In a COVID-19 patient cohort, predicted intermediate profiles improved retrospective disease-stage stratification relative to observed profiles alone, while expert programs highlighted immune and inflammatory signals associated with severity, and these results support biologically structured continuous-time modeling for prediction, interpretation and virtual-perturbation prioritization from sparse temporal omics data.
Abstract
Perturbation-omics experiments usually measure only a subset of molecular feature, intervention and time space, leaving many response trajectories, perturbation effects and disease- or differentiation-associated transitions unobserved. Here we present MEGA-ODE, a graph-constrained continuous-time framework for reconstructing sparse dynamic omics landscapes, predicting unmeasured molecular states and prioritizing virtual perturbations toward defined biological endpoints. MEGA-ODE integrates molecular-network priors, graph neural ordinary differential equations and context-adaptive mixture-of-experts routing. In L1000 transcriptomic perturbations and CPPA proteomic drug-response data, MEGA-ODE improved held-out-feature and unseen-perturbation prediction over baseline methods, and in SARS-CoV-2 infection time-series data it remained competitive for future-time-point forecasting. In a COVID-19 patient cohort, predicted intermediate profiles improved retrospective disease-stage stratification relative to observed profiles alone, while expert programs highlighted immune and inflammatory signals associated with severity. Across the MAPK drug-response and stem-cell differentiation case studies, graph- and expert-level attributions prioritized perturbation-associated MAPK edges, developmental regulators and TF-target relationships supported by independent promoter-proximal ChIP-seq overlap. In hESC-to-definitive-endoderm differentiation, MEGA-ODE prioritized candidate transcription-factor perturbations predicted to shift 12-36 h profiles toward 96 h definitive-endoderm marker signatures, framing trajectory navigation as a concrete hypothesis-generation task. Together, these results support biologically structured continuous-time modeling for prediction, interpretation and virtual-perturbation prioritization from sparse temporal omics data.
Single-cell perturbation profiling maps intervention-induced phenotypes, yet experiments measure only a fraction of the perturbation-context space. Learning context-dependent perturbation effects could enable response prediction beyond measured conditions. Here we introduce PerturbLDM, a latent-diffusion framework for conditional generation of single-cell transcriptional responses. Following Tahoe-100M pretraining, it predicted 13,942 held-out combinations of observed drugs, doses and cell lines more accurately than existing methods, with higher matched-control effect correlation than an additive marginal baseline in 95.2% of conditions. The Tahoe-100M-pretrained model was further used to rank PANACEA compounds by pathway similarity, placing shared-mechanism pairs among nearest neighbours. In smaller datasets, Per-turbLDM generated a mid-gestational fetal-colon state with 67% lower gene-wise error than Squidiff, retaining the balance between absorptive and BEST4/OTOP2-like epithelial programmes. In PBMCs, it captured six of seven interferon and antiviral programmes and the interferon-associated FAO–OXPHOS programme more accurately than scGen. Together, these results support conditional response generation across data scales and biological settings.
Lishan Yu, Kanglin Hsieh, Yan Chu et al.· bioRxiv· 0 citations
This work proposes MechGNN-Epi, a hybrid framework that couples a spatiotemporal graph encoder with a differentiable SIR update that yields epidemiologically constrained trajectories and produces region- and time-indexed parameter proxies that can be inspected as diagnostic signals, while not being guaranteed as causally identifiable mechanistic parameters.
Background The human gut microbiota is a highly complex ecological system closely linked to host health, yet the functional mechanisms underlying its dynamic behavior remain poorly understood. Accurate modeling of microbial community dynamics is essential for elucidating these mechanisms. However, most existing approaches rely on densely sampled time-series data and often lack biological interpretability. Methods To address these challenges, we propose gNODE, a framework that integrates the generalized Lotka-Volterra (gLV) model with neural ordinary differential equations (NeuralODEs) to jointly predict microbial community dynamics, infer species interactions, and quantify the functional contributions of key taxa. By embedding ecological equations into a neural architecture, gNODE incorporates biological constraints directly into its model structure, enabling biologically meaningful parameter estimation and accurate inference even under sparse temporal sampling. Results Through simulations and real datasets, gNODE demonstrates superior performance in parameter estimation, trajectory prediction, and perturbation response modeling compared with existing methods. In a Clostridioides difficile infection dataset, gNODE accurately captured post-infection community trajectories and identified key inhibitory taxa, highlighting its potential to discover microbes that suppress pathogens. In a probiotic cocktail colonization dataset, gNODE identified diet-specific keystone species, underscoring its utility for assessing perturbation responses and guiding the design of probiotic consortia. Conclusion gNODE provides a robust and interpretable framework for modeling complex microbial community dynamics, offering new mechanistic and functional insights into the ecological processes that shape host-associated microbiomes.
Xiaoxiu Tan, Feng Xue, Lu Xie et al.· Frontiers in Cellular and In...· 0 citations
Predicting cellular responses to unseen chemical perturbations is challenging due to unknown targets and mechanisms, high-dimensional expression responses, and limited experimental coverage of the large small-molecule design space. We propose PerturbPFN, a PFN-style amortized model for unknown-target perturbation prediction under a hierarchical synthetic structural prior. Instead of directly regressing high-dimensional expression responses, PerturbPFN infers a latent system graph, sparse atomic intervention targets, and intervention strengths, then propagates their effects through an SCM decoder. The model is trained entirely on prior-predictive synthetic episodes generated from biologically motivated graph and expression simulators, enabling structured in-context learning without test-time gradient updates. We evaluate PerturbPFN on both real single-cell perturbation data and synthetic benchmarks, covering effect prediction, target identification, and regulatory structure discovery. Our results show that PerturbPFN offers a complementary trade-off to specialized baselines, achieving competitive perturbation prediction with low inference cost while exposing interpretable intermediate estimates of targets, strengths, and system structure.
Yuche Gao, Jos'e Miguel Hern'andez-Lobato, Siyuan Guo· 0 citations
HyperODE is introduced, a surrogate capable of operating across an entire class of approximately mass-conserving compartmental models without retraining, by mapping the structure of ordinary differential equations into directed hypergraphs, which decouples the functional form of system interactions from the neural network architecture.
Abstract Motivation Understanding disease dynamics from longitudinal multi-omics is hindered by traditional approaches that focus on univariate trajectories and static networks while ignoring temporal evolution. We developed DiDyNet, a framework for identifying phenotype-specific temporal molecular networks by defining dynamic coupling as coordinated molecular trajectories. DiDyNet operates through four steps: (i) two-dimensional variance-based filtering to prioritize dynamic features; (ii) quantification of subject-specific coordination using Dynamic Time Warping to accommodate asynchrony; (iii) statistical testing for differential dynamic couplings; and (iv) linear mixed model-based post-hoc refinement to distinguish genuine coordinated dynamics from stochastic noise. Results Simulation studies showed that DiDyNet significantly outperformed static summary statistics, including the mean, median, and difference, which cannot capture dynamic signals. Dynamic Time Warping-based quantification also demonstrated greater robustness than Euclidean distance, correlation-based distance, and constrained alignment methods under temporal misalignment and signal sparsity. Application to an insulin resistance cohort identified a coordinated cross-omics network linking systemic inflammation with intracellular stress responses. Availability Source code is freely available at https://github.com/bioinfoliu/DiDyNet.
Zhe Liu, Ke Wu, Taesung Park· Bioinformatics Advances· 0 citations