Why does this matter? We believe the Almanac is the foundation for a new kind of scientific instrument that will enable AI-guided ensembles of MLIPs (with improving quality weekly) to be pointed at the hardest problems. Recent work from Sathya Edamadaka, Soojung Yang, Ju Li and @rgblabmit, which was a core motivator of the Almanac work, shows convergence in representation space between models. But, despite this convergence, no single model is yet enough for most research tasks. arxiv.org/abs/2512.03750 Demo coming soon: agent-driven MLIP committees orchestrated with Academy on @globus Compute - agents that query many potentials, quantify disagreement, and decide what to trust or what needs more detailed simulation.

Introducing Atomistic Language Models (ALMs), a new paradigm to unify atomistic understanding, materials generation, and natural language. ALMs set the SoTA on crystal structure prediction, de novo generation, and editing materials as instructed by text.

This paper shows that wildly different AI models for molecules, materials, and proteins are independently learning the same underlying representation of matter suggesting we’re converging on a shared physics-grounded “latent reality” that proves scientific foundation models might actually be generalizable across domains

The surprising ineffectiveness of molecular dynamics coordinates for predicting bioactivity with machine learning 1. The study challenges the assumption that molecular dynamics (MD)-derived coordinates are superior for machine learning-based bioactivity predictions, revealing that they often underperform compared to minimum-energy conformations. 2. Using over 2600 protein-ligand complexes, the authors systematically compared MD-derived and minimum-energy coordinates, employing three descriptor sets and machine learning algorithms like Random Forest, XGBoost, and Support Vector Regression. 3. Surprisingly, MD-derived conformations failed to consistently outperform minimum-energy structures, even though they provide dynamic representations of molecular interactions. 4. In certain cases, ensemble averaging of MD-generated snapshots improved predictive performance slightly, but the benefits were not proportional to the computational costs. 5. Extended Connectivity Fingerprints (ECFPs), a 2D molecular representation, outperformed MD-based models in many cases, questioning the utility of complex 3D data for predicting bioactivity. 6. The findings highlight a critical need for better 3D and dynamic molecular representations. The study suggests exploring geometric deep learning or incorporating protein information to improve machine learning models. 7. The authors propose a tiered approach: using fast, simpler methods like ECFPs for initial screening, followed by MD-based predictions for refining top candidates. 8. The study serves as a wake-up call for molecular machine learning, emphasizing the importance of balancing data complexity, computational cost, and predictive accuracy. @fra_grisoni @DerekvTilborg @Rza_ozcelik 📜Paper: chemrxiv.org/engage/chemrxiv… #MachineLearning #DrugDiscovery #MolecularDynamics #Bioinformatics #AI

Instead of the typical (somewhat technical) lecture, we discussed a couple of papers, including a favorite J. Chem. Ed. activity of mine: "Chemistry from Telephone Numbers: The False Isokinetic Relationship." It's a lot of fun. Do check it out. 2/3 pubs.acs.org/doi/abs/10.1021…