Project title
Yeast promoter designAI-driven closed-loop optimization
T4 Basic ScienceY. lipolyticaFluorescence & Validation
ProMeta trains four promoter-strength predictors and compares single models, stacking and simple averaging through internal cross-validation and two external datasets.
- Team name
- ProMeta
- Institution
- School of Bioengineering, Dalian University of Technology
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.
Team Motto
“Reading Sequences, Designing Gene Control ”
Background
Expression Levels That Can Be Chosen
Promoters regulate both the timing and strength of gene expression. Metabolic engineering needs balanced expression across pathway enzymes. The goal is a practical set of sequence-diverse regulatory elements with clear strength levels, supporting choice and reducing instability from repeated homologous sequences.
Research Approach
Four Predictors, One Internal Benchmark
Five-fold out-of-fold predictions on the same core dataset yield Pearson R values of 0.5290 for the CNN ensemble, 0.5370 for LSTM/GRU, 0.5030 for DNABERT2-W29 and 0.5658 for stacking. The stacking learner uses only base-model OOF predictions to avoid fitted-value leakage. External tests are still needed to assess generalization.
- Chassis
- Y. lipolytica
- External Datasets
- n=82、n=85
- Generator
- WGAN
Research Process
Development & Validation
- 01
Stage 1
External Model Comparisons
On Literature-82, the Base3 mean, CNN and DNABERT each achieve Spearman ρ = 0.60, versus 0.57 for final stacking. Base3 versus stacking yields Kendall τ of 0.43 versus 0.41 and discordant-pair rates of 0.29 versus 0.30. On Wang2023-rescued85, their Spearman values are 0.46 and 0.45. Neither external dataset shows stacking outperforming simple averaging. - 02
Stage 2
From 3,360 Sequences to 42
WGAN generates candidates at target strengths. Screening retains 3,222 valid sequences for frozen-model scoring, then 700 Top-100, 210 Top-30 and 42 Top-6 candidates across targets. Frozen scorers reduce the risk of generator–scorer co-adaptation. Further plots compare target versus predicted strength and GC content under balanced_only and strict_safe filters. - 03
Stage 3
Connecting Design and Experiments
Work spans model training, ensembles, candidate generation, experiments and feedback. Computational work handles literature, models and selection; laboratory work handles DNA synthesis, fluorescence and predicted-versus-measured strength comparisons. Project coordination links design and scheduling.
Project Evidence
Methods & Figures
Teamwork
