They are Yann LeCun’s solution to many of the limitations with LLMs
The website: AIDO Cell: A General-Purpose Simulator for Cell Biology
1 Like
i still cant properly make them model a MSN dopamine neuron or david goodsell diagram… (or pass topobench/knotbench). it also botched creating a world model of my MRI+CT scans (though its a good START!!)
World models for biomedicine
Biological systems respond dynamically to changes in their environment, and predicting these responses before complex interventions are applied could change how computational models are used in biology and medicine. In artificial intelligence, world models represent the state of a system and simulate how it evolves under alternative actions. We envision extending this paradigm to biomedicine, with world models that span molecular, cellular, tissue, and clinical scales and simulate biological trajectories under alternative interventions. Building such models will require methods that combine observational data with interventional and longitudinal measurements across modalities. Biomedical world models could open new modes of inquiry by predicting how biological systems respond to novel perturbations, searching for interventions that drive them toward desired states, and planning experimental campaigns over extended time horizons.
https://www.cell.com/action/showPdf?pii=S0092-8674(26)01005-6
AI World Models: Simulating the Future of Biological Interventions
A recent perspective paper outlines the conceptual transition from static artificial intelligence foundation models to dynamic “biomedical world models”. These advanced algorithms are designed to simulate how complex biological systems evolve over time in response to specific interventions such as genetic perturbations or drug treatments. By integrating observational data with interventional and longitudinal measurements, world models aim to predict biological trajectories across molecular, cellular, tissue, and clinical scales.
The scientific method relies on systematic intervention followed by observation. However, the combinatorial explosion of potential biological interventions far exceeds experimental capacity. For example, licensed drugs currently target less than 15 percent of the estimated druggable genome. To navigate this vast intervention space, researchers propose the development of biomedical world models. Unlike standard predictive or generative artificial intelligence that maps rich biological states to narrow targets, world models explicitly represent the state of a system and simulate its evolution under alternative actions.
The core architectural requirement of a world model includes a defined biological state, an externally specifiable action representation, and a biological dynamics transition model. These systems generate distributions over possible future states rather than single point predictions, thereby preserving inherent biological stochasticity and measurement uncertainty [Confidence: High]. The transition models process partial observations and predict branching future trajectories, enabling long horizon sequential planning.
Developing these models presents profound data challenges. Current single cell perturbation screens provide causal evidence but only cover a narrow range of cellular contexts. Developmental atlases capture natural biological dynamics but overwhelmingly rely on destructive measurements that prevent tracking the same individual cells over time. Electronic health records offer longitudinal patient trajectories but suffer from unmeasured confounding variables and selection biases.
To overcome these limitations, the authors argue for new data generation paradigms. These include non-destructive live cell imaging, molecular recorders, and automated self-driving laboratories capable of executing closed loop experimental designs. By continuously updating models with experimental feedback from programmable cloud laboratories and high throughput organoid screens, biomedical world models could eventually identify optimal interventions and sequentially plan complex treatments. The ultimate utility of these systems lies not merely in their predictive accuracy but in their decision utility for prioritizing real world experimental and clinical pathways [Confidence: High].
Insights
The provided paper outlines a theoretical computational framework rather than presenting novel experimental data. Consequently, there are no specific dietary, pharmacological, or physiological interventions tested, and no measurable effect sizes exist to calculate or extract. For individuals tracking physiological biomarkers or testing geroprotective protocols, the practical takeaway is an awareness of the current limitations of artificial intelligence in biology. The paper highlights that current models cannot yet reliably predict multi-step, longitudinal biological responses to novel compound combinations due to a lack of temporally linked interventional data. Reliance on artificial intelligence platforms to safely design complex longevity stacks remains premature until these systems master temporally compositional rollouts without compounding error drift [Confidence: High].
Context/Source
-
Full Title: World models for biomedicine.
Institution: Harvard Medical School, University of Oxford, Harvard University, Kempner Institute, Broad Institute. - Country: United States, United Kingdom.
- Journal: Cell.
- Impact Evaluation: The impact score of this journal is 45.5, evaluated against a typical high-end range of 0–60+ for top general science, therefore this is a Elite impact journal.