How artificial intelligence is driving innovation in the pharmaceutical

Although disease affects everyone, and trillions of pounds have been spent on research and development, there are still thousands of diseases without any treatment. There are over 300 million people suffering from rare diseases for which no medicine will be developed anytime soon, unless we dramatically disrupt the existing economic and development models.


Artificial Intelligence
Developing a drug and getting it to market can currently take 10-15 years at a cost of more than $2.5 billion, once the expense associated with compounds that fail during trials is taken into account. Furthermore, the top 10 selling drugs on the market today only work on average for just 30-50% of the patients for which they are prescribed.
Even when a molecule reaches clinical development, the chances of the drug making it to the market are less than one in 10. Long term, this is not sustainable. It also means that patients are not receiving inventive new treatments for diseases that are currently either untreatable or inadequately treated. So, we need to both increase our chances of success and cut the cost of failure -and artificial intelligence (AI) can play a role in both.

What is AI?
In essence, AI can be defined as a number of different computer sciences used to perform capabilities Although disease affects everyone, and trillions of pounds have been spent on research and development, there are still thousands of diseases without any treatment. There are over 300 million people suffering from rare diseases for which no medicine will be developed anytime soon, unless we dramatically disrupt the existing economic and development models.
How artificial intelligence is driving innovation in the pharmaceutical industry traditionally requiring human intelligence. These include machine learning, natural language processing and deep learning, and they have become ubiquitous in our daily lives-from powering suggestions that Amazon or Google make when we browse the Web or underpinning predictive text and translation apps on our smartphones.
Machine learning can be differentiated into supervised or unsupervised learning. In supervised learning, the machine is trained on a set of annotated data so it learns the features of the data that are important. For example, categorizing tumour subtypes from pathology samples that have already been labelled or categorized by a pathologist. The machine is then shown unlabelled tumour pathology slides and categorizes them according to what it has already learned. Machine performance depends heavily on the training set used. In unsupervised learning, the machine is given the data without stipulating which features are important and it learns to categorize the samples over time based on features it considers important.

Artificial Intelligence
Natural language processing allows the machine essentially to read and extract facts from text. Deep learning is a subset of machine learning where artificial neural networks, algorithms inspired by neuroscience, learn from large amounts of data. The deep learning algorithm performs a task repeatedly, each time tweaking it a little to improve the outcome with the neural networks having a number of layers that enable learning.

How can AI help drug discovery?
The ability of AI to ingest, analyse and extrapolate from huge amounts of structured and unstructured data means that AI will not only be able to improve the drug discovery process but can also be applied to drug development (see Figures 1A and B).
Around 90% of the digital data ever created in the world has been generated in just the past two years but only 1% of that data has been analysed. Every day over 10,000 bioscience publications are uploaded to the Web, not including the massive quantities of highly relevant genetic, 'omic' and imaging data that is created. Humans alone cannot use this information without help from technology since any one individual or team is only able to analyse a tiny fraction of this information.
To create new knowledge in the pharmaceutical industry from this wealth of data requires the use of AI to augment human insight. In doing this we should be able to choose better targets to challenge a disease, make better molecules with more predictive properties and select the right patients for a given approach.

The role of AI in choosing disease targets
Biology is complex and human biology is extremely complex. Most drugs fail either because of unacceptable side effects or lack of efficacy-in fact, most drugs fail in phase II and III clinical trials as a result of not modulating the right target in the disease. We have to create change by acknowledging that access to more relevant disease data and increased understanding of disease biology will lead to improved target selection.
We do this by ingesting the information, recognizing relevant entities in the information (e.g., genes, proteins, diseases, symptoms, chemicals) and extracting the relationships between the entities to build what is essentially a systems biology map or knowledge graph of all the known facts.
AI can do this extremely rapidly-manual annotation of the relationships articulated in just the abstract of a review article took one of our scientists a couple of hours-the computer 'read' and mapped the relationships in the whole large review article in just under seven seconds. So using this technology allows us to ingest and 'read' hundreds of millions of documents (papers, patents etc.) and create a knowledge graph containing over a billion relationships. We can use this knowledge graph in many different ways.
Given the known information, we can use diversified AI-based methods for target discovery, spotting Artificial Intelligence previously unrecognized links between potential targets and particular diseases. We can look at where these targets sit in the disease biology and pick the best assays to screen them in, e.g., inflammatory assays if the target is on an inflammatory pathway or autophagy screens if the target sits in an autophagy pathway. If a given target is not one that is chemically tractable we can look upstream or downstream for alternative targets that might be amenable to medicinal chemistry.

The role of AI in making the right molecule
There are two ways that AI can aid the design of new molecules. Firstly, machines can be trained using data from known compounds on developability criteria such as pharmacokinetic and safety parameters or binding affinity. The AI can then analyse compounds that medicinal chemists suggest for synthesis and predict whether the compounds are likely to have the desired properties. When the compounds are tested, the results are fed back into the system so it can learn by both its successes and mistakes.
The aim here is really to iterate in silico and make fewer physical molecules in the lifetime of a drug programme, this in turn will reduce costs-if we only need to synthesize 10% of the molecules we would have previously needed to get a candidate molecule suitable for humans, then costs should be reduced by 90%. In addition, it cuts the time it takes to get there-it's feasible to go from an initial 'lead' to a 'candidate' medicine in 12-14 months rather than the three to four years it would otherwise have taken.
In addition, AI can be used to explore new chemical space, suggesting the creation of molecules that a medicinal chemist would not necessarily have thought of. Previously, when a computer suggested molecules for synthesis there was usually a clear route to making them. Now we can use machine learning to suggest molecules that are synthetically tractable and to create a plan for how to build them.

How AI can help in selecting the right patient
It is well known that many diseases are heterogenous, e.g., depression, yet our ability to appropriately stratify patients in clinical trials has been limited. Ultimately this has led to variable outcomes and, consequently, increased trial size to account for this. In addition, many patients fail to respond even to drugs that have reached the market. A better understanding of the reasons for this heterogeneity would have a significant impact on cost and benefit.
There is a much greater wealth of data available to allow researchers to explore the links between multiple types of patient-derived data and a certain disease, to better understand which patients are most likely to respond to a given therapy, which patients might be more likely to have side effects and what the course of the disease would probably be in the absence of a treatment. This in turn enables better clinical trial design. However, without the power of AI it would be impossible to explore these vast data sets to enable these associations to be made.
A range of AI and machine learning techniques can be used to analyse this dimensional patient-level data ( Figure 2) and the resulting patient endotypes can be used to inform drug development in a number of ways. They can allow better identification of biomarkers and drugs with a biomarker in phase I that can be used for patient selection. These are over three times more likely to reach marketing approval.
Identification and selection of patients who have the same disease trajectory reduces patient variation, enabling smaller trials at lower cost. Understanding the mechanistic basis for a particular patient endotype can allow selection of those patients most likely to respond to the drug in question and this should increase the chances of success. In the field of oncology, it has now been demonstrated that the indications for a particular mechanism can be expanded from one disease subgroup of patients to other diseases where the same mechanism has also been shown to be important. • BenevolentAI starts AI collaboration with AstraZeneca to accelerate drug discovery. Accessed 22 July 2019: https:// techcrunch.com/2019/05/01/benevolentai-starts-ai-collaboration-with-astrazeneca-to-accelerate-drug-discovery/ • Industry experts with Jackie Hunter: using artificial intelligence in drug discovery. Accessed 22 July 2019: https://www. voxmarkets.co.uk/articles/industry-experts-with-jackie-hunter-using-artificial-intelligence-in-drug-discovery-ef01d69/ • The promise of using AI for target identification and working with AstraZeneca. Accessed 22 July 2019: https:// benevolent.ai/blog/the-promise-of-using-ai-for-target-identification-and-working-with-astrazeneca  Additionally, AI can play an important role in many other aspects of drug development, including clinical trial site selection, analysis of real-world evidence from wearables, extension studies and for drug discovery, e.g., via analysis of multiplex assays and screening data.
Fundamentally, the status quo of medicine discovery and development requires a dramatically different approach. We need to push the boundaries of AI and machine learning and unlock the power of decades of research, to truly understand diseases and to develop treatments for the millions of patients that need them. ■