Mixture of Models for Verticals: Biotech & Drug Discovery

The LLM era taught everyone to reach for one big model. The next era may not work that way. In every serious industry, the systems that actually deliver are mixtures of specialized models: one that sees, one that simulates, one that forecasts, one that checks, with a reasoning layer wiring them together. Each is best-in-class at a single job, and the system is the product of how they're wired together. That's a mixture of models, and every vertical wires it differently.

Biotech & Drug Discovery
Drug discovery is where "models, plural" is already the norm. A target gets characterized, a molecule gets designed, its properties get predicted, and a reasoning layer ties the literature together. No single network does all four. We read the papers behind that pipeline and argue about where it breaks.
WHY ONE MODEL ISN'T ENOUGH
An LLM can summarize a paper, but it can't fold a protein, dock a ligand, or predict toxicity. Those are different model families with different inductive biases, and language is just the layer that plans and explains.
A POSSBILE MIXTURE OF MODEL IN BIOTECH:
Structure prediction: AlphaFold-class models for protein and complex structure
Generative chemistry: diffusion and flow models that design candidate molecules
Property & ADMET prediction: GNNs that score binding, solubility, toxicity
Reasoning layer: an LLM that reads the literature and orchestrates the loop
It's deliberately small event. Real discussion, no slides. Join us Pizza's on the house.
Hosted by The Deep-tech Community and SigmoLabs.