework integrating generalized random forests with adaptive rank conditioning. Nonparametrically conditioning on the target variable's intrinsic delay coordinates systematically neutralizes shared manifold representations, suppressing false positives without surrogate data. Furthermore, an analytical time-series threshold allows instant, rigorous hypothesis testing directly from raw observations. We demonstrate F-CODEC's resilience against generalized synchronization and high-dimensional noise across canonical chaotic benchmarks and multi-species network motifs. Applying F-CODEC to historical lynx-hare cycles resolves delayed trophic forcing, while in-situ IoT honeybee monitoring captures real-time microclimate-acoustic causal surges preceding colony-level behavioral events. F-CODEC establishes a practical, scalable, and surrogate-free tool for causal discovery in nonlinear dynamical systems.

In many natural dynamical systems, tipping points emerge where slowly changing external factors induce shifts from one state to another, known as critical transitions or bifurcations. In epidemic systems, detecting these transitions is important for outbreak preparedness. Various statistical indicators like lag-1 autocorrelation and variance have been employed as early warning signals (EWS) of critical transitions by showing how the dynamics of the system exhibit critical slowing down (CSD) near the transition. More recently, deep learning-based algorithms have been developed that attempt to more accurately detect and classify bifurcations; however, they are limited by the systems simulated in the training dataset, the length of the time series, and have not been rigorously evaluated on real-world epidemic data spanning multiple geographic regions. In this paper, we evaluate existing deep learning-based bifurcation prediction models, on a CDC-curated dataset containing weekly influenza hospital admissions across all fifty U.S. states and the nation as a whole. We develop a pipeline for data preparation, reproduction number (Rt) estimation using the EpyEstim package (a Python implementation of the EpiEstim framework [4]), and bifurcation point identification. We systematically evaluate the models under several windowing strategies, including rolling, expanding, and fixed-length windows, and report accuracy, precision, recall, F1 score, and specificity. Our best-performing configuration (expanding window, length 100) achieves an accuracy of 90.09%, precision of 72.86%, recall of 96.23%, and F1 score of 82.93%. We additionally discuss Markov Regime Switching models as an alternative framework for detecting epidemic transitions. Our findings indicate that these models can generalize to real-world, state-level influenza data and are capable of online/real-time prediction.

Survival prediction frequently exhibit unstable and irreproducible performance, limiting their clinical applicability. Rather than proposing a new predictive architecture, this work presents a diagnostic investigation into the sources of this instability. Using a transformer-based deep fusion model combining RNA sequencing and histopathology data from the TCGA-BRCA cohort as a diagnostic probe, we demonstrate that global survival prediction is highly unstable across data splits. We then analyze the learned multimodal embeddings in an unsupervised manner, revealing stable latent patient subgroups with significantly divergent survival outcomes. These subgroups are not explained by canonical receptor status or immune infiltration, but instead align with tumor-intrinsic proliferative activity. Conditioning survival modeling on the largest discovered subgroup recovers strong and stable prognostic performance, substantially outperforming unimodal baselines. These findings indicate that survival predictability is conditional on latent heterogeneity and suggest that heterogeneity-aware modeling strategies are essential for robust multimodal prognostication. This work establishes a methodological framework for interpreting instability in multimodal survival learning.

Dynamic systems often evolve faster than reliable measurements can be collected. More frequent indirect observations may fill these gaps, but their noise and differing scales complicate estimation of the underlying state. We developed Bayesian Anchored Latent Learning (BALL) to combine sparse reference measurements with denser records on an interpretable scale. A bidirectional transformer teacher reconstructs trajectories from complete training histories, and a forward-only transformer student learns from those reconstructions and observed measurements. Across 30 simulations, BALL reduced composite trajectory error by 7% relative to the same transformer trained directly from measurements, achieving lower error across all datasets. In an applied example using psychiatric records from 1,609 patients, BALL improved the prediction of withheld questionnaire scores compared with direct transformer training at 21- and 28-day measurement schedules. Requests guided by estimated uncertainty further improved prediction. BALL connects retrospective learning with prospective state estimation and provides a framework for directing additional measurement.

ICD coding is the backbone of medical-record structuring, insurance payment, and clinical research. Today's best automatic coders are deep models, but they assume tens of thousands of labeled records, GPU servers, and machine-learning engineers. Many hospitals, especially small ones, have none of these. We study Chinese electronic medical records (EMRs) under a hard budget: 2,200 labeled records, CPU only, and fully interpretable decisions. Our method is lightweight and follows the way human coders work. First, codebook knowledge injection marks standard disease and operation names in the text as useful clues. Second, multi-view field weighting gives more weight to the clinical sections that matter most. Third, character n-gram TF-IDF features feed linear classifiers that score each candidate code. Fourth, a small set of clinical rules acts as a final safety check. Our main methodological finding is about this last step: naively applied rules do more harm than good (overall F1 drops from 0.8070 to 0.7995, and main-operation accuracy from 0.9727 to 0.9432). We therefore audit each rule on held-out data by counting how often it fixes a wrong prediction or breaks a correct one, and we remove rules that do more harm than good. This turns rule post-processing into a safety check that never hurts. On the CCL2026-Eval Task 6 dataset, our method reaches macro-F1 of 0.899 (main diagnosis) and 0.966 (main operation) in 5-fold cross-validation. A frequency-stratified analysis shows a clear long-tail limit: codes seen fewer than three times score zero F1. In a head-to-head test, a fine-tuned Chinese pretrained encoder scores lower than our method on all four tasks at this training size, while needing a GPU. Ablation tests show that the gains come mainly from field-weighted text. The framework trains in about 40 minutes on an ordinary CPU, and every prediction can be broken down into interpretable feature contributions.

Machine learning benchmarks often treat image files as independent observations, although multiple files may represent repeated views or other related samples. Prior work has shown that non-independent partitioning can inflate model performance, but the sensitivity of a biomedical benchmark to progressively increasing train-test relatedness is less often quantified. Here, we quantified how apparent classification performance changes when recoverable sample groups are progressively allowed to cross training, validation, and test partitions. Using source structure and filenames, we reconstructed analysis groups in a microscopy corpus compiled from multiple sources and generated five class stratified partitioning conditions, ranging from an intact group reference to random assignment of individual files. We applied the same training recipe within each model family to 129,173 images spanning 14 classes of blood cells and hematopoietic cell states. For a pretrained Vision Transformer, random assignment of individual files increased the mean test macro F1 from 0.853 to 0.977 and accuracy from 0.851 to 0.990 relative to the intact group reference. Apparent performance increased progressively across the intermediate conditions. The endpoint pattern also appeared with a ResNet-50 backbone, and the largest apparent gains occurred in immature granulocytic and blast classes. A descriptive analysis within the random assignment condition further showed higher scores among test images that had another member of the same group in the training set, although the exposed and unexposed subsets were not exchangeable. The grouping workflow, exact split definitions, code, and compact result files are released as HSD-Bench to support reproducible evaluation of hidden sample dependence.

Data curation-the process of collecting, cleaning, and joining raw datasets for unified analysis-is a time-consuming yet crucial part of any data science project. With the emergence of powerful agentic coding tools, it is tempting to "vibe curate"-i.e., to prompt AI agents to perform data curation operations and blindly trust their outputs in order to finish work more quickly. However, this can exacerbate two key challenges that already plague manual data curation workflows: reproducibility-the ability to trace the exact set of modifications applied to raw datasets-and verifiability-the ability to audit changes and confirm that data is processed properly. To address these issues, we introduce Scrub Data, a framework that enables the use of AI agents in data curation workflows while transparently maintaining both reproducibility and verifiability. The framework involves an interactive data curation loop between a user and an AI agent, backed by a data provenance graph that tracks and versions all dataset updates. The data graph requires that all updates are formatted as self-contained executable steps, allowing any version of a dataset to be reconstructed from raw data by playing forward the transformations stored in the graph. This workflow takes place within a lightweight web application that is designed to be modified by users and agents in real-time to create custom visualization tools for identifying data curation needs and verifying outcomes. We demonstrate the usefulness of the framework with a real data curation use case from animal movement ecology. Starting with 89M raw GPS coordinates, we use the framework to curate a benchmark dataset of 2.5M coordinates to be used for machine learning or statistical analysis. The framework is available for use and extension at https://github.com/justinkay/scrubdata.

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.

Motivation: Multi-center RNA-seq studies improve statistical power, but privacy regulations restrict patient-level data sharing. Meta-analysis methods avoid this restriction but lose power, especially under per-site imbalance. Federated learning allows sites to share only summary statistics. Among the three dominant differential expression (DE) frameworks, Flimma federates limma-voom and FedPyDESeq2 federates DESeq2. edgeR, still preferred for small-sample and high-variability designs, lacks a federated counterpart. Its iteratively reweighted least squares (IRLS) and Cox-Reid dispersion estimation make federation harder than limma-voom's single-pass fit. Results: We present FedEdgeR, a federated edgeR implementation protected by secure multi-party computation (SMPC), covering IRLS GLM fitting, three-level dispersion estimation, and the likelihood-ratio test. We evaluate it on four representative RNA-seq datasets: two tumor/normal cohorts, a multi-cohort anti-PD-1 immunotherapy study, and a 6-sample paired stress-test. FedEdgeR matches pooled edgeR under various metrics, such as Pearson r [≥] 0.99999 on -log10 p-values, 100% complete top-100 DE gene overlap, and F1 = 1.000 at the nominal FDR level 0.05. It consistently outperforms Fisher, Stouffer, random-effects, and RankProd meta-analysis on every tested dataset.

Aging is a network-level phenomenon, with its hallmarks interacting through dense feedback loops across vastly different timescales. Decades of reductionist research have produced thousands of mechanistic models of narrow subsystems, but no straightforward way to integrate them into one comprehensive system. Consequently, whole-cell and multi-hallmark aging models remain rare, manually constructed, and relatively small, despite broad agreement on their importance. The emergence of autonomous research agents offers a way to distribute this modeling effort across humans and artificial intelligence and thereby greatly accelerate it, but only if the underlying substrate allows agents to readily compose and analyze models with built-in validation checks and modeling guidance, without extensive custom code. We present hallsim, a JAX-native compositional simulation framework designed as the mechanistic substrate for agent-orchestrated research. It provides composability, end-to-end differentiability, GPU execution, and automated analysis and model-selection tools that facilitate composite model construction and fine tuning. The framework makes it easier to propose a mechanistic hypothesis, integrate it into an existing composite, reparametrize it against observed data, and batch-test it across different initial conditions. We demonstrate the co-simulation and in silico perturbation of three independently published kinetic models spanning genomic instability, nutrient sensing, and proteostasis. The models are connected by four edges and calibrated against a public dataset. Additionally, we train a Neural ODE surrogate and compose it alongside the mechanistic modules in a hybrid composite, demonstrating that mechanistic and neural models can function as complementary components of the same system.

Electronic health records provide a source of real-world information for disease risk prediction, but many machine learning models learn implicit associations that are difficult to inspect or use for clinical reasoning. Causal Bayesian networks (CBNs) offer a form of causal-interpretable machine learning by representing conditional dependencies, putative directional relations, and probabilistic evidence propagation in a directed acyclic graph. However, conventional CBN structure learning becomes unstable and computationally expensive when applied to large-scale EHR data. To address this limitation, we propose UPEBNL, a scalable framework for causal-interpretable cancer prescreening based on parallel CBN learning. UPEBNL integrates adaptive data slicing, quality-aware structure aggregation, and global DAG construction to learn stable and interpretable dependency structures from large-scale observational data. We evaluated UPEBNL in high-dimensional and multi-million-sample simulations and applied it to EHR-based prescreening for esophageal and colorectal cancer. In simulations, UPEBNL improved structural recovery accuracy by nearly 40% and achieved up to a 221.28-fold speedup over conventional CBN learning strategies. For cancer risk prediction, the learned CBNs provided interpretable evidence paths and achieved validation AUCs of 0.8171 for esophageal cancer and 0.784 for colorectal cancer. Calibration and decision curve analyses further supported the reliability and clinical utility of the models. These findings suggest that scalable CBN learning can support interpretable cancer prescreening from large-scale EHR data.

Forecasting and causal inference in complex, non-linear dynamical systems are fundamentally challenging due to noisy and high-dimensional time-series data. While Empirical Dynamic Modeling addresses this via state-space reconstruction in Euclidean space, it suffers from the "curse of dimensionality," distorting distance measurements as dimensions increase. This study proposes Forest-Weighted S-map (FORWS) and Forest-Weighted Causal Inference (FORWC), utilizing adaptive "forest weights" derived from random forest ensembles instead of Euclidean metrics. Using simulated and empirical ecological datasets, I demonstrate that these methods exhibit comparable or improved forecasting skill relative to conventional tools, while offering substantial resilience against dynamic process noise and mitigating the curse of dimensionality. Notably, FORWC enables multimodal causal inference, successfully extracting directed interactions between high-dimensional acoustic vectors and scalar temperature data monitored in a honeybee hive. By bridging chaotic physics with machine learning, this framework accurately extracts intrinsic manifold topology, providing a robust paradigm for uncovering causal networks in complex real-world environments.

Flow cytometry is central to clinical hematology, generating extensive archives comprising billions of single-cell profiles linked to expert-derived diagnostic summaries. While these records represent a largely untapped repository of collective expert knowledge, the underlying single-cell annotations required for supervised machine learning are typically not recorded in a standardized, accessible format. We present IMPGAT (implicit gating), a weakly supervised multiple-instance learning framework that reconstructs single-cell gating strategies directly from existing sample-level cell-type frequency distributions. The model utilizes an induced set attention mechanism to map single-cell marker profiles to specific cell populations, requiring neither explicit cell-level labels nor unsupervised clustering during training. We evaluate the framework on acute lymphoblastic leukemia (ALL) and acute myeloid leukemia (AML) cohorts, totaling 711 measurements from 268 individuals and 240 million single cells. In ALL, the model's quantification of cell-type abundances shows no statistically significant difference from human inter-expert variability. In AML, the framework achieves blast quantification on par with a recently established supervised state-of-the-art model that relies on explicit single-cell annotations for training. Error analysis shows that mismatches between model predictions and expert gating primarily involve expert-labeled non-cellular events, reflecting the non-incorporation of any explicit physical prior knowledge (such as the geometric relationships traditionally used to identify doublets and debris in scatter cytograms) in model training. Conversely, model certainty correlates with phenotypic borderline cases where manual bisection is inherently difficult. Our approach enables the systematic utilization of retrospective clinical records for the development of automated, standardized gating logic and is applicable to broader high-throughput contexts, such as large-scale immune monitoring or phenotypic drug screening.

Experimental assessment of developmental neurotoxicity (DNT) is costly, time-consuming, and difficult to scale across large chemical inventories. New approach methodologies (NAMs), including computational approaches based on machine learning, deep learning, and large language models (LLMs), offer new ways to screen and prioritize chemicals. However, these emerging computational approaches have largely focused on predicting toxicity endpoints, with less attention to the mechanisms underlying those predictions. For DNT assessment, understanding this mechanistic basis is equally important. In this study, we integrated LLM inference with the adverse outcome pathway (AOP) framework to connect DNT predictions with plausible mechanisms. We constructed an AOP-informed benchmark linking ChEMBL target-activity measurements for 5,861 compounds to 14 AOPs representing five neurodevelopmental outcomes and evaluated ten LLMs for the prediction of molecular initiating events (MIEs), key events (KEs), key event relationships (KERs), and adverse outcomes (AOs), as well as the generation of mechanistic explanations. Explanation quality was evaluated using both a naive LLM judge and a rubric-based LLM judge. Across the zero-shot models, the highest observed macro-F1 scores were 0.818 for MIEs and 0.463 for AOs, but only 0.056 for KEs and 0.034 for KERs, identifying downstream pathway reconstruction as the principal challenge. Few-shot prompting substantially improved performance, with Gemini 3.8 Flash under four-shot prompting achieving macro-F1 scores of 0.956 for AOs, 0.813 for MIEs, 0.492 for KEs, and 0.457 for KERs. Both LLM-judge configurations indicated improved mechanistic explanation quality with few-shot examples, although strong outcome prediction did not consistently ensure recovery of the recorded pathway. Our results demonstrate the potential of AOP-informed LLMs to support DNT assessment by linking predicted neurodevelopmental outcomes to plausible biological mechanisms and enabling predictions to be evaluated at both the outcome and pathway levels.

Inverse design has advanced the discovery of molecules and materials. Yet it remains largely unexplored for steering cell fate, where inputs are continuous, combinatorial culture conditions and outputs are distributions of cell states. Human hematopoiesis, in which combinations and doses of growth factors steer stem cells toward each blood lineage, is a demanding test case. While forward models can evaluate discrete conditions, the vast combinatorial space renders searching for a specific, target-producing design an intractable inverse problem. Here, we introduce LabCompass, an inverse-modeling framework that optimizes experimental designs toward candidates expected to produce specified target cell states, using a flow-matching generative model guided by cellular phenotypes predicted by a forward surrogate. We applied LabCompass to primary human CD34+ in vitro cultures, where continuous cytokine and small-molecule parameters govern hematopoietic cell fate. In a lab-in-the-loop setting, we experimentally validated model-driven designs and used these measurements to update the model across four successive loops. LabCompass recovered established lineage-inducing designs, improved predictions with new data, and enriched target populations across lineages. Incorporating new parameters expanded the design space and enriched rare populations beyond initial reach. Across loops, we generated a 78-million-cell spectral flow cytometry dataset and identified distinct culture conditions converging on the same target cell states. Our results establish generative inverse modeling as a compelling paradigm for navigating high-dimensional, continuous experimental spaces, and demonstrate that virtual cell models should be judged not only by how accurately they predict cellular responses, but also by how effectively they guide experiments toward desired biological outcomes.

This paper shows that it is possible to predict which newly published biomedical research papers will be connected to subsequent clinical trials. We use CAR-T cell therapy as a proof-of-concept setting because its development can be followed from published research through clinical testing, regulatory approval, and commercialization. We reconstructed connections between research papers and clinical trials and trained machine-learning models to predict which papers were associated with research that progressed to clinical testing. Using 65 paper-inherent indicators, including scientific topic, language, biomedical orientation, and team structure, the models achieved a cross-validated AUC of 0.908. The resulting rankings were also practically informative. In a retrospective analysis using no information generated after 2016, 80% of the top 20 papers came from laboratories associated with one of 13 predefined major CAR-T clinical or commercial outcomes. As an additional finding, first-year citations alone achieved an AUC of 0.797, while provenance indicators based on author, journal, and institutional reputation performed less strongly and added little to the paper-inherent model. Our results show that whether biomedical research papers will be connected to subsequent clinical trials can be predicted from newly published papers with high discrimination in CAR-T. Our findings could help research organizations and pharmaceutical companies identify papers, laboratories, and therapeutic programs that warrant closer evaluation before clinical outcomes are known.

Automated triage of whole-slide cervical cytology images remains constrained by the heavy dependence on single-cell bounding boxes or pixel-level annotations, which introduce significant bottlenecks, high expert overhead, and observer variability. To address these challenges, we present a weakly-supervised, detection-free Multiple Instance Learning (MIL) framework powered by a Gated Attention pooling engine that operates exclusively on slide-level predictions. By treating whole-slide images as bags of local instance patches, the dual-branch gated network dynamically assigns non-linear attention weights to highlight isolated dysplastic cells while suppressing benign background and debris, successfully resolving the "needle-in-a-haystack" problem inherent to high-grade lesions. Evaluated across multi-center cohorts including internal datasets (SIPaKMeD, Herlev, and CRIC) and the unannotated, out-of-distribution Mendeley LBC validation cohort processed via an unsupervised marker-controlled watershed pipeline our approach demonstrates robust generalization without requiring dense instance labels. Furthermore, our architectural comparison reveals a key trade-off: lightweight backbones like MobileNetV2 optimize in-distribution multi-center accuracy (90.96% Accuracy, 0.9800 ROC-AUC), whereas higher-capacity models like Xception ensure superior out-of-distribution robustness (80.43% Balanced Accuracy) under domain shifts. Ultimately, this detection-less pipeline provides a scalable and clinically viable solution to automate slide triage and alleviate cytopathologist shortages in resource-constrained environments.

Background Exposure-history data collected during epidemics or pandemics within digital contact-tracing applications hold immense potential for rapid, scalable epidemiological inference. The sensitivity of these data makes privacy-preserving designs essential to secure public trust and app uptake. However, during the COVID-19 pandemic, such approaches were not yet available at the scale or maturity required for deployment. We present a Bayesian decentralised framework to infer transmission dynamics from exposure-history data while protecting user privacy. Methods Our approach is based on iterative communication between user devices and a central server, computing log-likelihood gradients on device from exposure-history data and centrally updating parameter values using only aggregated gradients. Applying this framework to synthetic exposure-history data from a hypothetical respiratory pathogen epidemic simulated on dynamic contact networks, we assessed its ability to estimate the underlying transmission parameters, perform infector attribution, and reconstruct serial-interval (SI) distributions. We systematically explored how inference was affected by incomplete exposure-history data arising from partial app adoption, case under-ascertainment, and incomplete exposure logging, and by transmission scenarios. Findings Our decentralised framework accurately recovered transmission parameters, closely matching estimates derived from centralised data with no systematic bias in posterior means. Parameter recovery remained robust when hazards from unobserved exposures were known as a daily average across the population. Although incomplete exposure-history data and specific transmission scenarios introduced compensatory parameter biases, the inferred host infectiousness profile retained sufficient accuracy, and the weekly SI distributions reconstructed from infector attributions closely resembled the ground-truth distributions. Interpretation Our findings highlight that, beyond its primary role as a control measure, digital contact tracing could help characterise transmission dynamics almost in real time and inform public health interventions while preserving user privacy. Funding Oxford Martin School Programme on Digital Pandemic Preparedness; National Institute for Health and Care Research (NIHR) Health Protection Research Unit in Emerging and Zoonotic Infections.

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.