SilicoXplore: An integrated cloud platform coupling machine learning with physics-based modelling for end-to-end drug discovery, applied to the computational prioritisation of putative New Delhi metallo-β-lactamase-1 binders.

Modern drug discovery joins machine learning with physics-based simulation, but building such a pipeline needs Linux administration, dependency management, format conversion and scripting, which keeps out many of the chemists and biologists who ask the questions. We describe SilicoXplore, a cloud-hosted platform of 31 interoperable modules covering structure preparation, five docking engines, de novo generation, machine-learning prediction, ADMET estimation, molecular dynamics, free energy calculation and density functional theory, all driven through guided forms. Third-party engines are wrapped under their own licences rather than rewritten. It was applied to New Delhi metallo-β-lactamase-1, generating 752,436 molecules from 455 seed actives and narrowing them to four. Docking was validated by redocking, at 0.840 Å, and by enrichment against 4247 property-matched decoys, at an area under the receiver operating characteristic curve of 0.714. The four molecules and the reference were re-docked with both catalytic Zn2+ ions retained, then carried into triplicate 100 ns molecular dynamics with MM-GBSA and MM-PBSA analysis. Redocking reproduced both the crystallographic pose of the reference and its metal contact, at 2.215 Å, and three of the four molecules contacted a zinc ion directly. Replicate spread in the end-point energies matched the differences between molecules, so no ranking is claimed. The classifier, which reached an area under the curve of 0.985 within its training chemotype space, retained 76.5% of the library, because enforcing novelty places it outside the region in which the model was assessed. Selection was carried by the physics-based stages. All results are computational and require experimental validation.