Machine Learning\nRecognition of Artificial DNA Sequence\nwith Quantum Tunneling Nanogap Junction

Artificially synthesized DNA holds significant promise\nin addressing\nfundamental biochemical questions and driving advancements in biotechnology,\ngenetics, and DNA digital data storage. Rapid and precise electric\nidentification of these artificial DNA strands is crucial for their\neffective application. Herein, we present a comprehensive investigation\ninto the electric recognition of eight artificial synthesized DNA\n(<i>x</i>DNA and <i>y</i>DNA) nucleobases using\nquantum tunneling transport and machine learning (ML) techniques.\nBy embedding these nucleobases within a solid-state nanogap junction,\nwe calculated their fingerprint transmission and current readouts\nand also analyzed the influence of electronic coupling and molecular\norbital delocalization on these properties. The trained ML model achieved\na predictive basecalling accuracy of up to 100% for <i>x</i>DNA nucleobases and 99.80% for <i>y</i>DNA transmission\nreadout data sets. ML explainability study revealed that normalized\ndescriptors have a greater impact on nucleobase prediction than the\noriginal transmission function, proving more effective in disentangling\noverlapping artificial DNA nucleobase signals. Quaternary classification\nresults highlighted higher recognition accuracy for <i>x</i>DNA nucleobases than for <i>y</i>DNA nucleobases. Furthermore,\nprecise calling of complementary, purine, and pyrimidine base pair\ncombinations was demonstrated with high sensitivity and an F1 score.\nOur findings reveal the feasibility of highly sensitive and precise\nelectrical recognition of artificial DNA nucleobases, which can transform\ngenetic research and spur advancements in genetic data storage, synthetic\nbiology, and diagnostics.