EXPAR (EXPonential Amplification Reaction) is an isothermal amplification reaction first described by Jeffrey Van Ness et al in 2003.
The reaction start when a sequence termed "trigger-X" is present in the sample, which hybridizes to the complementary sequence called Template-X'-X'. This creates and exponential amplification loop in which this trigger-X sequence get amplified in a short amount of time.
Different ways of generating this trigger-X sequences to prime the reaction have been described. For instance, one implementation for detecting SARS-CoV-2 RNA relies on the generation of a DNA-RNA hybrid that is recognized and cut by a nickase that release the trigger-X sequence (Jake G. Carter et al, 2021).
Also, some implementations make use of the CRISPR-CAS technology to recognize the target and trigger the reaction (Mengqi Huang et al, 2017).
The mechanism for this implementation is summarized in the following figure:
The main drawback of EXPAR is its high background amplification, produced by the nonspecific trigger of the reaction. A well performing design is the one in which the positive sample amplifies in a relative short time, and the negative sample (no template control) yields a signal after a long time. A study have found that a good design of the Template X'-X' is key for obtaining a well performing reaction (Jifeng Qian et al, 2012). In this paper, the authors measured the performance of 384 templaete-X'-X' to create a dataset that was later used for training two different classifiers. One based is based on a SVM using a position weight matrix and the other uses a RELIEF attribute evaluation with Naïve Bayes classification.
Herein, we generated a pipeline for designing optimal Template-X'-X' sequences.
If the user has an specific target region, the pipeline will start in step 3.
-
The user inputs the sequences which will be considered as the target of the reaction. The sequences are aligned and gaps are removed.
-
Once conserved regions are found within the alignment, variable regions are filtered out (as they would not be good as targets).
-
PAM sequences (needed for CAS binding) are found and flanking regions are stored in a dataset.
-
The sequences obtained in step 3 are analyzed for the possible formation of hairpins with a bot that uses UNAFold. The thermodynamic parameters of each sequence (number of structures, melting temperature and ΔG) are stored in the dataset.
-
The sequences are then analyzed for the formation of self-dimers with a bot that uses the OligoAnalyzer (IDT). Again, the thermodynamic parameters of the sequences are stored in the dataset.
-
The best sequences are then analyzed by three different classifiers:
- A SVM (reproduction of Jifeng Qian et al, 2012)
- A Naïve Bayes classifier (reproduction of Jifeng Qian et al, 2012)
- A deep-learning algorithm developed for this project (Deep-EXPAR).
Then, a short-list of the sequences classified as well perfoming by these models can be manually selected for experimental evaluation.
Accuracy of classifiers used in this pipeline:
| SMV | Naïve Bayes | Deep-EXPAR* | |
|---|---|---|---|
| Sensitivity | 67.7% | 70.6% | 87.3% |
| Specificity | 80.4% | 77.5% | 90.2% |
*Deep-EXPAR is still being improved, so the accuracy of this model could increase in the future.