Integrating measurements across modalities and scales into unified quantitative frameworks of cellular function is a central goal in computational biology. Although powerful unimodal foundation models exist, the scarcity of large-scale paired data limits joint multimodal training. This constraint favours compositional foundation models, which fuse frozen unimodal experts through a learned interface. However, it remains unclear when such fusion provides task-relevant information beyond the strongest unimodal representation rather than merely aggregating redundant signal. Here we introduce Synergistic Information Score, a metric grounded in partial information decomposition that quantifies information gain arising only from cross-modal interactions. Extending theory from self-supervised learning, we show that standard alignment-based objectives with frozen encoders collapse to detect linear redundancies, which constrains their ability to capture nonlinear synergistic states. This distinction is especially important for tasks that connect tissue morphology with gene expression. Benchmarking ten fusion methods on spatial transcriptomics datasets, we demonstrate that tasks dominated by linear redundancies are well served by unimodal baselines, whereas complex niche definitions benefit from synergy-aware integration. Finally, we perform a scaling analysis and show that fine-tuning a dominant unimodal expert is the most sample-efficient strategy for standard tasks. Together, these results suggest that progress in multimodal modelling of cellular state requires shifting from alignment-based objectives to synergy-driven integration that exploits complementary cross-modal signal. Richter et al. show that simply aligning different types of biological data does not fully integrate them, revealing when data repeat the same information and, conversely, when they provide complementary insights.
Ecology aims to explain and predict how organisms interact with each other and their environments across space and time. Yet both ecological data and theory are fragmented, leading to models that generalize poorly beyond specific systems or scales. Empirical evidence spans diverse modalities, resolutions and contexts, while theory is distributed across partially overlapping frameworks that are rarely integrated within a single predictive model. We argue that ecological foundation models (ecoFMs), trained on large, multi-modal ecological datasets and informed by ecological theory, offer a route towards unifying data and theory within a common framework. Beyond their scientific value for ecology, ecoFMs present a challenging and consequential testbed for machine learning, demanding advances in multi-modal representation learning, theory-guided modelling and uncertainty-aware inference. By learning shared representations of organisms, environments and interactions, ecoFMs could improve generalization, link pattern to process and enable synthesis across ecological subdisciplines. We outline a roadmap for developing ecoFMs, including requirements for data infrastructure, model architectures, evaluation strategies and governance, and assess where current machine learning and ecological approaches fall short. If developed responsibly and collaboratively with ecological practitioners and other actors, ecoFMs could enable new modes of analysis and strengthen ecological forecasting, while simultaneously driving advances in machine learning for multi-modal integration, theory-guided learning and generalization in complex, data-limited systems. Ecology struggles to make broad predictions, as ecological data and theories are scattered across many different sources, scales and approaches. Gardiner et al. propose to create ecological foundation models, which combine large, multi-modal ecological datasets with ecological theory, as a way to create more generalizable and predictive models.
Artificial intelligence (AI) providers frequently update their models and system integrations without corresponding changes to their public identifiers. This undermines regulatory oversight and scientific reproducibility. We argue for meaningful AI versioning that meets the needs of regulators and researchers.
A neural network approach bridging Lagrangian and Eulerian perspectives transforms sparse observations into physically accurate, high-resolution maps of complex and chaotic fluid motions.
Knowledge-constrained network inference enables the reconstruction of regulatory programmes from high-dimensional biological data, yet the conditions under which shared and context-specific structures can be reliably identified remain poorly understood. Pablo Rodriguez-Mier et al. recently introduced CORNETO, a framework for multi-sample constrained inference. Here we reproduce, evaluate and extend CORNETO in terms of robustness, identifiability and bias-variance trade-offs. Using dynamic haematopoiesis and stable cortical datasets, we systematically assess four network reconstruction paradigms while preserving the original method-specific evaluation procedures and benchmark definitions. We show that multi-sample inference generally improves robustness and reproducibility, although gains in predictive performance depend on biological context and the completeness of the prior network. To enhance stability, we introduce bootstrap aggregation and selective re-expansion strategies that improve consensus robustness and recovery while preserving context-specific signals without modifying the underlying optimization scheme or sacrificing generality. Across datasets and methods, our results reveal distinct inference regimes governed by sample availability, prior knowledge and regularization. Together, these findings position network reconstruction as a regime-dependent learning problem and provide practical principles for robust multi-sample biological network inference.
