Identification of Active Oligonucleotide Sequences Using Artificial Neural Network

Alex Luke, Sarah Fergione, Riley Wilson, Brady Gunn, and Stan Svojanovsky

Volume 9, Issue 2 (December 2018), pp. 30–36

PDF icon Download PDF

  author={Alex Luke and Sarah Fergione and Riley Wilson and Brady Gunn and Stan Svojanovsky},
  title={Identification of Active Oligonucleotide Sequences Using Artificial Neural Network},
  journal={The Journal of Computational Science Education},
Copied to clipboard!

In this project we designed an Artificial Neural Network (ANN) computational model to predict the activity of short oligonucleotide sequences (octamers) with important biological role as exonic splicing enhancers (ESE) motifs recognized by human SR protein SC35. Since only active sequences were available from the literature as our initial data set, we generated an additional set of complementary sequences to the original set. We used back-propagation neural network (BPNN) with MATLAB® Neural Network Toolbox™ on our research designated computer. In Stage I of our project we trained, validated and tested the BPNN prototype. We started with 20 samples in the training and 8 samples in the validation sets. Trained and validated BPNN prototype was then used to test the unique set of 10 octamer sequences with 5 active samples and their 5 complementary sequences. The test showed 2 classification errors, one false positive and the other false negative. We used the test data and moved into Stage II of the project. First, we analyzed the initial DNA numerical representation (DNR) and changed the scheme to achieve higher difference between the subsets of active and complementary sequences. We compared the BPNN results with different numbers of nodes in the second hidden layer to optimize model accuracy. To estimate future model performance we needed to test the classifier on newly collected data from another paper. This practical application included the testing of 41 published, non-repeating SC35 ESE motif octamers, together with 41 complementary sequences. The test showed high BPNN accuracy in the predictive power for both (active and inactive) categories. This study shows the potential for using a BPNN to screen SC35 ESE motif candidates.