AI-Driven Drug Discovery to Address Antimicrobial Resistance and Glioblastoma: Stoked Bio's MOSAIC Platform Approach

In a recent webinar, Jon Stokes and Jeff Skinner of Stoked Bio presented their MOSAIC platform, an AI-driven approach designed to address two of the hardest problems in medicine: antimicrobial resistance (AMR) and recurrent glioblastoma (GBM). Their strategy is rooted in biology-first discovery, a focus on synthesizable chemistry, and an iterative cycle where data from the lab continually refines the machine learning models.

The Challenge: Resistance & Unmet Need

Both AMR and GBM present relentless challenges due to their ability to evolve against existing therapies. Every new antibiotic has a finite lifespan as bacteria adapt, and in oncology, relapse is nearly inevitable as tumors change during treatment. Today, an estimated 1.5 million people die each year from infections that were once easily treated, a number projected to rise to 10 million annually by 2050 if new drugs are not developed. Despite this urgency, incentives for antibiotic development remain limited because treatments are short, cures are quick, and recouping billion-dollar R&D investments is difficult.

Strategy: Biology-First + AI Design

1. Phenotype as the Anchor

• Instead of targeting a specific protein, they aim to kill microbial or cancer cells in a disease-relevant context.
• Simultaneously test compounds on human cell lines to reject broadly toxic “anti-life” molecules.

2. Generate in Synthesizable Space

• Use fragment-based generative design rather than atom-level synthesis.
• Leverage ~150,000 building blocks + ~50 allowable reactions → ~46 billion candidate molecules that are synthetically tractable in 1–2 steps.
• Volume is high but quality is constrained to feasible chemistry.

3. Test & Learn Loop

• Top candidates flow into synthesis and biological testing.
• Results feed back into the model to improve predictions for potency, selectivity, ADMET, and brain–penetration filters (for GBM).
• The loop accelerates and tightens optimization cycles.

Key Engineering Decisions & Trade-offs

The team emphasizes that success depends less on novel algorithms than on robust data. High-quality, biologically relevant phenotypic data provides the foundation for all modeling. By restricting generative design to tractable chemistry, they ensure that most AI-proposed molecules are actually makeable, saving time and resources. Working under resource constraints has also become an advantage. Limited budgets force sharper decision-making and encourage workflows designed to maximize information gained per experiment.

Operational Best Practices

To manage a growing pipeline of projects and partnerships, Stoked Bio relies on CDD Vault as a centralized system of record. The shift away from ad-hoc notes and fragmented files has improved reproducibility and collaboration. Equally important is the culture they have built. The team operates with low ego and high transparency, encouraging direct feedback while maintaining trust. This openness allows the science to move faster and ensures continuity between the academic lab where discoveries begin and the company where they advance toward the clinic.

Future Milestones & Outlook

Looking ahead, the near-term milestone is securing a pharma partnership built on preclinical proof-of-concept data packages. The mid-term goal is achieving first-in-human dosing and early safety and efficacy signals. Long-term, the ambition is to reach full approval for programs guided by the MOSAIC pipeline. At each stage, the team hopes to compress timelines and reduce costs, aiming first for “twice as fast at half the cost” and then iterating beyond.