Artificial Intelligence (AI) continues to revolutionize many aspects of life science and healthcare industries. Applications of AI include automated cancer detection in medical imaging, structural predictions for proteins from amino acid sequences, and rapid analysis and understanding of medical notes and clinical records. The collection and analysis of research and health data plays an equally critical role in addressing future treatments as the industry works towards developing solutions to accelerate R&D cycles and develop systems for better patient care. In particular, AI applications in drug discovery have increased immensely in the past several years. After years of incremental progress, recent advances in protein structure & function prediction have opened the floodgates to AI-fueled structure prediction for drug discovery, leading to rapid improvements and innovation in the space. AI will become an integral part of the drug discovery workflow in the near future in every life science and pharmaceutical organization. In this article, we explore a high-level overview of deep learning fundamentals and discuss why these techniques have become viable for drug discovery after years of slow adoption by research and chemical scientists.
The foundational concepts enabling the recent explosion of neural networks have been around for decades. Modern deep neural networks are comprised of artificial neurons, which were first introduced in 1943. These neurons are nodes through which data and computations flow. Neurons work to receive input features and these input signals can come from either the raw data set or from neurons positioned at a previous layer of the neural net. The output is a vector of numerical values which flow to the following layer in the network. Models from the 1950s utilized tens of neurons in a single layer to learn simple patterns. As computing power increased, additional neurons were added to the networks, improving performance. Neurons were added horizontally: each neuron received input from the input data itself (Figure 1).
This horizontal grouping or layer has become fundamental to AI. In the 1980s, additional horizontal layers were added to neural networks, where the output of the previous layer becomes the input of the next layer (Figure 2). This “deep” layered structure for neural networks gave the field of deep learning its name and allowed networks to derive higher-level features from the input data, enabling deep neural networks to approximate a wider range of patterns.
Despite the ideas underpinning deep learning existing since the 1980s and their foundations since the 1950s, these algorithms did not start having a real-world impact until the 2010s, when industry funding and implementations increased exponentially. In the realm of drug discovery, we are still in the infancy of deep learning’s applications, with AlphaFold2’s release in 2020 representing the first revolutionary advance in AI-based proteomics for drug discovery.
The delay in the adoption of deep learning for commercially viable drug discovery can be attributed to several factors: Data availability, storage costs, computation costs, and architecture enhancements. Larger data sets, new types of data, and scalable, on-demand IT and research tools allow organizations to query data and perform analyses with more technical feasibility, fueling R&D efforts at a lower cost. As biological and chemical workflows have become digitized, enormous amounts of data are created every day: Protein structures, chemical interactions, genome sequences, etc., Growth of viable data has led to AI learning a wide array of biochemical patterns. Even with the large amounts of data produced, training deep neural networks would not be possible without a cost-efficient method to store the data. Deep neural networks can now access terabytes of training data at costs that are reasonable for commercial applications.
Deep learning algorithms now make use of a massive number of layers with hundreds of billions of parameters. The number of parallel computations required to optimize these parameters is enormous and would be completely unfeasible on a central processing unit (CPU). With the advent of CUDA programming as well as the cost reduction & performance increases in graphics processing units (GPUs), it is now possible to train models spread across thousands of GPUs, with millions of computations occurring in parallel. Improvements in computational power have enabled new breakthroughs in key deep learning architectures which have formed a major part of the recent advances in drug discovery.
These factors, along with novel services and algorithms, have converged to produce a wave of AI applications for drug discovery. AlphaFold2 in 2020 represented a massive leap in protein structure prediction, outperforming the competition by leaps and bounds at CASP14 and achieving accuracies comparable with experimental methods. MaSIF, also released in 2020, was able to use geometric deep learning techniques to predict protein-protein interaction hotspots on a protein surface. Graph neural networks and transformer models have also made strong gains in retrosynthesis, where the chemical reactions required to produce a molecule are derived from the molecule itself.
These new applications are just a sampling of the immense possibilities enabled by the advent of AI in drug discovery and analyzing multimodal data will power the next generation of healthcare outcomes & treatment. As computation and storage costs continue to decrease, GPU processing power increases, and data becomes increasingly available, we will continue to see new AI architectures revolutionize drug discovery, from structure prediction to drug-protein interaction identification. As AI rapidly advances in the drug discovery space, it is imperative for researchers to familiarize themselves with the available models and understand the concepts underpinning these innovations. With time, the newly-developed algorithms will become tools as indispensable as a microscope to the bench scientist in coming years.