Computer generation of fruit shapes from DNA sequence

The generation of realistic plant and animal images from marker information could be a main contribution of artificial intelligence to genetics and breeding. Since morphological traits are highly variable and highly heritable, this must be possible. However, a suitable algorithm has not been proposed yet. This paper is a proof of concept demonstrating the feasibility of this proposal using ‘decoders’, a class of deep learning architecture. We apply it to Cucurbitaceae, perhaps the family harboring the largest variability in fruit shape in the plant kingdom, and to tomato, a species with high morphological diversity also. We generate Cucurbitaceae shapes assuming a hypothetical, but plausible, evolutive path along observed fruit shapes of C. melo . In tomato, we used 353 images from 129 crosses between 25 maternal and 7 paternal lines for which genotype data were available. In both instances, a simple decoder was able to recover expected shapes with large accuracy. For the tomato pedigree, we also show that the algorithm can be trained to generate offspring images from their parents’ shapes, bypassing genotype information. Data and code are available at .