Fleming: An AI Agent for Antibiotic Discovery in Mycobacterium Tuberculosis

Fleming is an integrative AI agent that combines discriminative and generative models with molecular optimization and ADMET prediction to successfully identify and design novel lead compounds for *Mycobacterium tuberculosis* inhibition with high in vitro hit rates and favorable safety profiles.

The Gist

Imagine you are trying to find a specific key that can unlock a very stubborn, ancient, and heavily fortified door. This door is Tuberculosis (TB), a disease caused by a bacterium called Mycobacterium tuberculosis. For decades, scientists have been trying to make new keys (antibiotics) to open this door, but the process has been slow, expensive, and often unsuccessful. The "key-making factory" (traditional drug discovery) is like a giant warehouse where workers randomly grab keys from a shelf, hoping one fits. Most don't, and the ones that do are often rusty or break easily.

Enter Fleming, a new kind of "Super-Intelligent Architect" designed to solve this problem.

The Problem: The Needle in a Haystack

Finding a new TB drug is like looking for a needle in a haystack, but the haystack is the size of a galaxy, and the needle is invisible. Traditional methods test thousands of chemicals one by one, but they mostly test chemicals that look very similar to ones we've already tried. It's like trying to open a new lock by only using keys that look exactly like the old ones.

The Solution: Meet Fleming

Fleming isn't just a computer program; it's an AI Agent. Think of it as a highly skilled project manager who doesn't just guess, but actually thinks and collaborates.

Fleming is built like a high-tech command center with a central "Chief Architect" (the main AI) who leads a team of four specialized experts:

1. The Locksmith (Inhibition Agent): This expert knows exactly what the TB lock looks like. It has studied over 114,000 different chemical keys and learned which ones successfully jam the lock (kill the bacteria). It's incredibly good at spotting patterns that humans miss.
2. The Dreamer (Generative Agent): Instead of just picking keys from a shelf, this expert imagines brand new keys from scratch. It uses a "diffusion" process (like a sculptor slowly revealing a statue from a block of marble) to create entirely new chemical structures that have never existed before.
3. The Safety Inspector (ADMET Agent): A new key might fit the lock, but what if it's made of poison? This expert checks if the new key is safe for the human body. It asks: "Will this hurt the liver? Will it dissolve too fast in the stomach? Is it toxic?"
4. The Polisher (Optimization Agent): If a key is almost perfect but slightly too big or too heavy, this expert tweaks it. It makes small adjustments to ensure the key is not only effective but also easy to manufacture and safe to carry in your pocket (the body).

How Fleming Works: The "Super-Team" Approach

In the past, scientists used these tools separately. They might ask the Locksmith for a list, then hand that list to the Safety Inspector, who would reject 90% of them. Then they'd ask the Dreamer for new ideas, but the Dreamer didn't know what the Safety Inspector liked.

Fleming connects them all.
When you ask Fleming, "Design a new TB drug," the Chief Architect coordinates the team in real-time:

• The Dreamer sketches a new molecule.
• The Locksmith immediately checks: "Will this kill TB?"
• The Safety Inspector checks: "Is this safe for humans?"
• The Polisher tweaks the design to make it better.
• The Chief Architect even reads scientific books and papers (using a literature search tool) to make sure the design makes sense in the real world.

The Results: A Miracle in the Lab

The paper reports some stunning results that sound almost too good to be true:

• The Hit Rate: In traditional drug discovery, if you test 100 random chemicals, maybe 1 or 2 will work (a 1-5% success rate). Fleming predicted 435 new chemicals, and 83% of them actually worked in the lab. That's like finding 83 needles in a haystack instead of 1.
• The "De Novo" Success: When Fleming designed molecules completely from scratch (never seen before), 100% of the ones they tested worked.
• Safety: Not only did they work, but they were also safe. Most of the new designs passed strict safety checks for the liver and heart.
• Novelty: These weren't just copies of old drugs. They were structurally unique, meaning the bacteria likely won't be able to build a defense against them easily.

The Analogy of the "Smart Co-Pilot"

Think of Fleming as a co-pilot for a scientist.

• Old Way: The scientist is flying a plane blindfolded, throwing darts at a map, hoping to hit a target.
• Fleming Way: The scientist is now the captain, but they have a co-pilot (Fleming) who has a radar, a weather map, and a database of every flight path ever taken. The co-pilot says, "Captain, if we turn left here, we'll avoid the storm (toxicity) and hit the target (kill TB) with 90% certainty."

Why This Matters

This isn't just about one drug; it's about a new way of doing science.

• Speed: What used to take years of trial and error can now happen in days.
• Cost: It saves millions of dollars by stopping bad ideas before they are even built.
• Hope: With antibiotic resistance (superbugs) becoming a global crisis, tools like Fleming give us a fighting chance to stay ahead of the bacteria.

In short, Fleming is an AI that acts like a master chemist, a safety inspector, and a creative artist all rolled into one. It doesn't just guess; it reasons, optimizes, and designs, proving that when we combine human curiosity with artificial intelligence, we can solve some of the world's toughest health challenges.

Technical Summary

1. Problem Statement

The development of new antibiotics for Mycobacterium tuberculosis (Mtb), particularly for multidrug-resistant (MDR) strains, faces significant hurdles:

• High Failure Rates & Costs: Traditional high-throughput screening (HTS) is expensive, labor-intensive, and often yields low hit rates (typically 1–5%).
• Chemical Space Limitations: HTS libraries often sample redundant chemical spaces, missing novel scaffolds necessary to overcome intrinsic bacterial resistance mechanisms (e.g., the waxy cell envelope of Mtb).
• Model Generalizability: Existing AI models for molecular property prediction often fail when applied to structurally novel compounds (out-of-distribution) and lack calibrated uncertainty estimates.
• Fragmented Workflows: Drug discovery traditionally requires siloed expertise in medicinal chemistry, ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) profiling, and literature review, which is difficult to integrate into a single efficient pipeline.

2. Methodology: The Fleming Framework

Fleming is an integrative AI agent system designed to orchestrate a multi-objective drug discovery workflow. It utilizes a central "Medicinal Chemist" agent (powered by the o4-mini Large Language Model) that coordinates four specialized sub-agents and a suite of 9 molecular AI models and 11 tools.

Core Components:

1. Mtb Growth Inhibition Agent:

   • Model: A Directed Message Passing Neural Network (DMPNN) trained on a diverse dataset of 114,933 compounds and fragments.
   • Innovation: Uses an evidential framework (eDMPNN) with Dirichlet loss to quantify prediction uncertainty. This allows the system to filter out low-confidence predictions, improving reliability on novel structures.
   • Data: Includes 113,919 diverse drug-like molecules, 1,000 fragments, and 14 known TB antibiotics.

2. Generative Agent:

   • Models: Combines SyntheMol (a fragment-based generative model) and an Equivariant Conditional Diffusion Model.
   • Function: Generates de novo molecular structures conditioned on desired inhibition levels. The diffusion model specifically assembles atomic-level building blocks to maximize structural novelty.

3. ADMET Agent:

   • Models: Utilizes 37 distinct DMPNN models trained on the Therapeutics Data Commons (TDC) to predict absorption, distribution, metabolism, excretion, and toxicity (including hERG, liver injury, and carcinogenicity).
   • Function: Filters and ranks candidates based on drug-likeness and safety profiles.

4. Molecular Optimization Agent:

   • Tool: Uses SynSpace to generate chemical neighbors of initial designs.
   • Function: Iteratively refines molecules to improve ADMET properties while maintaining predicted inhibition potency.

5. Literature & Reasoning Engine:

   • Integrates PaperQA2 and web-search capabilities to retrieve scientific literature regarding specific substructures or mechanisms of action, grounding AI predictions in existing biological knowledge.

Operational Modes:

• Co-pilot Mode: Interactive, real-time assistance for chemists to design or analyze specific molecules.
• Batch Mode: Automated processing of large libraries to identify and prioritize top candidates for synthesis.

3. Key Contributions

• Integrated Agentic Workflow: First demonstration of a unified AI agent that seamlessly combines discriminative prediction, generative design, ADMET optimization, and literature reasoning for antibiotic discovery.
• Evidential Uncertainty Modeling: The application of evidential deep learning to Mtb inhibition prediction significantly improves generalizability to unseen chemical spaces compared to standard classifiers.
• Conditional Diffusion for Antibiotics: Successfully adapted 3D equivariant diffusion models to generate novel antibiotic candidates with high structural diversity.
• Human-in-the-Loop Validation: The system includes a plausibility check via LLM and independent review by human medicinal chemists before experimental validation.

4. Results

Model Performance:

• Inhibition Prediction: The eDMPNN achieved an AUROC of 0.76 (improving to 0.79 with uncertainty filtering) and an AUPRC of 0.20. It outperformed standard models (Random Forest, Transformers) and showed superior generalizability (AUSPC) on structurally distinct test sets.
• Differentiation: Fleming demonstrated 17% higher discrimination between known Mtb leads and leads for other diseases compared to a generic LLM agent, and 13% higher than molecular property prediction alone on challenging ADMET tasks.

Prospective Validation (Wet Lab):

• Hit Rate: In a prospective screen of 435 diverse molecules, Fleming predicted 6 inhibitors. 5 of 6 (83.3%) were confirmed as active against Mtb in vitro.
• De Novo Design: Fleming generated novel molecules using diffusion and SyntheMol. From these, 6 molecules were selected for synthesis based on plausibility and feasibility.
  • 100% Hit Rate: All 6 synthesized molecules (including 2 designed de novo and 4 modified analogs) inhibited Mtb growth in a dose-dependent manner.
  • Potency: EC50 values ranged from 3.6 µM to 75.4 µM.
  • Novelty: All 6 molecules had a Tanimoto similarity < 0.33 to the training set and existing antibiotic databases, confirming they are structurally novel.
  • Safety: 5 of the 6 candidates passed all of Lipinski's Rule of 5. They showed low cytotoxicity to HepG2, HEK293T, and HDF cell lines, with selectivity indices (SI) ranging from 2.1 to 10.8.
  • Pharmacokinetics: Assayed compounds showed robust metabolic stability (half-lives up to 1072 minutes) and favorable plasma protein binding.

Optimization Success:

• The molecular optimization agent reduced the proportion of molecules failing >25% of ADMET tasks from 42.5% to 5.0% without compromising predicted inhibition.

5. Significance

• Accelerated Discovery: Fleming achieved an 83% hit rate for de novo generative designs, a success rate orders of magnitude higher than traditional HTS (1–5%).
• Overcoming Resistance: By exploring underrepresented regions of chemical space and generating structurally novel scaffolds, the system addresses the critical need for new mechanisms of action against MDR-TB.
• Scalability & Accessibility: The framework is designed to be adaptable to other diseases by retraining specific component models. The use of open-source components and natural language interfaces suggests a path toward democratizing drug discovery for resource-limited settings.
• Paradigm Shift: This work validates the "AI Agent" paradigm in drug discovery, moving beyond simple prediction tools to autonomous systems capable of reasoning, planning, and executing complex multi-step discovery workflows.

In conclusion, Fleming represents a breakthrough in AI-driven antibiotic discovery, successfully translating computational designs into potent, novel, and safe anti-tubercular compounds, thereby offering a viable pathway to combat the global threat of drug-resistant tuberculosis.