March 12, 2019 PAP-Q1-2019-NI-006
Computer algorithms have advanced to the point where machines are able to “learn” as they analyze data. In machine learning, otherwise known as deep learning, systems are able to identify predictive patterns and then output results. Learning also takes place through natural language processing (NLP) — in which machines “read” diverse types of written information — and via robot process automation (“bots”). However, there is some debate about whether these examples can be considered pure artificial intelligence, as these algorithms cannot function without human inputs.1 Instead, the term augmented intelligence is preferred to describe the technology currently in place.
Another way to define artificial intelligence process is with the term “narrow AI.” This refers to artificial intelligence that performs specific tasks and can learn while doing them — but without self-awareness. Again, for “true AI” to be achieved, there would have to be genuine intelligence or observed capabilities like those of the human brain.2 NLP and artificial neural networks designed to mimic the way our brains make sense of the world are two prime examples.3 There are many forms of narrow AI in use today, including computer vision (image recognition), voice analysis, route selection and controlling self-driving cars.4
Machine learning algorithms learn in three different ways, typically referred to as supervised, unsupervised and reinforcement learning.5 Unsupervised learning finds hidden patterns, supervised learning can be used to improve the efficiency of predictions and reinforcement learning is used in modeling. Google DeepMind, which uses deep learning on a convolutional neural network with a form of model-free reinforcement learning, is a leading example of an advanced machine learning system. Google’s most advanced AI is AlphaGo Zero.5
An AI like AlphaGo Zero requires a powerful supercomputer that can rapidly process large, complex sets of data to run properly. As computing power increases, the power of AI algorithms also increases. Two of the fastest supercomputers today are China’s Sunway TaihuLight (93 petaFLOPS) and IBM’s Watson (IBM Power 750, a cluster of 90 supercomputers with 80 teraFLOPS).5 A computer with the capability to perform one billion calculations per second “1 exaFLOP” — which is on the same order of processing power as the human brain — could be created in the near future. The next step is quantum computers, which use single particles, or qubits, to encode information and have the potential to exponentially increase computing power. Qubits are highly unstable, creating a significant engineering challenge. D-Wave introduced the first quantum computers, which are being used for various types of research.5
AI — whether augmented or artificial — has numerous potential applications in the pharma industry, from drug discovery and development to medical imaging, diagnostics, disease diagnosis, therapy planning and hospital workflow design.1,3,5 AI also may help determine the causes of diseases and facilitate the development of personalized medicines.3 The success of any AI initiative depends on the ability of the algorithm, the power of the computer and the quality and quantity of the data.5 The impacts can often be even greater when AI is combined with automation.3
The use of artificial intelligence in drug discovery, in particular, has been a topic of growing interest in the industry. This is because the application of machine learning and NLP techniques can lead to improved predictive modeling and simulation capabilities. This can be applied to drug discovery through the integration of real-world data and electronic medical records from disparate sources. When integrated, this data could mean improved candidate screening and trial selection, optimization of clinical trial designs and better prediction of drug demand.6
AI has the potential to facilitate the identification of new candidate compounds with the greatest likelihood of being druggable, selective and efficacious (molecular design). Synthesis route planning and existing drug selection may also be facilitated by AI,7 as well as biomarker development and the repurposing of known compounds — both existing approved drugs and molecules that failed to be commercialized for their intended indications.4,7
This is all made possible by a combination of advances in deep learning, the increased availability of data and new frameworks for implementing deep neural networks (DNNs), which now are more accurate than the human brain in areas such as in image, voice and text recognition.8 AI imagination deep generative models are also enabling new applications.
When used effectively, AI may also be programmed to improve pharmaceutical manufacturing operations. According to Constantin Loghinov, Managing Director of MILS Group, LLC, “To make machine learning effective in biopharmaceutical manufacturing, gargantuan internal mindset and business process change around data collection, analysis, and use is necessary. There are, however, some quicker Band-Aid solutions that can be initially deployed.”9 AI is already being employed for machine learning–based visual inspection and bacterial culture yield optimization. AI could also be used to leverage data from multiple manufacturing sites within a single organization. Analysis of data generated by Internet-of-Things (IoT) sensors (cameras, thermostats, chemical sensors, etc.) could help achieve more predictable, high-quality, flexible, low-cost manufacturing processes. The ultimate goal — as with many initiatives in the pharma industry — is to reduce the cost and time required for drug development, commercialization and production.
The huge opportunities presented by the application of AI in the pharma industry come with challenges — not just with the technology, but with the data and the way work is conducted. Companies must be convinced that investing in AI technologies will provide value and then have sufficient knowledge to select the technologies most appropriate to the specific applications they are considering.6
Most pharmaceutical researchers lack expertise in data science and thus an understanding of the potential benefits that AI can bring. In a recent survey of 330 drug-discovery scientists conducted by BenchSci, which has developed a machine-learning tool for antibody drug discovery, over 40% of respondents said they were unfamiliar with potential applications of AI.10 They also lacked knowledge of both the technology and the companies offering AI tools and services. Nonprofit innovation advocate the Pistoia Alliance similarly found in a separate survey of 374 life scientists that lack of expertise is seen as the top barrier preventing wider adoption of AI.10
Pharmaceutical companies looking to develop their own internal AI expertise face the challenge of finding talented people who want to join the industry. Most AI experts are in their early twenties and are often most attracted to jobs in well-financed tech startups and businesses like Google’s Deep Mind, as opposed to health and medicine.9,11
In addition, deep learning systems are only as good — or “smart” — as the data used to train them.5 While the amount of data generated grows exponentially, much of it is diverse, uncurated and/or unavailable. Clinical trials today are often for targeted patient populations and involve 1000 (or fewer) people and often do not provide sufficient data for algorithms to identify patterns.11 Patient data is also fraught with privacy issues.
The use of AI in pharmaceutical drug development clearly has significant potential to accelerate the process and enable the discovery of novel medicines that can address real unmet medical needs. As the use of AI increases across all aspects of the pharmaceutical industry from drug discovery to patient selection of clinical trials and post-market safety monitoring, regulatory implications will need to be addressed.
Machine learning, NLP and other AI technologies clearly have the potential to have such positive impacts. In recognition of the potential of AI and other digital health tools, the FDA formed the Digital Health Innovation Action Plan in 2017, which is focused on modifying the agency’s approach to digital health products.18 In 2018, the agency formed an internal data science incubator called the Information Exchange and Data Transformation (INFORMED). It has developed a streamlined path for digital health products, has committed to enabling the use of digital health in drug development and is currently building a flexible framework and new software validation tools to address the unique regulatory concerns associated with AI-based technologies.
In April 2018, the agency approved the first medical device that combined a special camera and artificial intelligence to detect a greater than mild level of diabetic retinopathy in adults who have diabetes in a primary care setting.19 A month later, the FDA permitted the marketing of an artificial intelligence algorithm for aiding providers in detecting wrist fractures.20
The FDA is also working within the industry (including pharmaceutical, medical device and computer technology companies) and academia to develop guidance documents regarding the use of AI in pharmaceutical products.21 One example is the Good Machine Learning Practices (GmLP) document for the evaluation and use of continuously learning systems (CLS) proposed by the Xavier Health CLS Working Team. The goal of the group is to identify ways in which companies could demonstrate confidence in a CLS while maximizing the benefits of AI and minimizing risk to patients.
Many pharmaceutical companies are already tackling these challenges. Most have partnered with AI startups or businesses within established tech firms, even going so far as to make investments in some of them. A few are also investing in internal expertise. Many companies are focusing on the application of AI to drug discovery and development, such as candidate selection, identification of optimum combination therapies and drug repurposing,7 but some are pursuing initiatives designed to improve clinical trials or disease diagnosis.
The healthcare artificial intelligence market is estimated to be growing at a compound annual growth rate of approximately 40–50% and is predicted to reach a value of $8 billion by 20225 or $10 billion by 2024.12 Drug discovery applications are thought to account for 35–40% of revenues. As of August 2018, Clevis Research estimated that close to 30 different start-ups were specializing in the application of AI to drug development.13
This explosion of startups focusing on AI for pharma will likely lead to both more outsourcing of AI-related R&D and acquisitions of some of these highly specialized firms.4 Other drivers include the lack of access to skilled data scientists, the need for complex and sophisticated IT infrastructure and the rapid rate at which advances are being made in the field.
Some of the leading startups include:
Some of these startups are not just developing AI technology but using their platforms to identify drug candidates and take them to the clinic. Berg, for instance, uses deep learning to evaluate patient-driven data; model unknown cancer, diabetes and Parkinson’s disease mechanisms; and identify potential treatments. Its drug candidate BPM31510 is currently in a phase II clinical trial for advanced pancreatic cancer.5 Verge Genomics recently raised $32 million in funding from WuXi AppTec’s Corporate Venture Fund DFJ and others to advance drug candidates for the treatment of neurodegenerative diseases (e.g., Alzheimer’s, Parkinson’s, amyotrophic lateral sclerosis [ALS]) using machine learning models trained on patient and lab data.14
Large tech companies have also elected to develop AI solutions for use in the pharma industry. Perhaps the best known is IBM’s Watson Health, although the business has suffered several setbacks recently; Google’s DeepMind Health is another. Google is exploring the use of AI technology to develop drugs to treat cancer, eye disorders and other diseases.2
Most major pharmaceutical companies have begun investigating the use of AI to some degree across the entire spectrum of potential applications. Most of these programs are conducted in partnerships with companies specializing in AI technology. A few examples include:15
Other interesting activities include the investment by Amgen and Biogen in quantum computing, AbbVie’s application of AI to monitor patient adherence in clinical trials, the partnerships between biotech companies Celgene and Genentech with precision medicine startup GNS Healthcare, the collaborations between Boehringer Ingelheim, Merck, Servier and Takeda with Numerate and the licensing deal between Janssen (a Johnson & Johnson business) and BenevolentAI.
An example of a recent acquisition is that of NextCODE Health, a spinoff of deCODE Genetics, by WuXi AppTec. Now called Wuxi NextCODE, the business is focused on using AI to better understand genes and how they function in order to identify their roles in particular diseases.
Several pharma companies have also made investments to establish internal expertise in AI and develop proprietary technologies. In addition to its external partnerships, GSK formed a new drug discovery group in 2017 to evaluate the potential benefits of an integrated artificial intelligence and machine learning approach.16 The group applies deep learning to identify targets/pathways, optimum drug candidates and the most appropriate patient populations.
AstraZeneca (AZ) is using AI in a wide variety of applications: assay evaluation to accelerate drug discovery, image data analysis to match the right drugs to the right patients and data monitoring in clinical trials.17 AZ developed a prototype Design-Make-Test-Analyze (DMTA) platform that applies AI and laboratory automation to accelerate the development of experimental hypotheses and reliably predict the results of routine assays. The company also developed new computational algorithms that enable accurate and efficient segmentation of large quantities of mass spectrometry imaging data to better understand the link between the tissue microenvironment and drug localization, efficacy and safety. AZ plans to combine these deep learning algorithms with image analysis to accelerate the evaluation of animal models of chronic kidney disease.
A different deep learning algorithm developed by AZ automatically evaluates tissue biomarkers using digital pathology that was shown to score the breast cancer biomarker human epidermal growth factor receptor 2 (HER2) as well as identify samples at risk of misdiagnosis. It intends to make automated analysis of digital pathology images a high-throughput process and to incorporate AI algorithms into the development of diagnostic tests. AZ’s AI-based decision-support system Watcher continuously monitors safety data in early-phase clinical trials to identify potential problems early on. The company intends to augment the system’s capabilities with machine learning and clinical rule sets so that it can be used by patients in their homes.
Pfizer is using AI to analyze electronic medical records to identify clinical markets useful for the diagnosis of the rare heart disease transthyretin cardiomyopathy (TTR CM).1 It is also using AI to improve the marketing of its smoking cessation drug Chantix (varenicline) by identifying patients with characteristics associated with people who have successfully quit smoking.
Amgen, meanwhile, is piloting an NLP-based AI tool that is designed to enhance its ability to identify trends and patterns in manufacturing deviations.18 The tool will access data in a “data lake” comprising raw and transformed data that can be structured, semi-structured or unstructured. The tool finds not only obvious trends but weaker patterns that would not typically be detected through human analysis. The ultimate goal is to achieve real-time predictive models that are not only based on deviations, but rather consider the various factors in manufacturing operations that can lead to deviations.
These types of investments are going to be necessary for all pharmaceutical companies going forward, according to some industry analysts.1 Drug manufacturers will need to have a minimal level of understanding and experience with deep learning, NLP and other AI technologies in order to be able to explain the results they are leveraging and to remain competitive.
Dr. Challener is an established industry editor and technical writing expert in the areas of chemistry and pharmaceuticals. She writes for various corporations and associations, as well as marketing agencies and research organizations, including That’s Nice and Nice Insight.