Proceedings: DIA 2019 Global Annual Meeting

AI in Drug Discovery and Development:
Emerging Technologies and Applications
Sandra Blumenrath
DIA
I

f the steady decline in new first-in-class drugs coming to market is any indication, the current system of drug discovery is inefficient and insufficient to meet modern challenges. Automating the many decision-making processes currently carried out by researchers is hailed as the linchpin of the discovery and development of novel therapeutics, especially for difficult-to-treat conditions. In addition to data mining, another hope for AI in pre-clinical drug development is to capture patterns that are difficult to identify, simulate quantum states, or even suggest the structure of a new therapeutic molecule.

Assembling experts from the application and research side of drug discovery, the session AI in Drug Discovery and Development: Emerging Technologies and Applications addressed recent innovative ideas and tools within the realm of AI that could potentially shake up drug discovery as we know it, moving past generalist applications of AI to much more specific, purpose-built tools.

Key Takeaways

  • Quantum computing tools have the potential of addressing highly complex computational modeling problems that are otherwise intractable with classical computing methods. Examples include problems in quantum chemistry that are applicable to pharmaceutical use cases, such as the accurate simulation of atoms and molecules, combinatorial optimization of, e.g., molecular structures and interactions, and machine learning.
  • Metabolomics, which captures the metabolites that result from molecules processed through a metabolic pathway, is an emerging field that—together with AI-based data processing and analysis—can be used as a highly efficient drug discovery tool. In 2018, an estimated 20 different metabolites were discovered that could be used as drugs to treat diseases.
  • One of the most powerful use cases for AI on the clinical trial side of drug development is the identification of digital biomarkers to power new and less fragile endpoints. Protocol development (design, optimization, and automation) also lends itself well to AI-based support. The Common Protocol Template has been a critical enabler for efficient, automated protocol development using AI.

Near-Term Challenges and Opportunities for Quantum Computing in Drug Development

Quantum computing is an exciting field of research with the promise of big payoffs. The basis for this type of computing is provided by quantum mechanics, a subfield of physics that describes the behavior of very small (i.e., quantum) particles. It was first proposed in the 1980s to improve computational modeling of complex quantum systems. Since then, there has been some progress in building small-scale devices processing quantum information, but significant technical advances are still needed to achieve a large-scale functional and practical quantum computer.

In contrast to classical computers whose computing powers are based on the mechanical, thermodynamical, or electromechanical features of their substrates, quantum computers (QCs) are built using super-conductive qubits, trapped ions and photons, and other physical mechanisms. As such, QCs are inherently susceptible to external disturbances and therefore somewhat error-prone—a concern that was not pointed out during this session but has been hotly debated in the computing world. But if these and other challenges are addressed effectively, QCs have the potential to handle highly complex and specific modeling problems with exceptional efficiency and accuracy.

Although they will not replace classical computers (much of what we do with classical computers will not benefit from quantum computing), QCs’ ability to handle otherwise intractable problems opens the door for use cases where classical computers would fail. Examples relevant to drug discovery and development involve quantum chemistry, such as the accurate simulation of atoms and molecules, combinatorial optimization of, e.g., molecular structures and interactions, and machine learning. The rationale here is that a model of computation with built-in quantum mechanics will be best suited to simulate quantum systems. Here are some examples:

  • The properties of molecules or atoms are guided by quantum mechanics. Finding exact solutions to these extraordinarily difficult quantum calculations using classical computers is intractable even for very small molecules, such as methane (CH4); modeling the exact properties of this molecule would challenge even the most powerful super-computers. Current methods therefore use ad hoc approximations and empirical fits—an approach that is unreliable and, at best, applicable to only very specific conditions. Quantum computers, on the other hand, provide an efficient and accurate alternative for simulating these systems.
  • The analysis of molecular structure and interactions, which falls under so-called “combinatorial optimization,” requires the selection of a subset of applicable features—a task that easily becomes unmanageable for large problems using classical computers.
  • Machine learning is an area of application for quantum computing that has emerged only within the last year or so. Two important contexts where machine learning can benefit from the use of QCs is supervised learning in data classification and as a sampling device. In the latter case, QCs can generate quantum states that are otherwise impossible to describe with classical computing approaches.

Metabolomics-Based Drug Discovery Using Cognitive Computing Tools

Metabolomics is an emerging field that can be used as a drug discovery tool. Just like genomics and transcriptomics looks at the genome (or genes) and RNA expression, respectively, metabolomics captures the metabolites that result from the processing of molecules through metabolic pathways using mass spectrometry methods. The metabolome is highly dynamic and responds to environmental influences, making it closely related to the phenotype. For this reason, metabolomics is also often referred to as “functional genomics.”

Metabolites are essentially small-molecule biomarkers of processes and actors. To understand the underlying molecular mechanisms of various disease processes, researchers at the Scripps Center for Mass Spectrometry and Metabolomics are conducting global profiling experiments by looking at all metabolites simultaneously. AI is then used at the data processing and analysis level to search through the scientific literature and other databases and identify the genes and proteins related to metabolites with key disease activity, their biological function, and how the disease phenotype can be treated. Conversely, another approach is to start with the cognitive processing tool and leverage the scientific literature to predict the expected metabolome of different diseases.

Use Cases for AI in Clinical Trials

AI can find various applications on the clinical trial side of drug development, including process automation, “chat,” content generation and quality control, and prediction and measurement.

Process automation is one of the most accessible areas for many organizations. It’s relatively easy to identify and implement the processes that benefit from automation, and the use cases are typically very narrow. However, an exciting category of AI use cases is the “chat” use case to improve the clinical trial experience of all stakeholders involved. It can help meet information needs in real time and gather data on patient engagement and experience as well as the experience of site and investigator staff participating in a trial.

SPEAKERS AND PANELISTS

Scott Spangler, Distinguished Engineer, Chief Data Scientist for Life Sciences, IBM Watson (Chair)

Yudong Cao, Co-Founder and CTO, Zapata Computing

Erica Majumder, Postdoctoral Research Fellow, Scripps Center for Mass Spectrometry and Metabolomics

Craig Lipset, Former Head of Clinical Innovation, Pfizer

Among the many applications around content generation and extraction, natural language processing that turns unstructured data (such as from electronic health records) into meaningful content is among the most potent ones. More powerful than that, however, might be the ability of AI to power digital diagnostics (i.e., to identify appropriate patients for a research study) or digital biomarkers. Most endpoints used in today’s studies are very fragile and dated. Taking advantage of new types of sensors and combining all available data to identify new endpoints is an exciting opportunity for AI. Protocol development (design, optimization, and automation) is another example where AI-powered processing has great potential. Protocol design is data-intensive, and, currently, too much of the diverse data used to inform design requires manual curation, interpretation, and aggregation. Here, AI tools can support smarter protocol designs, and, enabled by the move towards a common language as to what a protocol should include (such as the recently developed Common Protocol Template), can accelerate the process significantly.

Final Thoughts

Although AI and the various tools it offers are not magic bullets that solve all issues slowing down the pace of drug discovery and development, it holds much promise for the future of healthcare. While very much a work in progress, AI tools provide clinical researchers with the ability to use large amounts of diverse data put together in smarter ways. To ensure AI’s long-term role in preclinical drug development and discovery, however, it is critical to address intellectual property issues and issues related to data privacy, accessibility, and sustainability as well as the ethics of applying AI techniques to sensitive medical information.

Another common theme in discussions of AI applications in drug discovery is that these tools should not simply be black boxes that provide answers; they require a partnership with the scientist to make discovery work. The next generation of AI experts in healthcare will need broad knowledge in analytics, algorithm coding, and technology integration.