From Cancer Cells to Planetary Orbits: Can One Predictive Core Learn Different Worlds?

Introduction

Weather systems, molecules, cells and planetary orbits appear to obey unrelated rules. Yet they share a practical modeling problem: given the current state, can a system predict how it will change? JEPA-Anything, presented by PhAI Labs and collaborating universities, explores whether one predictive principle can be reused across such domains. It is not an attempt to throw every type of data into a single input format.

The core idea: separate predictable changes

Joint-Embedding Predictive Architectures, or JEPAs, predict target representations in latent space rather than reconstructing every detail of the original observation. JEPA-Anything adds OPF, or Orthogonal Predictive Factorization. The target state is divided into complementary subspaces, each handled by a separate predictive branch and recombined afterward.

Orthogonality constraints are intended to reduce redundant learning across branches. Factor-activity and encoder-variance constraints help prevent inactive factors and representation collapse. Importantly, the factors are not assigned in advance to concepts such as velocity, temperature or a particular biological pathway. They are learned from structures that are predictable in the data. Domain-specific encoders and observation formats remain in place, while the predictive core and latent interface are shared.

Results across seven system types

The report evaluates the approach in visual, biological, clinical, control, molecular, physical-field and weather settings. In a controlled intervention benchmark, it reduces prediction error for both individual interventions and previously unseen combinations of factors. On a matched dynamics benchmark, nine of ten prediction tasks show an improvement over standard JEPA. The report also describes lower errors on Burgers’ equation, shallow-water dynamics and WeatherBench 2.

Long-horizon molecular rollouts involving liquid water, alpha-quartz, paracetamol and benzene also favor JEPA-Anything in the reported comparisons. The results are not uniformly positive, however. Standard JEPA performs better on the Hopper control task, and the advantage of the new method becomes smaller as rollout horizons increase. The evidence therefore supports a reusable predictive principle, not a finished universal world model.

When latent factors become scientific interfaces

The researchers also inspect the learned factors instead of treating them as an opaque optimization artifact. In a liver-cancer study, analysis of the latent space led to a candidate intervention combining IL-18 with NT5E/CD73 blockade. The reported testing progressed through cell-based systems, patient-derived organoids, tumor tissue fragments and immunocompetent mice, where the combination showed strong tumor-killing and immune-activation signals in the tested samples. This remains a research finding that requires continued and independent validation.

A separate planetary-orbit experiment provides a physics check. The model receives simulated positions and velocities without being given Kepler’s law. Researchers then analyze latent modes and obtain a frequency–semi-major-axis slope of -1.4991, close to the theoretical value of -1.5. The example suggests that latent representations may be compared with known laws, not merely judged by predictive error.

Why it matters—and what it does not prove

JEPA-Anything’s significance is more modest than the phrase “AI understands everything.” It proposes that different sciences can retain their own data interfaces while sharing a mechanism for learning predictable change. If internal factors can be inspected, connected to hypotheses and tested through experiments or established laws, world models could become intermediate tools in scientific discovery. Their cross-domain robustness, long-horizon stability and independent reproducibility remain open challenges.

Source: QbitAI