Original Authors: Wei Z, Ektefaie Y, Zhou A, et al.
Published: bioRxiv, March 12, 2026

Abstract

Fleming is an integrated AI agent for tuberculosis antibiotic discovery developed by Harvard University and other institutions. The system coordinates four specialized sub-agents through a central medicinal chemist agent, integrating 9 molecular AI models and 11 tools to achieve end-to-end workflows from molecule generation, activity prediction, ADMET optimization to literature reasoning.

Supported by a DMPNN inhibition prediction model trained on 114,900 compounds, Fleming achieved an 83% hit rate in prospective validation of 435 molecules. More importantly, all 6 de novo designed generated molecules showed Mtb inhibitory activity in wet-lab validation (100% hit rate), with 4 possessing pharmacokinetic properties suitable for hit-to-lead projects.

Key Metrics

  • Prospective validation hit rate: 83%
  • De novo design hit rate: 100% (6/6)
  • Enrichment improvement: 17-83x
  • Hit-to-lead candidates: 4/6 molecules

1. Background: Challenges in TB Antibiotic Discovery

Multidrug-resistant Mycobacterium tuberculosis (MDR-Mtb) caused approximately 160,000 deaths in 2022, accounting for over 13% of global antimicrobial resistance deaths. Mathematical models estimate that nearly 19 million people carry latent MDR-TB infections, at risk of progressing to active disease.

In this context, Fleming was developed—a collaboration between Harvard University, Broad Institute, FutureHouse, and Texas A&M to accelerate next-generation tuberculosis antibiotic discovery.

2. Technical Architecture and Methods

2.1 Agent Architecture

Fleming employs a hierarchical agent architecture with a central "medicinal chemist agent" coordinating four specialized sub-agents: Mtb growth inhibition agent, molecule generation agent, ADMET agent, and molecule optimization agent. Each sub-agent has direct access to 9 molecular AI models and 11 tools.

2.2 Core AI Models

The inhibition prediction model is based on Directed Message Passing Neural Networks (DMPNN), trained on 114,933 diverse compounds and fragments. Model AUROC improved from 0.76 to 0.79.

2.3 ADMET Prediction and Optimization

The ADMET agent integrates 37 independent DMPNN models. After optimization, the proportion predicted to fail more than 25% of ADMET tasks decreased significantly from 42.5% to 5.0%.

3. Wet-Lab Validation Results

3.1 Prospective Validation

The research team conducted prospective validation on 435 structurally diverse molecules. Wet-lab results showed that 5 of the predicted inhibitors indeed inhibited Mtb growth (83.3% hit rate), and 382 predicted non-inhibitors showed no inhibitory activity (89.0% accuracy). This achieved 17-83x enrichment improvement.

3.2 De Novo Design Validation

100% Hit Rate Milestone

All 6 molecules inhibited mc2-7000 Mtb strain growth in a dose-dependent manner, with EC50 ranging from 3.6 µM to 75.4 µM. The probability of this result occurring randomly is extremely low (assuming random hit rate 1-5%, probability < 10^-10 to 10^-6).

3.3 Toxicity and PK Assessment

Of the 6 candidate molecules, 5 showed no to low cytotoxicity. Considering moderate to high selectivity, good safety, and PK properties, 4 of the 6 generated molecules reached levels suitable for hit-to-lead projects.

4. Discussion

4.1 Technical Contribution

Fleming represents the first end-to-end validation of an agent framework in antibiotic discovery. The 100% generative design hit rate and 17-83x enrichment improvement demonstrate that integrated agent approaches can significantly accelerate preclinical lead identification.

4.2 Limitations

4.3 Comparison with Related Work

Compared to fragment-based methods like SyntheMol, Fleming's diffusion model explores broader chemical space with higher novelty. Compared to Latent-Y, Fleming focuses on small-molecule antibiotics, demonstrating agent framework value in different domains.

5. Conclusion

As a wet-lab-validated integrated antibiotic design agent, Fleming demonstrates end-to-end capabilities from chemical space exploration to preclinical candidate identification. Its 100% generative design hit rate and 17-83x enrichment improvement show that agent frameworks can effectively coordinate molecular AI models, optimization tools, and literature knowledge.

Fleming should be viewed as a complement to traditional drug discovery workflows rather than a replacement. Its true value lies in accelerating early lead identification, providing high-quality starting points for subsequent medicinal chemistry optimization and clinical development.

References

← Back to Blog