Applications of temporal graph learning for predicting the dynamics of biological systems
Manuel Dileo (✉), Andrea Sottoriva · 2026
Read Applications of temporal graph learning for predicting the dynamics of biological systems by Manuel Dileo (✉), Andrea Sottoriva (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-28659.txt.
Introduction
Manuel Dileo and Andrea Sottoriva explore temporal graph learning for predicting living-system motion. Static foundation models apply transformers to gene tables but miss how growth programs change over time. Their approach infers pseudotime paths from single-cell data, splits cells into growth snapshots, rebuilds one gene control network per snapshot, and predicts states with time-based graph neural networks. On two mouse growth sets, graph models beat scGPT and scFoundation, and time-based learning captures control motion including key gene hubs.
Catalog record
| CREATOR | Manuel Dileo (✉), Andrea Sottoriva |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40518 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.28659
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Applications of temporal graph learning for predicting the dynamics of biological systems Manuel Dileo (✉), Andrea Sottoriva Abstract Recent biological foundation models have shown strong performance in single-cell representation learning by applying transformer architectures directly to gene-expression matrices. However, these approaches predominantly operate in static settings and do not explicitly model the temporal evolution of developmental programs in the cell. Modeling such dynamics is important for understanding how cellular states progressively emerge, differentiate, and reorganize during development or disease progression. In this work-in-progress paper, we investigate an alternative temporal graph-based perspective in which cellular states are represented through pseudotime-resolved gene regulatory networks and modeled as evolving graph structures over persistent gene identities. Starting from single-cell transcriptomic data, we infer pseudotime trajectories, discretize cells into developmental snapshots, reconstruct one gene regulatory network per snapshot, and apply temporal graph neural networks to forecast future biological states. We evaluate this framework on two publicly available mouse developmental datasets, erythroid gastrulation and pancreatic endocrinogenesis, considering three complementary tasks: gene-expression forecasting, future link prediction, and future out-degree centrality prediction. Our results show that graph-based models outperform well-known biological foundation-model such as scGPT and scFoundation, suggesting that explicitly modeling evolving regulatory structure provides useful information beyond static pretrained representations. For link prediction and centrality forecasting, temporal graph learning captures non-trivial regulatory dynamics and enables the identification of temporally important gene hubs. Overall, our findings support temporal graph learning as a promising direction for modeling dynamic biological systems and as a complementary paradigm to current foundation model approaches in single-cell biology. 1 Introduction Recent advances in single-cell transcriptomics and biological foundation models have significantly improved the ability to learn representations of genes, cells, and cellular states from large-scale transcriptomic datasets. In particular, transformer-based architectures such as scGPT [ 9 ] , scFoundation [ 14 ] , and Geneformer [ 31 ] have shown strong performance across a broad range of downstream tasks, including cell annotation, perturbation prediction, and gene-expression imputation. These approaches typically operate directly on gene-expression matrices and learn contextualized representations through masked reconstruction objectives analogous to masked language modeling. Despite these advances, most existing approaches remain fundamentally static and do not explicitly model the temporal evolution of cells. This limitation is particularly relevant in developmental biology, where cellular states continuously emerge, differentiate, and reorganize over time and during diseas progression. Modeling such dynamics remains challenging because standard single-cell RNA sequencing technologies are inherently destructive: each cell can only be observed once, preventing direct temporal tracking of individual cellular trajectories. As a result, developmental progression must instead be reconstructed computationally from large populations of asynchronously observed cells. Recent advances in pseudotime inference and RNA velocity analysis have partially addressed this limitation by enabling the reconstruction of ordered developmental trajectories from single-cell transcriptomic data [ 6 , 13 ] . However, most existing methods still focus on modeling cellular states independently and do not explicitly represent the evolving regulatory structure underlying developmental transitions. At the same time, recent progress in temporal graph learning has shown that temporal graph neural networks (TGNNs) can effectively model evolving relational systems by jointly capturing structural dependencies and temporal dynamics across graph snapshots [ 19 ] . This naturally raises the question of whether developmental biological systems can be reformulated as evolving graph processes suitable for temporal graph learning. In this work-in-progress paper, we investigate temporal graph learning as a framework for modeling and predicting the evolution of biological systems from single-cell transcriptomic data. Starting from single-cell gene-expression matrices, we first reconstruct developmental trajectories through diffusion pseudotime [ 13 ] , discretize cells into developmental snapshots, and infer one gene regulatory network (GRN) per pseudotime bin using SCENIC [ 1 ] . This procedure transforms developmental single-cell systems into discrete-time temporal graphs with persistent gene identities and evolving regulatory interactions. Temporal graph neural networks are then applied to learn temporal node representations and forecast future biological states. We evaluate this framework on two publicly available mouse developmental datasets, erythroid gastrulation and pancreatic endocrinogenesis, considering three complementary forecasting tasks: future gene-expression prediction, future link prediction, and future out-degree centrality prediction. Our results show that graph-based temporal models outperform widely used biological foundation models such as scGPT and scFoundation on gene expression forecasting, suggesting that explicitly modeling evolving regulatory structure provides useful information beyond static pretrained representations. Interestingly, spectral graph architectures such as ChebNet and GCRN-GRU consistently emerge as some of the strongest-performing models across tasks, suggesting that spectral graph convolutions may be particularly well suited for the heterophilous and hub-centered structure of gene regulatory networks. Overall, our findings support temporal graph learning as a promising and complementary paradigm to current foundation-model approaches for modeling dynamic biological systems. 2 Background Virtual cell modeling. Recent advances in foundation models for single-cell biology have stimulated growing interest in virtual cell modeling, namely the development of computational systems capable of simulating cellular states, perturbations, and transcriptional responses [ 8 , 25 ] . Most existing approaches are based on large pretrained transformer architectures trained through masked gene reconstruction objectives or perturbation prediction tasks [ 9 , 14 ] . These models primarily focus on learning contextualized representations of genes and cells from large-scale transcriptomic corpora, enabling downstream applications such as cell-state annotation, perturbation prediction, and gene-expression imputation. Graph-based approaches such as GEARS [ 26 ] and HEIST [ 20 ] have introduced explicit modeling of gene interactions. However, these methods still operate in essentially static settings and do not explicitly model the temporal evolution of regulatory programs across developmental trajectories. In contrast, temporal graph learning provides a natural framework for modeling evolving biological systems. Temporal Graph Neural Networks. Temporal graph neural networks (TGNNs) are deep learning models designed to learn, predict, and reason over evolving relational structures. A comprehensive overview of the main architectures and design choices is provided in recent surveys [ 19 ] . Existing TGNN approaches are commonly categorized according to how temporal information is represented. In continuous-time models, dynamic graphs are treated as streams of timestamped events, whereas in discrete-time or snapshot-based approaches, temporal networks are represented as ordered sequences of static graphs observed at a chosen temporal granularity [ 15 ] . These two formulations target different types of temporal dynamics and lead to substantially different modeling assumptions and evaluation protocols. Event-based approaches are particularly suitable for high-frequency interaction streams and fine-grained temporal reasoning tasks, whereas snapshot-based approaches are better aligned with systems evolving through relatively coherent temporal phases [ 21 ] . In this work, we focus on discrete-time TGNNs, since developmental biological systems naturally evolve through progressive transcriptional states that can be represented as ordered pseudotime-resolved snapshots. Among discrete-time TGNNs, some of the most effective architectures combine graph message passing with explicit temporal update mechanisms. Common design strategies include parameter evolution [ 21 ] , recurrent embedding evolution [ 29 ] , and temporal memory modules inspired by sequence modeling [ 34 ] . Alternative lines of work incorporate temporal point process (TPP) components to model event intensities and temporal dependencies, offering a complementary perspective on temporal dynamics that can be adapted to discretized settings [ 24 ] . Overall, these approaches highlight the importance of jointly modeling structural dependencies within snapshots and temporal evolution across snapshots. Gene Regulatory Networks.