Keep pulling the thread on Joshua Meyer.
AI models for biology have advanced from predicting the structure of single proteins, as seen with models like AlphaFold, to predicting the structure of entire macromolecular complexes, including protein-protein, protein-small molecule, and protein-DNA/RNA interactions.
Joshua Meyer predicts a future capability where AI models can generate "zero-shot" drug candidates, where a molecule from the first designed batch is viable for clinical trials.
In one instance, a Chai Discovery AI model solved a drug discovery problem in two weeks that another company had spent $5-10 million working on.
A "third wave" of AI-bio companies, including Chai Discovery, has emerged in the last few years, characterized by leveraging recent advances from NLP and computer vision.
The FDA has indicated it will use AI to review regulatory documents submitted by pharmaceutical companies.
Chai Discovery's Chai-1 model is an "all-atom" model capable of predicting interactions across different therapeutic modalities, including small molecules and proteins.
Predicting the 3D structures of DNA and RNA interactions is currently more difficult for AI models than predicting protein and small molecule interactions, potentially due to a lack of training data.
Chai Discovery released its Chai-1 model, an "atomic level microscope" for predicting molecular interactions, as open source in late 2023.
Chai Discovery's Chai-1 is a structure prediction model, while Chai-2 is a generative model for de novo molecule design.
Chai Discovery's internal goal for its molecule design model was a 1% success rate for the year, a benchmark they considered transformative for the field.
To validate its results, Chai Discovery tested its Chai-2 model across 50 different biological targets, a scale approximately 10 times larger than typical AI biology validation projects.
Chai Discovery believed that the adoption of diffusion models, which can sample multiple distinct hypotheses instead of averaging them, would be a key driver for improving performance in protein structure prediction.