The International Genetically Engineered Machine Competition (iGEM) has built a platform for undergraduates to show and exchange their creative ideas about Synthetic Biology. Project Modelling nowadays becomes a more and more important part in the competition and many iGEM models worth further studying and analyzing.
We now consider to collect representative models concerning CRISPR system and machine learning which are mostly related to our work. Next, we intend to organize iGEM models about key scientific issues on our Github page.
List of outstanding models about CRISPR system and machine learning in the last 3 year on iGEM Team Wikis.
-
Name: Target Recognition Model
Type: Kinetic Model
Functions: Find the numbers and locations of lure sequences and give coresponding intructions to the project
Evaluation: The approximation works well for single mismatch situation. A computational solution for multi-mismatch situation is provided.
-
Name:
-
Regulatory Model
-
Stochastic Model of dCas9 Binding and Replication Initiation
-
Plasmid Copy Number Hacking Model
-
Gene Expression Noise Control Model
-
Productivity Model
-
Quorum Sensing Model
Type:
- ODE
- Markov chain
- Logistic process
- Yule-Furry process
- ODE
- PDE
Functions:
-
Explain the mechanisms of their system's regulation parts.
-
Show that their system hacks the bacteria's replication rate.
-
Discuss how their system can control the plasmid copy number.
-
Illustrate how their system can control the expression noise in a cell.
-
Explain why their system is able to increase the cells' productivity to multiple types of products like GFP and indigo.
-
Illustrate how their system works coupling with a quorum sensing system.
Evaluation: Model 5 and 6 were validated by experiments.
-
-
Name:
-
CRISPR SHERLOCK Kinetic Model
-
Glutathione Detection Kinetic Model
-
Eicosane Detection Kinetic Model
Type: Kinetic Model
Functions:
-
Understand how different CPLX1 concentrations impact the time until maximum active GFP was established
-
Follow glutathione degradation into glycine.
-
Understand the amount of RFP produced in the presence of varying concentrations of either glutathione or eicosane.
Evaluation:
- The model would give a more true reflection of the system if crRNA degradation by trans-cleavage was also included. Rates used were equal to degradation of CPLX1 mRNA degradation and GFP activation. Inclusion of crRNA degradation did not impact time until maximum GFP concentration.
- The model should be re-parameterised in the future to include enzyme rates for E. coli become available.
- The model allowed greater understanding of the system and careful consideration for the other two systems to ensure both would be functional.
-
-
Name: ASFAST
Type: Kinetic model
Functions: Find out the reaction rate of ASFAST under different virus D.A concentration and take this as a reference to set its detection limit。
Evaluation: More researches have to be done on the dissociation rate of trans activity to determine the exact rate of its reverse reaction.
-
Name:
-
Dynamic Model of Heavy Metal Detection Biosensor
-
Free Energy Model of Off-target Problem
Type:
-
ODE
-
Probability theory and dynamic deduction
Functions:
-
Characterize the pathways quantitatively and predict their performance.
-
Investigate the off-target problem in gene editing by the CRISPR-Cas system.
Evaluation:
-
The concentration of smURFP is relatively sensitive to parameters such as ktx3,ktl3,ktx4,kb4,kb6,kd2,kd5, kd6,kd7,kd8,kd11, etc. However, due to the lack of previous modeling studies on dCas9-RNAP, some kinetic parameters may not be very accurate.
-
After testing, they find its rate can reach approximately 2e8 base/h(under parallel computing in 4 cores). Besides the default para:meters, they hope their model can hit more true data.
-
-
Name: System Modeling
Type: ODE
Fucntions: Predict the time point when the all antibiotic resistance genes(ARGs) are cut off and the cell auto-lysis is going to begin.
Evaluation: Just a quite weak pulsed light signal is enough to start the operation of their system. The system is safe because the accidental leakage of bacteria will cut off all the ARGs and autolysis for the natural light.
-
Name: SpliceMIT – Splice Modelling Intronic Technology
Type: Non-parametric statistical method
Fucntions: Generate antisense oligonucleotides (ASO) for a given DNA/RNA sequence, and then analyze and output the most effective ASOs.
Evaluation: The probabilities of secondary structure of gRNA sequence and pre-gRNA sequences are compared. The correlation efficient is 0.91, with an extremely low p-value.
-
Name: CRISPR Activity Model
Type: Biophysical model
Fucntions: Provide the program with the appropriate genome and target gene sequence, and return a list of potential guide sequences ranked from best to worst.
Evaluation: Using this model, they have designed sgRNAs to target essential virulence regions in Agrobacterium. The guides obtained from running the model were shown to be effective in vitro.
-
Name: LacILov System Model
Type: ODE
Functions: Capture the dynamics of construct developed in the Wet Lab to further characterize its behaviour.
Evaluation: Adjusted R-squared=0,8238 p-value=9.741e-16.
-
Name: Outreach Database
Methods: Random Forest, Gradient Boosting, Neural Networks, Ridge Regression, C-SVM with classifier chains
Functions: Automatically label successful outreach events that happened in 2018 based on their descriptions.
Evaluation: The best models are using C-SVM with classifier chains. These models has an accuracy of 95.20% for labeling the "Project Tags", an accuracy of 93.25% for labeling the "Audiences", and an accuracy of 92.11% for labeling the "Goals".
-
Name: Emulsion Construction Prediction
Methods: SVC, kNN, MLP
Functions: Find an emulsion which allows for the maximum removal of chlorophyll from the oil by finding the variables of temperature and concentrations of oil, water, and surfactant.
Evaluation: The phase diagram model generated with the SVC algorithm produced the most physically accurate results.
-
Name: Mutagenicity Prediction
Methods: SVM
Functions: Predict mutagenicity based on chemical substructures.
Evaluation: They import the linear regression of the result of Ames test to our model prediction. The result shows that the two has a
$R^2$ of 0.9699, which indicates their high relativeness. -
Name: Growth Curve Model
Methods: Polynomial Regression
Functions: Predict the doubling time.
Evaluation: The prediction quality of the model is poor. The degree of the polynomial is influencing the performance of the model, but there is no clear trend visible.
-
Name: Future
Methods: BP Neural Network
Functions: Give a prediction of their method on human body, and provide a treatment model with higher precision.
Evaluation: The total set AUC is approximately 0.7.
-
Name: Efficacy of Toehold Switches Prediction
Methods: Multivariate Regression
Functions: Predict the efficacy of toehold switches according to their dynamic range.
Evaluation: The 10-fold cross validated model metrics output an adj.
$R^2$ value = 0.59. Their model outperforms the only other existing multivariate model proposed by CUHK’s iGEM 2017, whose reported$R^2$ value was 0.22. -
Name: Pre-counseling
Methods: SVM
Functions: Take in this set of features, and output the likelihood of testing positive.
Evaluation: The patient database can be replaced with a real database representing real and unique patients. This will give more accurate results for a given pre-counseling input.
-
Name: EC Number Prediction
Methods: SVM, kNN
Functions: Predict probabilistic EC number and provide synthetic biologists with more help on finding right enzymes.
Evaluation: The results of the tests indicated the effectiveness of their prediction tool.
-
Name: Determining pathogenicity of the Xanthomonas Species
Methods: Random Forest
Functions: Find a genetic basis for non-pathogenicity and select a set of non-pathogens that conform to this genetic basis.
Evaluation: The performance of 100 models was combined to estimate the sensitivity and specificity of the model, resulting in an average sensitivity (non-pathogen prediction rate) of 0.90 ± 0.12 and average specificity (pathogen prediction rate) of 0.81 ± 0.11.