AI-driven computational platform for prioritizing herbicide-tolerant protein variants before synthesis

Overview

This is an integrated computational platform that prioritizes herbicide-tolerant enzyme variants before any wet-lab synthesis. It combines AI-based literature and kinetics mining, structure-aware protein language model embeddings, and physics-based scoring to rank thousands of candidate mutations by predicted stability, folding, and herbicide binding or turnover. The platform directly addresses the core bottleneck in variant discovery — blind synthesize-and-test cycles — by reducing large mutation libraries to a short, high-confidence set ready for experimental validation.

Technical specifications

Key components:

• CatLog: agentic literature and kinetics mining (27,000+ curated records) that identifies herbicide-target and detoxification enzymes and their kinetic parameters
• EnzymeShell/ShellChop: shell-resolved ESM-C protein language model embeddings combined with PyRosetta rotamer repacking to score thousands of mutations for stability and activity
• Seq2Bind: structure-free binding-affinity ranking that outperforms structure-based docking (HADDOCK) at recovering binding interfaces without requiring 3D structures
• AnionPDB: cofactor- and anion-pocket engineering for metalloprotein design
• openPTM: post-translational modification-aware stability and function scoring
• cyclome: enhanced-sampling molecular dynamics for conformational state analysis
• AlphaFold3: structural modeling of modified residues

Validation workflow:

• Top-ranked candidates are expressed via cell-free protein synthesis, with co-advised graduate students running the experiments
• Kinetic assays (LabChip) and hydrogen-deuterium exchange mass spectrometry (HDX-MS) confirm folding, stability, and herbicide tolerance or detoxification activity
• A closed feedback loop continuously recalibrates predictions against real experimental data

Technology readiness level

The platform is at TRL 3–4 (proof-of-concept). Phase 1 involves computational mining of herbicide-target literature, building structural models of active and binding sites, and scoring candidate mutation libraries to produce a ranked shortlist. Phase 2 validates top candidates through cell-free expression and biophysical assays. Phase 3 iterates predictions against experimental data in a closed loop, expanding validated hit families and delivering ranked, lab-confirmed variants along with open-source scoring tools to downstream purification and screening partners.