The exponential growth of scientific literature presents an increasingly acute challenge across disciplines. Hundreds of thousands of new chemical reactions are reported annually, yet translating them into actionable experiments becomes an obstacle1,2. Recent applications of large language models (LLMs) have shown promise3-6, but systems that reliably work for diverse transformations across de novo compounds have remained elusive. Here we introduce MOSAIC (Multiple Optimized Specialists for AI-assisted Chemical Prediction), a computational framework that enables chemists to make use of the collective knowledge of millions of reaction protocols. MOSAIC is built on the Llama-3.1-8B-Instruct architecture7, training 2,498 specialized chemical experts in Voronoi-clustered spaces. This approach delivers reproducible and executable experimental protocols with confidence metrics for complex syntheses. With an overall 71% success rate, experimental validation demonstrates the realizations of more than 35 new compounds, spanning pharmaceuticals, materials, agrochemicals and cosmetics. Notably, MOSAIC also enables the discovery of new reaction methodologies that are absent from the expert's training, a cornerstone for advancing chemical synthesis. This scalable model of partitioning vast domains into searchable expert regions enables a generalizable strategy for AI-assisted discovery wherever accelerating information growth outpaces efficient knowledge access and application.

Rumination, or perseverative negative self-referential thinking, is a hallmark of depression. In adults, a dynamic resting-state fMRI model of trait rumination was recently identified through predictive modelling. In adolescents, a development period during which rumination and depression increase, the neurobiological correlates of ruminative thinking are less clear. In the current preregistered study, we examine dynamic connectivity correlates of self-reported rumination in the largest sample of adolescents to date (n = 443, containing clinical and non-clinical individuals). Notably, the adult model failed to generalize to our sample. In addition, linear models trained on default-mode network (DMN) connectivity, as well as whole-brain connectome models, failed to generalize to held-out data. In an exploratory random forest analysis, we found significant prediction performance of a model where increased variability between DMN-cerebellum, DMN-dorsal attention network, and DMN-DMN connections was nominally associated with higher rumination. However, the model did not generalize to an external sample with lower rumination scores and a distinct scanner protocol. Our findings illustrate the difficulty of characterizing the neurodevelopment of risk factors for depression.