Nat. Commun. | 准确刻画化学键断裂的从头算波函数基础模型

TL;DR - A Nature Communications study introduces Orbformer, a transferable deep-QMC foundation model for ab initio electronic wavefunctions. By reusing learned electronic-structure patterns across molecules, it accurately models bond breaking and other strongly correlated processes at substantially lower computational cost.

• Orbformer was pretrained without external energy labels on 22,350 equilibrium and nonequilibrium molecular geometries, then jointly fine-tuned across related structures or reaction pathways.
• On five bond-dissociation curves, it matched or exceeded the accuracy–cost Pareto frontier of conventional quantum-chemistry methods and converged systematically toward roughly 1 kcal/mol chemical accuracy.
• Joint fine-tuning yielded about a 20× efficiency gain; pretraining provided further speedups, especially when the target chemistry resembled the pretraining distribution.
• The model generalized from systems with at most 24 electrons during pretraining to systems with up to 106 electrons and learned physically meaningful localized orbitals and reusable local electronic-structure patterns.

TL;DR - A Nature Communications study presents an AI-guided pipeline for designing and experimentally optimizing minibinders against cancer cell-surface proteins. It shows that practical performance depends not only on predicted binding affinity but also on target-specific designability and non-interface biochemical properties such as isoelectric point.

• RFdiffusion, ProteinMPNN, and structure predictors were combined with cell-surface display, FACS, and sequencing to screen minibinders targeting PD-L1, CD276, and VTCN1.
• Design success was strongly target-dependent: PD-L1 yielded multiple validated binders, including a 2.12 nM candidate, whereas CD276 and especially VTCN1 were more difficult.
• Chai-1 ESM ipTM scores correlated better with experimental binding than the other evaluated structure scores and identified many disruptive interface mutations.
• Minibinders worked as PD-L1 detection reagents and CAR recognition domains, while optimizing non-interface residues toward a moderate pI improved CAR surface expression, activity, and target selectivity.

TL;DR - A Nature Communications study combined patient-derived iPSC neural cultures, high-content drug screening, machine learning, transcriptomics, and electrophysiology to identify potential repurposed treatments for MPS IIIA childhood dementia. It found nine neuroprotective candidates, but the evidence remains preclinical and in vitro.

• Models derived from five patients reproduced heparan sulfate accumulation, lysosomal abnormalities, neuronal death, reactive astrocytosis, and dysfunctional neural-network activity.
• XGBoost and CNN models classified drug-treated cells as “healthy-like” or “MPS IIIA-like,” largely agreeing with conventional phenotypic screening of 63 compounds.
• Nine drugs improved disease-associated phenotypes across multiple analyses, including acetyl-D-leucine, allopurinol, deferoxamine, hydroxychloroquine, lithium, probenecid, and theophylline.
• Allopurinol–probenecid and four-drug combination treatments produced neural-network improvements that persisted for two weeks after withdrawal, warranting further dose, safety, blood-brain-barrier, and in vivo studies.