Wei, Z.; Ektefaie, Y.; Zhou, A.; Negatu, D.; Aldridge, B. B.; Dick, T. B.; Skarlinski, M.; White, A.; Rodriques, S. G.; Hosseiniporgham, S.; Inna, K. V.; Sacchettini, J.; Zitnik, M.; Farhat, M. R.

Tuberculosis (TB) remains a critical public health challenge, causing over 1.5 million deaths annually. Rising resistance to existing TB antibiotics threatens to reverse decades of progress in combating the disease. Developing new antibiotics is challenging due to high development costs and failure rates. Artificial intelligence offers a promising solution to streamline early drug discovery by predicting inhibitory properties of novel compounds, generating molecules with desired properties, and automating workflows through large language model agents. In this work, we introduce Fleming, an AI agent designed for TB antibiotic discovery. We curated and generated the largest dataset of TB inhibitors to date (n=114,933), and used these data to train both discriminative and generative models for TB inhibitor identification. Fleming orchestrates four specialized agents-a bacterial inhibition prediction agent, a molecular generation agent, a molecular optimization agent, and an ADMET agent-to perform key tasks in early TB antibiotic discovery. These agents incorporate literature search capabilities and chemical optimization tools, creating an integrated platform for TB antibiotic discovery. Using experimental data and human medicinal chemist review, we demonstrate how Fleming enhances the antibiotic discovery process, and mirrors the decision-making of medicinal chemists through an intuitive natural language interface.