Amin Tavakoli

We use the GenSLM protein language model (PLM) to perform sequence-conditioned generation of TrpB enzyme variants, addressing a major challenge in biocatalyst design: finding functional starting points for optimization. By combining generative modeling with computational filtering, the approach produces sequences that are stable, expressible, and catalytically active, some operating independently of their natural partner. Several AI-generated variants outperformed both natural and lab-optimized enzymes, showing that the model captures latent functional patterns and can propose non-trivial, high-performing sequences beyond evolutionary examples. This demonstrates that PLMs can act as powerful generative priors for functional protein engineering, enabling AI-driven exploration of functional sequence space with minimal experimental iteration.

While deep learning has advanced quantum chemistry, most models remain limited to neutral, closed-shell molecules. In contrast, real-world systems involve varying charges, spins, and environments. We present our model—a geometry- and physics-informed deep learning framework that incorporates spin-polarized orbital features and SE(3)-equivariant graph neural networks to represent arbitrary molecular systems. Our model accurately predicts properties of charged, open-shell, and solvated molecules, and generalizes to much larger systems than seen during training. It reaches chemical accuracy with 10× less data than competing models and offers a 1,000–10,000× speedup over DFT, thanks to its physics-grounded design.

DeepRXN is a specialized platform designed to advance the integration of deep learning into chemoinformatics, hosting predictive chemoinformatics software and public chemical reaction databases. Its unique focus on representing chemical reactions through elementary step reactions offers numerous advantages, with applications spanning reaction prediction, synthetic planning, atmospheric chemistry, drug design, and beyond. This innovative perspective holds the potential to reshape and elevate various aspects of chemical research and applications within a singular framework.

Machine learning models for protein engineering typically use sequence-based, structure-based, or combined representations rooted in the idea that sequence and structure inform function. While effective in capturing evolutionary patterns, these approaches often overlook protein dynamics. We introduce a dynamic-aware representation derived from unsupervised analysis of molecular dynamics simulations. By encoding temporal and spatial behaviors, our method captures key interactions within the protein chain, offering valuable insights for protein design.