How Generative AI Is Crafting New Antibiotics from Scratch

Background

Generative artificial intelligence can amplify biases present in training data, raising safety concerns. MIT bioengineering researchers applied generative AI to the opposite problem—designing antibacterial molecules.

Study Overview

The work, titled “A generative deep learning approach to de novo antibiotic design” and published in Cell , trained a deep‑learning model on a large curated dataset of known antibacterial compounds. The model was used to propose millions of novel structures, several of which showed activity in mouse infection models.

Methodology

The researchers employed two complementary generation strategies:

Fragment‑guided generation : A graph neural network (GNN) screened >45 million chemical fragments for predicted selective activity against Neisseria gonorrhoeae and Staphylococcus aureus . Selected fragments were expanded into full molecules using two generative models—a chemistry‑aware genetic algorithm (CReM) and a variational auto‑encoder (VAE).

De novo generation : The same CReM and VAE models were run without fragment constraints, allowing the AI to design entirely novel scaffolds based solely on learned chemical knowledge.

Generation Results

Combined, the CReM and VAE pipelines generated >36 million previously unrecorded compounds predicted to possess antibacterial activity. Most of these candidates were synthetically inaccessible, creating a major bottleneck for experimental validation.

Synthesis and Validation

After applying synthetic feasibility filters, 24 molecules were synthesized and screened in vitro. Seven exhibited measurable antibacterial activity; two—designated NG1 and DN1—showed high potency and selectivity, with mechanisms distinct from existing clinical antibiotics. Both compounds reduced bacterial load in mouse infection models.

Implications

This study demonstrates that generative AI can serve as a true molecular design engine, creating novel scaffolds that are absent from existing chemical libraries. By expanding the searchable chemical space, such approaches may accelerate the discovery of antibiotics with new mechanisms of action.

Code example

来源：ScienceAI
本文
约1000字
，建议阅读
5
分钟
该研究不仅能够助力抗菌化合物的研发，还能推动对庞大未知化学空间的探索。