Current progress and open challenges for applying deep learning across the biosciences - Nature Communications

The recent success of AlphaFold21 in predicting the 3D structure of proteins from their sequences highlights one of the most effective applications of deep learning in computational biology to date. Deep learning (DL) allows for finding a representation of the data with multiple layers of abstraction using complex models that are composed of several layers of nonlinear computational units (Figure 1). Observed through the success of DL in a broad variety of application domains, the efficacy of using DL depends on the development of specialized neural network architectures that can capture important properties of the data such as spatial locality (convolutional neural networks – CNNs), sequential nature (recurrent neural networks – RNNs), context dependence (Transformers), and data distribution (autoencoders – AEs). Figure 1 illustrates six DL architectures that have found the most applications within the realm of computational biology. We refer the reader to LeCun et al. 2 for a complete review of DL methods and architectures and keep the focus of the paper on computational biology applications. These DL models have revolutionized speech recognition, visual object recognition, and object detection and have lately played a key role in solving important problems in computational biology.