The convergence of mind and matter, the coming together of digital and material sciences lies at the heart of the fourth industrial revolution. From search algorithms to self driving cars, artificial intelligence is everywhere.

Artificial Intelligence (AI) is an area of computer science that enables machines to work and think like humans. Generally viewed as the simulation of human intelligence by machines, the AI comprises of learning, reasoning, problem solving, planning and self correction. Machines are increasingly going beyond the physical repetitive work and include planning, strategising and decision making as a part of their life routine.

AI is based on a key technology called machine learning that allows systems ability to self learn and improve from experience without being explicitly programmed. The artificial neural networks of such as Deep Neural Networks (DNN) or Recurrent Neural Networks (RNN) are two key aspects of machine learning driving the evolution of technology.

Application to drug discovery In search of a novel drug, the clinical candidate molecules must have right potency for a given biological target, avoid undesired targets and exhibit favourable absorption, distribution, metabolism, excretion and toxicity properties. Thus, optimisation of a therapeutic molecule along these constraints is a significant multidimensional challenge. People have used machine learning technologies like support vector machines, random forest and bayesian learning to find the right molecule for clinical trials.

In a recent study, AI based models were generated to accurately predict affinity, selectivity and cellular activity of >130,000 compounds across the kinase family of 92 members (Martin et al 2011). In another study prediction models for over 280 kinases were generated using Random Forest method of machine learning algorithms with to perform a more efficient virtual screening for compound repurposing and detection of the off-target effects.

Deep neural networks (DNN) have been used in predicting properties of substances that are sometimes difficult to catch. DNNs belong to the class of artificial neural networks, which are brain-inspired. Deep neural networks have been successftdly used to accurately predict properties of therapeutic molecules. Here multiple nodes are interconnected like the neurons in the brain with a hidden layer in between the input and the output. DNNs use a number of hidden layers to compute a significantly large number of parameters for training the neural network.

Within the space of biology, deep learning has made impressive strides in vision and speech applications pushing technologies to a new level. However, the impact of Deep Learning on drug discovery is beginning to unfold. The 2012 Merck Kaggle molecular activity Challenge was to identify the best statistical techniques for predicting biological activities of different molecules, both on- and off-target, given numerical descriptors generated from their chemical structures. An encouraging outcome of this study showing the superiority of AI approaches was followed by a large benchmark study on a dataset from ChEMBL.

ChEMBL is a manually curated chemical database of bioactive molecules with drug-like properties, maintained by EMBL’s European Bioinformatics Institute (EMBL-EBI). ChEMBL is an important open access resource used for the discovery of new drugs. In a significant benchmark study (Lenselink et al (2017), on high quality ChEMBL data ver 20, DNNs turned out to be top performing classifiers, highlighting their value (27% more) over conventional methods of computational drug discovery.

Drug discovery is now at the point where huge publicly available libraries have moved the domain of discovery close to big data analysis. Machine learning methods are increasingly used for virtual screening of compounds from high throughput screening resources and deliver enhanced value for predicting ADME/Tox properties. In a recent study (Korotcov et al 2017) data representing whole cell screens, individual proteins, physicochemical properties and so on, was used. DNNs showed superior performance in predicting biological activity, solubility and ADME properties over traditional methods.

Designing a new future

The entire process of drug discovery is very expensive and lengthy. Even if one has billions of dollars to spend, it can take upto a decade and a half to graduate an idea from its bench to bedside. This is in contrast with other industry sectors where products are quickly visible and market value can be computed over a brief x-axis of time. Here the drugs that didn’t make to the finishing line also have to be paid for. It has been estimated that more than 30 Pharma companies, including Amgen, AstraZeneca, Bayer, Genentech, GSK, Merck, Novartis, Pfizer, Roche, Sanofi and so on use Artificial intelligence in drug discovery.

Machine Learning has been used for the last two decades in drug discovery. A recent offshoot of machine learning is an amazing science of Deep Learning (DL). Though deep learning has proven to be superior compared to traditional methods, it generates very large training sets that require more efficient computational power. Nevertheless, DL has been found superior for de novo molecular design and reaction predictions e.g, bioactivity prediction.

Not only the established companies but startups are also using AI tools to enhance discoveries, arrive at fast data-driven decisions, extract causal associations from published literature, analyse multiomics data to visualise latent patterns, predict targets and biomarkers, manage inventory, optimize clinical trials, detect weak signals of innovation, find patterns from biomedical data and build hypotheses, leading to personalised medicine, automate diagnosis, curate scientific papers, understand disease mechanisms, repurpose existing drugs and so on. It is hoped that further technological developments AI based platforms will bring a paradigm shift in delivering drugs quicker, cheaper and personalised.

Additional reading Martin, E et al. Profile-QSAR: A novel meta-QSAR method that combines activities across the kinase family to accurately predict affinity, selectivity, and cellular activity. J. Chem. Inf. Model. 2011, 51, 1942-1956 Chen, H et al. The rise of deep learning in drug discovery. Drug Discov. Today 2018, 23,1241-1250 Lenselink et al: Beyond the hype: deep neural networks outperform established methods using a ChEMBL bioactivity benchmark set. J Cheminform 2017: 9, 5 . 86-018-05267-x

Related posts

Leave a Comment