Longevity InTime: Autonomous AI Institute. Anti-Aging Digital Health Immortality Transhumanist AI Channel

Mol-JEPA predicts drug‑like molecule properties from structure and biology Mol‑JEPA, a model for forecasting the properties of medicinal molecules, was released as a preprint on 23 August. The authors compiled data for 4.69 million small molecules, encompassing 14 types of information ranging from chemical structure to cellular measurements and ADMET profiles (absorption, distribution, metabolism, excretion, and toxicity). During training, Mol‑JEPA randomly masks one data type and learns to reconstruct its compressed numerical description from the remaining modalities. This forces the model to link a molecule’s structure with its biological and pharmacological observations without distorting the underlying structure. The approach addresses cases where a candidate binds strongly to a target protein but is rapidly cleared or toxic, and where minor structural tweaks can dramatically alter ADMET outcomes. By preserving the exact structure and training on multiple observation types, Mol‑JEPA learns a unified representation that captures diverse molecular behaviors. Mol‑JEPA encodes each data type—atomic bond schemata, calculated chemical features, biological assay results, cellular profiles, and ADMET data—into a short numerical vector of uniform format. A shared module then receives the available vectors and predicts the representation of the masked type, thereby training the model to connect structure with function. In property‑prediction tasks on small subsets, Mol‑JEPA achieved lower mean absolute error than competing methods, and its advantage widened as test molecules became more structurally distinct from the training set. On public temporal splits of the TabICLv2 benchmark (a tabular‑data model), Mol‑JEPA was sometimes outperformed in both MAE and head‑to‑head comparisons. Random splits can place molecules sharing a core scaffold in both training and test, leading to overly optimistic estimates; the authors note that evaluating on structurally distant compounds provides a stricter test of transferability. An ablation study showed that adding more information improves accuracy: a version trained only on molecular graphs and ECFP4 fingerprints was outperformed by the full model using all data types, which reduced MAE by 14 % on a simple task and 13 % on a more complex one. For drug development, linking molecular structure to accumulated behavior data is essential. Mol‑JEPA consolidates this information into a single numerical descriptor that can be used to predict the properties of new drug candidates. 🔗 Read original →