Team name
NJTech-SynCAR
SynCARengineering a key enzyme for putrescine synthesisSemi-Rational Design and AI to Improve CAR Activity on GABA
School of Biotechnology and Pharmaceutical Engineering, Nanjing Tech UniversityT3 Biomanufacturing
Project title
SynCARengineering a key enzyme for putrescine synthesis
T3 BiomanufacturingCAR EngineeringActivity & Cross-Validation
NJTech-SynCAR first screens carboxylic acid reductase variants by semi-rational design, then trains a protein language model on mutation and activity data to propose the next candidates.
- Team name
- NJTech-SynCAR
- Institution
- School of Biotechnology and Pharmaceutical Engineering, Nanjing Tech University
Team media is for event showcases and reporting only. Portrait rights remain with the individuals. For concerns about portrait use, contact the organizers. We will investigate and take appropriate action.
Background
CAR in Biomanufacturing
Carboxylic acid reductases convert carboxylic acids to aldehydes. The project's route runs from L-glutamate through GABA to 1,4-diaminobutane, a monomer for polymers such as nylon-46 currently made mainly from petrochemicals. Improving the intermediate enzyme step is central to this biological route.
Research Approach
Experiments Before Model Training
Sequence analysis, structure prediction and mutation libraries first produce measured activity data. A protein language model then learns from mutation sites, substitutions and activities to suggest further designs. An experiment–learn–redesign loop links both stages; PSSM-normalized secondary-screen results provide the initial ranked variants.
- Target Enzyme
- MAB2962
- Mutation Sites
- 342、514、281、420
- Model Backbone
- ESM2: 35M, 150M and 650M Compared
Research Process
Development & Validation
- 01
Stage 1
Testing Sites and Combinations
Bar charts compare single mutations at position 342 and additional position-514 mutations on the D281P parent. Some variants show higher whole-cell catalytic activity than their corresponding parent. - 02
Stage 2
Model Selection and Split Sensitivity
The selected esm2_t33_650M_UR50D backbone has 33 layers and a 1,280-dimensional hidden state. Layers 0–21 are frozen: 22 layers, or 66.7%. Of 651.0M parameters, 218.1M are trainable. There are 105 training and 35 test examples. Four-fold Spearman correlations are 0.618, 0.705, 0.455 and 0.227, showing substantial sensitivity to the data split at this sample size.
Project Evidence
Methods & Figures
Teamwork
Computational, experimental and candidate review records
Project Stories
Project & Team Records
Structures, Contacts and Training Data
The interface places a 3D protein structure beside a residue-contact heatmap, with a table in another view. Preparation scripts parse mutation notation such as A123B into sites and substitutions, separate single and double mutants, and match activities. These steps determine what evidence enters the model.
Toward a Transferable CAR Platform
The next proposed stage combines general CAR mutations, knowledge transfer and the SynCAR platform. Expected outputs include improved variants, AI-guided design, greener diaminobutane production, transferable workflows and lower reaction costs. Transferability will require evidence on another design task, beyond improving one enzyme.
