Alzheimer’s disease is characterized by the aggregation of the Aβ peptide into amyloid fibrils. According to the amyloid hypothesis, pharmacologically targeting Aβ aggregation could result in disease-modifying treatments. The identification of inhibitors of Aβ aggregation, however, is complicated by complex technical challenges, which typically restrict to tens of thousands the number of compounds that can be screened in experimental aggregation assays. Here, we report a computational route to increase by 4 orders of magnitude the number of screenable compounds. We achieve this result by developing an open source pipeline version of the Deep Docking protocol, and illustrate its application to the discovery of secondary nucleation inhibitors of Aβ aggregation from an ultra-large chemical library of over 539 million compounds. The pipeline was used to prioritize 35 candidate compounds for in vitro testing in Aβ aggregation assays. We found that 19 of these compounds inhibit Aβ aggregation (54% hit rate). The two most potent compounds showed potency better than adapalene, a previously reported potent inhibitor of Aβ aggregation. Consistent with the intended mechanism of action, these two compounds also proved to be high-affinity binders of Aβ fibrils with an equilibrium dissociation constant in the low nanomolar range in surface plasmon resonance experiments. These results provide evidence that structure-based docking methods based on deep learning represent a cost-effective and rapid strategy to identify potent hits for drug development targeting protein misfolding diseases.
Biomolecular function often depends on conformational ensembles, yet modern diffusion-based structure generators are biased toward the compact conformations prevalent in structural databases, limiting their ability to explore broad conformational landscapes. This work introduces metadiffusion, where an additional meta-energy biasing layer on top of diffusion steers pretrained biomolecular diffusion models through gradient-guided denoising. Without retraining, metadiffusion generates diverse conformational ensembles whose residue-level flexibility patterns closely match molecular dynamics simulations. The method supports three complementary modes: optimisation, steering to user-specified targets, and exploration via inter-sample repulsion. This approach enables controlled exploration of collective variables, enumeration of alternative binding poses across proteins, nucleic acids and ligands, and conformational ensemble generation consistent with SAXS and NMR chemical shifts. Metadiffusion thus provides a practical route to connect diffusion-based structure generation with ensemble-level, experimentally-restrained structural analysis.
Off-target binding remains a leading cause of drug attrition, yet no method exists for rapidly quantifying small-molecule promiscuity across the human proteome. Here, we define promiscuity as the mean predicted binding affinity over 15,405 human proteins and derive a specificity score combining target affinity with this proteome-wide distribution. To make this assessment tractable at scale, we introduce OmniBind, a message-passing neural network that predicts promiscuity directly from a SMILES string at about a thousand compounds per second, several orders of magnitude faster than proteome-wide profiling. OmniBind promiscuity scores correlate with experimental binding data near assay reproducibility limits. Ranking candidates by specificity rather than affinity alone improves enrichment of approved drugs across all thresholds tested, an advantage robust to the choice of affinity predictor. OmniBind fills an unoccupied niche in the early-stage screening landscape as a fast, proteome-scale complement to traditional safety panels, with accuracy that will scale as the underlying affinity predictors continue to improve.