Generative artificial intelligence model can easily design billions of novel antibiotics

[bnew]

Generative artificial intelligence model can easily design billions of novel antibiotics

Researchers at McMaster University and Stanford University have invented a new generative artificial intelligence model which can design billions of new antibiotic molecules that are inexpensive and easy to build in the laboratory.

www.news-medical.net

Generative artificial intelligence model can easily design billions of novel antibiotics​

• Download PDF Copy
  

This article has been provided by McMaster University and subjected to News-Medical.Net's review protocols, complying with its guidelines. To guarantee the article's authority, our editing team has highlighted the following features: verified accuracy, undergone scholarly review, sourced from a reliable authority, and meticulously scrutinized for errors. Modifications may be made to the article's style and length.

" style="box-sizing: border-box; background-color: transparent; color: rgb(95, 95, 95); text-decoration: none; transition: all 0.15s ease-out 0s; cursor: pointer; border: 0px; overflow-wrap: break-word;">Reviewed

Mar 22 2024 McMaster University

Researchers at McMaster University and Stanford University have invented a new generative artificial intelligence model which can design billions of new antibiotic molecules that are inexpensive and easy to build in the laboratory.

The worldwide spread of drug-resistant bacteria has created an urgent need for new antibiotics, but even modern AI methods are limited at isolating promising chemical compounds, especially when researchers must also find ways to manufacture these new AI-guided drugs and test them in the lab.

In a new study, published today in the journal Nature Machine Intelligence, researchers report they have developed a new generative AI model called SyntheMol, which can design new antibiotics to stop the spread of Acinetobacter baumannii, which the World Health Organization has identified as one of the world's most dangerous antibiotic-resistant bacteria.

Notoriously difficult to eradicate, A. baumannii can cause pneumonia, meningitis and infect wounds, all of which can lead to death. Researchers say few treatment options remain.

"Antibiotics are a unique medicine. As soon as we begin to employ them in the clinic, we're starting a timer before the drugs become ineffective, because bacteria evolve quickly to resist them," says Jonathan Stokes, lead author on the paper and an assistant professor in McMaster's Department of Biomedicine & Biochemistry, who conducted the work with James Zou, an associate professor of biomedical data science at Stanford University.

"We need a robust pipeline of antibiotics and we need to discover them quickly and inexpensively. That's where the artificial intelligence plays a crucial role," he says.

Researchers developed the generative model to access tens of billions of promising molecules quickly and cheaply.

They drew from a library of 132,000 molecular fragments, which fit together like Lego pieces but are all very different in nature. They then cross-referenced these molecular fragments with a set of 13 chemical reactions, enabling them to identify 30 billion two-way combinations of fragments to design new molecules with the most promising antibacterial properties.

Each of the molecules designed by this model was in turn fed through another AI model trained to predict toxicity. The process yielded six molecules which display potent antibacterial activity against A. baumannii and are also non-toxic.

Synthemol not only designs novel molecules that are promising drug candidates, but it also generates the recipe for how to make each new molecule. Generating such recipes is a new approach and a game changer because chemists do not know how to make AI-designed molecules."

James Zou, co-author, associate professor of biomedical data science at Stanford University

The research is funded in part by the Weston Family Foundation, the Canadian Institutes of Health Research, and Marnix and Mary Heersink.

Source:

McMaster University

Journal reference:

Swanson, K., et al. (2024). Generative AI for designing and validating easily synthesizable and structurally novel antibiotics. Nature Machine Intelligence. doi.org/10.1038/s42256-024-00809-7.

---

SyntheMol/README.md at main · mivanovitch/SyntheMol

Combinatorial antibiotic generation. Contribute to mivanovitch/SyntheMol development by creating an account on GitHub.

github.com

 

[bnew]

Chai Discovery Team Releases Chai-2: AI Model Achieves 16% Hit Rate in De Novo Antibody Design

Chai-2 achieves 16% hit rate in zero-shot antibody design, revolutionizing AI-driven molecular discovery and protein engineering

www.marktechpost.com

Chai Discovery Team Releases Chai-2: AI Model Achieves 16% Hit Rate in De Novo Antibody Design​

By Asif Razzaq

July 5, 2025

TLDR: Chai Discovery Team introduces Chai-2, a multimodal AI model that enables zero-shot de novo antibody design. Achieving a 16% hit rate across 52 novel targets using ≤20 candidates per target, Chai-2 outperforms prior methods by over 100x and delivers validated binders in under two weeks—eliminating the need for large-scale screening.

In a significant advancement for computational drug discovery, the Chai Discovery Team has introduced Chai-2 , a multimodal generative AI platform capable of zero-shot antibody and protein binder design. Unlike previous approaches that rely on extensive high-throughput screening, Chai-2 reliably designs functional binders in a single 24-well plate setup, achieving more than 100-fold improvement over existing state-of-the-art (SOTA) methods.

Chai-2 was tested on 52 novel targets , none of which had known antibody or nanobody binders in the Protein Data Bank (PDB). Despite this challenge, the system achieved a 16% experimental hit rate , discovering binders for 50% of the tested targets within a two-week cycle from computational design to wet-lab validation. This performance marks a shift from probabilistic screening to deterministic generation in molecular engineering.

AI-Powered De Novo Design at Experimental Scale​

Chai-2 integrates an all-atom generative design module and a folding model that predicts antibody-antigen complex structures with double the accuracy of its predecessor, Chai-1. The system operates in a zero-shot setting , generating sequences for antibody modalities like scFvs and VHHs without requiring prior binders.

Key features of Chai-2 include:

• No target-specific tuning required
• Ability to prompt designs using epitope-level constraints
• Generation of therapeutically relevant formats (miniproteins, scFvs, VHHs)
• Support for cross-reactivity design between species (e.g., human and cyno)

This approach allows researchers to design ≤20 antibodies or nanobodies per target and bypass the need for high-throughput screening altogether.

Benchmarking Across Diverse Protein Targets​

In rigorous lab validations, Chai-2 was applied to targets with no sequence or structure similarity to known antibodies . Designs were synthesized and tested using bio-layer interferometry (BLI) for binding. Results show:

• 15.5% average hit rate across all formats
• 20.0% for VHHs ,13.7% for scFvs
• Successful binders for 26 out of 52 targets

Notably, Chai-2 produced hits for hard targets such as TNFα , which has historically been intractable for in silico design. Many binders showed picomolar to low-nanomolar dissociation constants (KDs) , indicating high-affinity interactions.

Novelty, Diversity, and Specificity​

Chai-2’s outputs are structurally and sequentially distinct from known antibodies. Structural analysis showed:

• No generated design had <2Å RMSD from any known structure
• All CDR sequences had >10 edit distance from the closest known antibody
• Binders fell into multiple structural clusters per target, suggesting conformational diversity

Additional evaluations confirmed low off-target binding and comparable polyreactivity profiles to clinical antibodies like Trastuzumab and Ixekizumab.

Design Flexibility and Customization​

Beyond general-purpose binder generation, Chai-2 demonstrates the ability to:

• Target multiple epitopes on a single protein
• Produce binders across different antibody formats (e.g., scFv, VHH)
• Generate cross-species reactive antibodies in one prompt

In a cross-reactivity case study, a Chai-2 designed antibody achieved nanomolar KDs against both human and cyno variants of a protein, demonstrating its utility for preclinical studies and therapeutic development .

Implications for Drug Discovery​

Chai-2 effectively compresses the traditional biologics discovery timeline from months to weeks , delivering experimentally validated leads in a single round. Its combination of high success rate, design novelty, and modular prompting marks a paradigm shift in therapeutic discovery workflows.

The framework can be extended beyond antibodies to miniproteins, macrocycles, enzymes , and potentially small molecules , paving the way for computational-first design paradigms . Future directions include expanding into bispecifics, ADCs , and exploring biophysical property optimization (e.g., viscosity, aggregation).

As the field of AI in molecular design matures, Chai-2 sets a new bar for what can be achieved with generative models in real-world drug discovery settings.

---

Check out the Technical Report. All credit for this research goes to the researchers of this project.