AI-built molecule library aims to speed tuberculosis drug discovery

Tuberculosis remains a global health crisis, with more than 10 million new cases and 1.25 million deaths reported in 2023, while treatments have changed little in four decades and require long courses. In response, a team led by Jinjiang Guo created GenVS-TBDB, a target-aware database of small molecules designed specifically against Mycobacterium tuberculosis (M.tb) essential proteins. To build the resource, the researchers first integrated multiple data sources to identify 460 probable small-molecule binding pockets across 377 essential M.tb proteins. They then used a target-aware molecule generative model to produce over 1.2 million novel small molecules tailored to those pockets. For each generated compound the team computed key physicochemical properties to check medicinal chemistry tractability. Those compounds were evaluated further with molecular docking and an anti-TB specific graph neural network model to rank binding propensity and to prioritize likely whole-cell activity. The result is a large, searchable library of AI-designed and computationally screened candidate molecules aimed at expanding the chemical space available for TB drug discovery.

The GenVS-TBDB workflow combined computational filtering and experimental follow-up to test whether the AI-driven predictions could point to real bioactive molecules. After virtual screening and property calculations, the team selected 30 compounds for practical testing: 22 AI-designed molecules synthesized de novo and 8 commercially available analogs. Validation experiments showed that 2 synthesized compounds produced significant thermal stabilization of FtsZ, confirming target engagement with that protein. In whole-cell testing, 6 compounds exhibited cellular inhibition below 50 µM, with the most potent showing activity at 12 µM. One compound, GDI-11785, showed binding to the cell wall biosynthesis pathway and had 35 µM cellular activity, highlighting a promising starting point for pathway-focused optimization. The computational steps used — including molecular docking and the anti-TB specific graph neural network model — provided binding propensity ranking and whole-cell activity prioritization for each target. The GenVS-TBDB small library is publicly accessible for download at https://datascience.ghddi.org/database/view.

The work behind GenVS-TBDB demonstrates how target-aware AI generation, combined with virtual screening and focused lab validation, can expand the set of molecules scientists can try against M.tb. By producing more than a million tailored candidates and narrowing them with modeled physicochemical properties, docking and machine learning, the project creates a bank of starting points that could shorten the initial, exploratory phase of tuberculosis drug discovery. The experimental finding that some AI-designed compounds engage FtsZ and that GDI-11785 affects the cell wall biosynthesis pathway gives proof-of-concept that the computational predictions can translate into measurable biological effects. Making the library freely available aims to let medicinal chemists and microbiologists worldwide test and optimize these hits, potentially feeding new series into later-stage development. While much work remains to turn hits into safe, effective medicines, GenVS-TBDB provides a practical, open resource to diversify and accelerate early TB drug research.