With the increasing global population, ensuring stable food supply has become critical. This requires plant breeding to achieve high rates of genetic gain. Genomic selection, a technique that leverages DNA variation and phenotypic data to predict performance, has been shown to boost selection gains and reduce breeding cycles in various crops. Additionally, deep learning techniques, a type of artificial intelligence, are increasingly being used in genomic prediction, providing the potential to revolutionize fields like precision medicine and agriculture.
Deep learning architectures, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Autoencoders, and Generative Pretrained Transformers (GPT), have been successful in processing biological data. However, they can be expensive to train, particularly for genomics tasks with significant data requirements and privacy concerns.
Deep learning has significant applications in genomics, including gene expression characterization, regulatory genomics, functional genomics, and structural genomics. It has been used to analyze complex biological data and uncover insights into genetic mechanisms. Methods like CNNs and RNNs have been applied in characterizing gene expression and identifying regulatory motifs, achieving high accuracy in these tasks. Further, they have also shown promise in structural genomics, aiding in protein structure classification and homology detection.
Generative models, including Wasserstein Generative Adversarial Networks (WGAN-GP), Restricted Boltzmann Machines (RBM), and Variational Autoencoders (VAE), were employed in the study mentioned in this paper. These models were trained on two datasets from the 1000 Genomes project and used to generate artificial genomic sequences. The models’ performance was evaluated through Principal Component Analysis and a measure called the nearest neighbor adversarial accuracy.
While the VAE model was not as successful, the WGAN and CRBM models generated substantial artificial genomic sequences. However, Linked Desequilibrium (LD) decay analysis revealed that both models had lower LD than real genomes. Also, it was revealed that the CRBM model performed better than the WGAN in the 3-point correlation analysis but showed anomalies, potentially indicating sequences outside the real data space.
Despite deep learning’s potential, more work is needed to confirm its superiority over conventional models in predictive power. Challenges such as computational complexity and model optimization also persist. Furthermore, privacy concerns need to be addressed. Nevertheless, advancements in model training and privacy safeguards could pave the way to artificial genome banks, leading to broader access to genomic data.
Deep learning proves promising in revolutionizing genomics. However, navigating challenges around predictive accuracy and interoperability are essential for meaningful breakthroughs. Assuming these hurdles can be overcome, deep learning could hold the key to meeting the increasing demand for food and revolutionizing the fields of precision medicine and agriculture.
