This talk will focus on understanding IP rights, inventorship, and ownership of AI/ML-based inventions including platform aspects and therapeutic modalities. Deciding between patenting and maintaining trade secret aspects of the platform and modalities will be discussed. It will review life cycle management strategy for a biologics therapeutic program and Freedom to Operate.

Protein language models trained on evolutionarily diverse sequences implicitly model the sequence-function landscape of proteins. These models learn to generate diverse sequences, but must be steered for protein design tasks. We first discuss the generation of diverse CRISPR-Cas effectors for genome editing applications through fine-tuning on curated natural sequences. Next, we present a strategy for steering protein language models through conditioning on structural and functional context.

Snakebite envenoming remains a devastating and neglected tropical disease, claiming over 100,000 lives annually and causing severe complications and long-lasting disabilities for many more. Three-finger toxins (3FTx) are highly toxic components of elapid snake venoms that can cause diverse pathologies, including severe tissue damage and inhibition of nicotinic acetylcholine receptors (nAChRs) resulting in life-threatening neurotoxicity. Currently, the only available treatments for snakebite consist of polyclonal antibodies derived from the plasma of immunized animals, which have high cost and limited efficacy against 3FTxs. Here, we use deep learning methods to de novo design proteins to bind short- and long-chain α-neurotoxins and cytotoxins from the 3FTx family. With limited experimental screening, we obtain protein designs with remarkable thermal stability, high binding affinity, and near-atomic level agreement with the computational models. The designed proteins effectively neutralize all three 3FTx sub-families in vitro and protect mice from a lethal neurotoxin challenge. Such potent, stable, and readily manufacturable toxin-neutralizing proteins could provide the basis for safer, cost-effective, and widely accessible next-generation antivenom therapeutics.

Methods from artificial intelligence can now “write” proteins de novo, without starting from proteins found in nature. I will discuss our recent progress with developing deep learning models for de novo protein design, demonstrating that they generalize beyond the training space, and applying them to difficult problems, including atomically accurate design of dynamic proteins. Exciting frontiers lie in constructing synthetic cellular signaling from the ground up using de novo proteins.

Coarse-grained (CG) models are an alternative to the expensive all-atom models, but reaching high predictive power has been a longstanding challenge. By combining deep learning methods with a diverse training set of protein simulations, we have developed a CG force field which can be used for molecular dynamics on new sequences not used during model parametrization. This showcases the feasibility of a universal and efficient CG model for proteins.

Proteins are essential for cellular function, and missense variants can cause genetic disorders by destabilizing proteins or disrupting key interactions. While prediction of deleterious variants has progressed, understanding of the molecular mechanisms behind these variants remains limited. Thanks to advances in sequence- and structure-based computational predictors, we can now unravel the molecular mechanisms behind loss-of-function and quantify the role of stability in disrupting protein function.

• Why benchmarking is needed
• Designed competitions, and accidental ones
• Lessons from CACHE  
• The AIntibody competition to assess computational methods in antibody discovery ​

• What types of protein design problems are generative AI models useful for?
• How should we leverage biophysical principles when using generative AI?
• When should we use sequence- vs. structure-based models?
• What types of new data would be most useful for generative AI design platforms?

• Motivation: The urgent need for novel biologics is driving the exploration of AI in drug discovery
• Focus: AI's potential in accelerating biologic drug discovery, particularly de novo antibody design
• Showcase: Successful AI-driven VHH and mAb designs
• Discussion: AI's strengths in predicting antibody structures, challenges in translating designs into functional molecules, achieving industrial-scale reliability, and closing the gap between computational and experimental results

AI has become increasingly powerful but can be overhyped. Our lab has used AI methods to develop antibody language models, antibody structure prediction models, protein-protein docking models, and antibody design models. I will share recent results, including testing language models’ comprehension of biological antibody maturation processes, benchmarking antibody developability models, and bringing physical energies back into AI predictions. Our results suggest how to use AI tools with appropriate caution.

We present JAM, a generative protein design system that enables fully computational design of antibodies with therapeutic-grade properties for the first time. JAM generates antibodies that achieve double-digit nanomolar affinities, strong early-stage developability profiles, and precise targeting of functional epitopes without experimental optimization. We demonstrate JAM's capabilities across multiple therapeutic contexts, including the first fully computationally designed antibodies to multipass membrane proteins - Claudin-4 and CXCR7.