Prediction of mycotoxin response of DNA-wrapped nanotube sensor with machine learning

DNA-wrapped single-walled carbon nanotubes (DNA-SWCNTs) have demonstrated great versatility as optical sensors. SWCNTs emit a near-infrared fluorescence that is responsive to the slightest changes in their environment, enabling the creation of sensors that can respond to single-molecule fluctuations within the vicinity of their surfaces. The fluorescence response and surface interactions of these sensors are determined by the DNA wrapping sequence. However, the lack of information on the relationship between the DNA sequence and its effect on the SWCNT fluorescence remains a bottleneck for designing sensors specific to analytes of interest. The use of directed evolution was recently demonstrated in order to evolve SWCNT sensors towards mycotoxins through iterative cycles of DNA mutation, screening and selection. In the current work, we use the data acquired during the directed evolution of DNA-SWCNT sensors to train machine learning (ML) algorithms. Artificial neural network (ANN) and support vector machine (SVM) methods were used to predict the response of DNA-SWCNT sensors to the mycotoxin. The reliability of the models was assessed through cross-validation. The cross-validated ANN and SVM models were able to accurately classify the various DNA-SWCNTs as yielding either a high or low fluorescence response with an accuracy of 73 and 81%, respectively. The models were further tested on alternative similar and dissimilar DNA sequences outside of the initial training dataset. The ANN model showed a better ability to predict dissimilar DNA sequences resulting in a high sensor response in comparison with the SVM model. In addition, the possibility to combine the two SVM and ANN models with directed evolution methods was investigated. The experimental results showed that the SVM model was able to predict the response of DNA-SWCNT sensors with 95% accuracy. Finally, the Hierarchy and k-means++ clustering methods were used to examine the similarity and dissimilarity of each DNA sequence at every stage of our investigation. In this work, we show that the application of ML algorithms to directed evolution libraries of DNA allows one to accurately map the performances of DNA-SWCNT sensors within a particular DNA sequence space. Moreover, the computational success of this mapping provides a framework for replacing current empirical approaches with the rational design of DNA sequences for SWCNT sensing.