Accelerating ligand screening for tuberculosis: predicting binding energies with a 3D deep neural network.

The computational burden of molecular docking becomes prohibitive when receptor flexibility is explicitly considered, despite the accuracy gained in virtual screening workflows. To mitigate this limitation, we introduce a three-dimensional deep neural network designed to estimate binding energies before exhaustive docking across the complete fully flexible receptor ensemble. We target the InhA enzyme, a validated target of Mycobacterium tuberculosis, through the use of a fully flexible receptor model (with 20,000 poses) generated by computational simulation. The model leverages volumetric representations of electrostatic charge distributions in the InhA binding pocket after positioning a candidate ligand. Training was conducted with a curated collection of 60 ZINC compounds previously ranked as promising ligands. The approach reproduces docking trends with a Pearson correlation close to 0.70 and a Spearman correlation above 0.66, and processes more than 150,000 receptor-ligand conformations in approximately 25 min. These results highlight the feasibility of employing a deep neural network model as a pre-filtering mechanism within a curated ligand set and a specific fully flexible receptor workflow, aiming to reduce the number of docking simulations required, thereby accelerating rational drug discovery against tuberculosis.