AI Fleming discovers promising tuberculosis drug leads

Antibiotic development is notoriously slow, expensive, and prone to failure, creating a pressing need for tools that can find better candidate molecules faster. Researchers led by Maha Farhat describe Fleming, an integrative artificial intelligence agent built to meet that need for Mycobacterium tuberculosis (Mtb). Fleming was designed to search novel chemical space and to pick out lead compounds that satisfy multiple desirable criteria at once. Instead of relying on a single prediction step, Fleming combines discriminative and generative AI models trained on experimental data, and layers on molecular optimization, ADMET prediction, and literature search capabilities so that any promising molecule is evaluated from several angles before advancing. The training set behind Fleming included 114,900 diverse compounds and fragments tied to in vitro growth inhibition, giving the system a broad empirical base for recognizing chemical features linked to Mtb activity. The result is an agent that does more than flag likely inhibitors: it proposes new molecules and weighs their drug-like properties, aiming to move efficient candidates toward preclinical lead identification.

Fleming’s approach pairs two types of AI models focused on Mtb inhibition: discriminative models that assess whether a compound is likely to stop bacterial growth and generative models that design new chemical structures. Both were trained on a set of 114,900 diverse compounds and fragments based on in vitro growth inhibition, and then integrated with molecular optimization routines, ADMET prediction, and literature search functions to form a single preclinical lead identification workflow. In head-to-head comparisons, Fleming achieved 17% higher discrimination between known Mtb leads and leads for other diseases than a generic LLM agent, and it showed 13% better discrimination than molecular property prediction alone on challenging ADMET tasks. In laboratory tests, Fleming delivered an 83% in vitro hit rate for predicted inhibition and a 100% hit rate for compounds produced by its de novo generative design. Its generated designs also demonstrated an 83% rate of favorable ADMET profiles, indicating that the system can both propose active molecules and anticipate many safety and pharmacokinetic concerns.

The work on Fleming suggests a practical route to accelerate early-stage antibiotic discovery for tuberculosis by combining prediction, design, and prioritization in one tool. By exploring new regions of chemical space and filtering candidates through ADMET and literature checks, Fleming aims to reduce the number of dead-end compounds that waste time and resources in the lab. The high in vitro hit rates and the strong performance on ADMET-related tasks imply that this integrated approach can find candidates that are not only active against Mtb but also more likely to have favorable drug-like properties. Because Fleming is built for preclinical lead identification, it does not replace laboratory testing but is intended to make those experiments more focused and efficient. If adopted, such agents could speed up which molecules enter costly development steps and help research teams prioritize compounds that stand a better chance of becoming effective, safe treatments for tuberculosis.