Generative AI for Biomolecular Design: Antibiotics, Proteins & Autonomous Labs

Revolutionizing Biomolecular Design with AI

• 💡 Generative AI is transforming biomolecular design, including drug discovery and protein structure prediction.
• 🎯 The goal is to design molecules with specific desired properties, such as antibacterial activity, antiviral effects, target binding, or controlled toxicity.
• 🧩 A typical biomolecular design system involves three components: a generative model for creating molecules, a scoring model for evaluating designs, and an experimental platform for validation.

Advancing Antibiotic Discovery

• 🦠 Generative models are crucial for combating antibiotic resistance, a growing global health crisis.
• 🔬 The junction tree VAE (JT-VAE) is introduced as a hierarchical generative model for small molecule design, representing molecules as graphs.
• ✅ This approach has successfully enabled lead optimization and de novo design of new antibiotics, validated in mice models.
• 📈 It addresses the challenges of the vast chemical space and the limitations of traditional, expensive experimental screening.

Enhancing Protein Binder Design

• 🧬 Scoring models are critical for evaluating designs in protein binder design, which is vital for therapeutics.
• 🧠 pTMEnergy is presented as a differentiable statistical energy function derived from AlphaFold’s confidence prediction module.
• 🚀 Integrating pTMEnergy into BindEnergyCraft significantly improves the in silico binder success rate compared to traditional methods like ipTM scores.
• 📊 This new scoring model provides a smoother gradient for optimization, leading to more effective binder design.

Towards Autonomous Closed-Loop Systems

• 🤖 The talk highlights the development of a fully autonomous design system that integrates generative models with automated physical labs.
• ⚖️ This system balances exploitation and exploration by quantifying uncertainty in predictions and assessing synthesizability of candidate molecules.
• 🧪 Demonstrated in molecular dye design, the system aims to optimize properties like absorption wavelength, solubility, and photostability over multiple experimental rounds.
• 🔄 This closed-loop approach accelerates experimental testing and continuously improves model performance by learning from new data.