Computer Science · Computational Theory and Mathematics
A study finds that large language models can recognize patterns in material categories but consistently struggle to use coordinate information, even as the models and datasets grow. For properties shaped mainly by crystal structure, specialized models that represent geometry performed better.
Researchers evaluated three AI systems on hundreds of molecules bound to a SARS-CoV-2 protein and in three prospective drug screens. The results suggest that combining AI-based predictions with traditional docking may help prioritize candidates, although no single approach worked best in every test.
A model trained on results from 73 laboratory studies linked chemical structures with changes in the transfer of the RP4 antibiotic-resistance plasmid. Quaternary ammonium and long-chain alkyl compounds were associated with reduced transfer, while sulfonamide- and amine-containing compounds were associated with increased transfer.
Researchers systematically reviewed 43 English-language studies on using machine learning and multiple real-world data sources to monitor possible side effects of mRNA vaccines. Some models showed good performance in particular datasets, but the studies differed substantially and direct applications to mRNA vaccine safety were often retrospective or limited. This review describes research methods and potential uses; it does not establish how well these tools work in routine healthcare.
Researchers developed Orbformer, an AI model that learns transferable descriptions of molecules from 22,000 structures, including molecules with breaking bonds. After fine-tuning, it reached chemical accuracy on established tests and more demanding bond-breaking and Diels–Alder reaction problems, while reducing the need to solve each new molecule from scratch.
