AI meets wet lab: how collaboration is delivering new antibiotics
Researchers are pairing computational models with new experimental platforms to tackle disease, energy demands, and climate change. That mix is driving progress in engineered cells and new therapeutics - including antibiotics that target drug-resistant pathogens.
James J. Collins, a founder of synthetic biology and a leader in systems biology, has helped spark this shift. His work spans diagnostics and therapeutics for Ebola, Zika, SARS-CoV-2, and antibiotic-resistant bacteria, and it consistently centers on one theme: collaboration that fuses AI predictions with rigorous biological testing.
Why collaboration is the force multiplier
At the MIT Jameel Clinic, Collins teamed up with Regina Barzilay and Tommi Jaakkola to apply deep learning to antibiotic discovery. That effort led to halicin, a new antibiotic active against a broad set of multidrug-resistant bacteria, published in Cell in 2020. It's a clear example of AI, network biology, and systems microbiology doing better work together than apart.
With Donald Ingber at the Wyss Institute, Collins' group uses organs-on-chips to test AI-discovered and AI-generated antibiotics in tissue-like environments. These platforms complement animal studies and help clarify therapeutic potential before costly development.
- Cell: Deep learning approach to antibiotic discovery (2020)
- Wyss Institute: Organs-on-Chips overview
Generative AI for first-in-class molecules
In 2025, Collins' lab reported a generative AI pipeline that designs antibiotics from scratch. Using genetic algorithms and variational autoencoders, they generated millions of candidate molecules across fragment-based designs and unconstrained chemical space.
After computational filtering, retrosynthetic modeling, and medicinal chemistry review, the team synthesized 24 compounds and tested them. Seven showed selective antibacterial activity. One lead, NG1, was highly narrow-spectrum against multidrug-resistant Neisseria gonorrhoeae, sparing commensals. Another, DN1, targeted MRSA and cleared infections in mice via broad membrane disruption. Both were non-toxic with low resistance emergence.
From in silico promise to drug-like reality
The group is now optimizing for drug-like properties early - absorption, stability, safety - not just activity. By pairing AI design with high-throughput biology, they're compressing iteration cycles and raising the odds that candidates survive real-world development.
Translational engine: Phare Bio and the Antibiotics-AI Project
Phare Bio, a nonprofit co-founded by Collins, advances top AI-derived antibiotics from MIT's Antibiotics-AI Project toward the clinic. The model is collaborative by design: work with biotech, pharma, AI companies, philanthropies, nonprofits, and even nation states to bridge discovery and development.
Backed by an ARPA-H grant, the team is designing 15 new antibiotics and maturing them as preclinical candidates. The focus is a reliable pipeline - from computational design to experimental validation - that can respond faster to antibiotic resistance and deliver therapies to patients sooner.
What researchers can apply now
- Build cross-functional teams early: Pair ML experts with microbiologists, chemists, and clinicians. Align on objective functions that reflect activity, selectivity, toxicity, and developability.
- Run closed-loop discovery: Generate → filter → synthesize → test → learn. Feed biological results back into your models every cycle.
- Use physiologically relevant assays: Layer organs-on-chips or similar systems alongside standard in vitro and animal models to reduce false positives.
- Plan for synthesis and scale: Include retrosynthetic feasibility and med-chem review during in silico triage, not after.
- Target spectrum with intent: Decide where narrow-spectrum (e.g., NG1 for N. gonorrhoeae) or broader action (e.g., DN1 for MRSA) is clinically advantageous.
- Track resistance from day one: Measure resistance rates and mechanisms during hit validation, and design around liabilities.
- Set translational checkpoints: Define go/no-go gates for PK, safety, and manufacturability to avoid sunk-cost drift.
- Form translational partnerships early: Nonprofits like Phare Bio can de-risk the path to the clinic and attract the right downstream partners.
Outlook
The playbook is clear: combine smarter models with smarter experiments, and move promising molecules through a disciplined pipeline. As teams adopt this approach, the field can shift from reacting to resistance to anticipating it - and design antibiotics prepared for real clinical use.
If your lab is upskilling in AI for model building, screening, or data analysis, explore focused options by role here: Complete AI Training: Courses by Job.
Your membership also unlocks:
