Can Deep Learning Crack the Genetic Code? A Data-Driven Approach

The genetic code is textbook scientific knowledge established without AI. This study tests whether a neural network can re-discover, autonomously, the mapping between codons and amino acids and build the complete deciphering dictionary from transcript–protein training pairs. We compare Deep Learning architectures and quantitatively estimate the size of the required human transcriptomic training set to achieve the best possible accuracy in codon-to-amino-acid mapping. We investigate the effect of a codon embedding layer (semantic similarity between codons) on the training accuracy rate, and the benefit of using the unbalanced representations of amino acids in human proteins to speed deciphering of rare amino-acid codons. Deep neural networks require large datasets; deciphering the genetic code is no exception. Achieving high test accuracy and unequivocally identifying rare codons (e.g., tryptophan and stop codons) requires on the order of millions of cumulative codon–amino-acid pairs presented over tens of epochs, depending on architecture and settings. We confirm that the generic capacities and modularity of deep neural networks allow efficient customization to learn the genetic code deciphering task.

Marc Joiret, Gabriele Gianini, Rocco Zaccagnino, Antonio José Jimeno Yepes, Giulia Lorini, Giulia Lalli, Francesca Rossi, Pierre Collignon, Luc Leybaert, Matteo Leoni