Building virtual cell dynamics that evolve with biological measurements

Zhang, Z.; Wang, Z.; Tian, Q.; Zhou, B.; Zhou, T.; Nie, Q.; Zhou, P.

doi:10.64898/2026.10.23.761350

Home/Archive/bioRxiv

bioRxiv · 10.64898/2026.10.23.761350

Building virtual cell dynamics that evolve with biological measurements

Zhang, Z.·Wang, Z.·Tian, Q.·Zhou, B.·Zhou, T.·Nie, Q.·Zhou, P.

View original source ↗Export metadata ↓

Abstract

Virtual cell models built from single-cell measurements need to predict how cells change over time, but new biological measurements often reveal dynamical relationships that existing models cannot express. Translating such measurements into dynamics therefore requires deciding what biological relation should be modeled, where it should enter the formulation and how its recovery should be validated. Here we present CellCompass, an agentic AI system that couples a scientific agent with an extensible dynamical engine for modeling state transitions and population change, along with fate and molecular readout. CellCompass reuses existing formulations when sufficient and designs, implements and refines new ones when experimental measurements introduce additional constraints. To assess this capability, we developed DynBench100, a benchmark of 100 simulated single-cell time-course scenarios with hidden dynamical ground truth. CellCompass achieved the highest aggregate performance among the tested general-purpose and biomedical agents across six dynamical tasks. Across four real single-cell studies, it retained an existing velocity-growth model to reconstruct context-dependent epithelial-mesenchymal transition and developed three new dynamical formulations for distinct experimental measurements. In drug-response data, incorporating metabolic RNA labeling reduced pulse-control identifiability error by 31.1%, revealing staged response programs and generating testable predictions for perturbing a candidate adaptive program. In hematopoietic lineage tracing, incorporating clonal barcodes improved clone-fate agreement by 54% over the strongest dynamical baseline, identifying early lineage priming among transcriptionally similar progenitors. Finally, in time-resolved CITE-seq data, CellCompass reduced future surface-protein distribution error by 13.2% for RNA-initialized trajectories, uncovering lineage-specific temporal offsets between transcriptional and protein programs. Together, CellCompass provides a framework in which virtual-cell dynamics can expand with the biological measurements used to constrain them.

Explore related subjects

bioinformatics Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Zhang, Z., Wang, Z., Tian, Q., Zhou, B., Zhou, T., Nie, Q., Zhou, P.. 2026-10-27. Building virtual cell dynamics that evolve with biological measurements. https://doi.org/10.64898/2026.10.23.761350

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Network analysis reveals converging molecular mechanisms in genetic epilepsies

Background: Rare genetic epilepsies are often difficult to diagnose and treat. They can arise from changes in many different genes, and the molecular basis linking these changes to epilepsy remains poorly understood. This complexity makes it difficult to select effective treatments and develop new therapies, leaving many patients with poorly controlled seizures and reduced quality of life Methods: To address this gap, we first identified epilepsy-associated genes and conducted a network-based analysis of biomolecular interactions in genetic epilepsy to identify converging pathways with high translational relevance and potential therapeutic targets. To run the analysis, we constructed a human brain-specific molecular interaction network by integrating all human pathways from the WikiPathways database. HotNet2 was then applied to identify highly perturbed clusters through evidence-based weight diffusion of EAGs in the human brain network. Results: We identified 2,496 epilepsy-associated genes (EAGs) from multiple resources. These genes encode coding for metabolite interconversion enzymes (20%), transporters (15%), and protein-modifying enzymes (11%). Furthermore, the network analysis identified 19 subnetworks with more than two nodes and 125 unique genes. Gene enrichment analysis revealed convergence of processes including action potential, synaptic function, nervous system development, TOR signaling, cilium organization, DNA repair, and cell fate specification. We identified 44 first-order neighbors connecting the subnetworks. By integrating drug-target interactions from DrugBank using CyTargetLinker, we highlighted potential therapeutic targets within the S4 subnetwork, associated with cerebral cortex development, as well as CREB1, a connector between subnetworks. Conclusion: Network analysis organized heterogeneous epilepsy-associated genes into biologically related subnetworks and generated hypotheses for further research.

bioinformatics↗

bioRxiv2026

GaussF: An Unbiased Gaussian Framework for High-Precision Transcript Quantification with a Programmable Self-Benchmark Model

Short-read RNA sequencing has revolutionized transcriptome analysis, but accurate isoform-level quantification faces substantial challenges from overlapping sequences among complex isoforms and inherent sequencing biases, leading to uneven coverage. Current alignment-based and alignment-free methods, while valuable, largely rely on empirical, data-derived factors for bias correction and isoform resolution. This dependency limits their capacity to fully capture the intricate spectrum of sequencing data. We introduce GaussF, a novel alignment-free methodology specifically designed for precise isoform quantification by directly addressing these biases and the isoform recognition problem without relying on empirical data. GaussF is predicated on the observation that k-mer counts categorized by GC content across a natural transcript exhibit an intrinsic Gaussian distribution. It establishes a transcript-specific self-benchmarking workflow utilizing a theoretical Gaussian cumulative distribution function (CDF). Crucially, the parameters defining this CDF are derived independently of any empirical sequencing data. This parameter-fixed theoretical CDF serves as an unbiased benchmark, capturing only the genuine contribution of abundance while systematically excluding all bias-related data. To improve isoform resolution despite shared regions, GaussF characterizes isoforms through identifying minimal-shared regions and restructures the k-mer reference accordingly. This approach simplifies complexity and enhances accuracy. Rigorous validation using artificial and real RNA samples, alongside comparative analysis with leading quantification tools like Kallisto, Salmon, Cufflinks, StringTie, and RSEM, demonstrates GaussF's superior performance in accurate isoform detection and quantification. GaussF is available on GitHub and PyPI, representing a significant advancement in RNA-seq analysis.

bioinformatics↗

bioRxiv2026

Predicting Differential Protein-Protein Interactions

Motivation Analysis of context-specific protein-protein interactions (PPIs) provides deeper insight into protein function, revealing how cellular pathway components interact in certain perturbation conditions. However, identifying such interactions experimentally often requires combining affinity purification with mass spectrometry, proving costly and time-consuming. On the other hand, it has been shown that an increase in protein abundance does not always correspond to an increase in protein interaction, making it an unsuitable surrogate for interaction studies. Results To address this, a random forest model was created using PPI-specific MINT embeddings and condition-specific mass spectrometry protein abundance information to predict differential interactions across two experimental conditions. The model was trained and tested on validated context-specific PPIs from eight distinct cell-lines, demonstrating proficiency in predicting differential PPIs for both novel proteins and cell lines. To our knowledge, we provide the first computational model that predicts differential PPIs between experimental conditions. Availability and implementation The code is available on GitHub and backed up on Zenodo (DOI: 10.5281/zenodo.23758127). The model is freely available on Zenodo (DOI: 10.5281/zenodo.23757927).

bioinformatics↗