Recently, I was fortunate to chair a panel discussion where the potentials of Artificial Intelligence (AI) and Machine Learning (ML) were fiercely discussed. AI/ML seems to be talked about everywhere at the moment, heralded as a solution for many challenges in pharma and healthcare (and indeed many other industries). But, does AI/ML really warrant this level of excitement and hype?
AI/ML leaders in advanced statistics, NLP, deep learning, neural networks come together
One thing that is often overlooked in these discussions is that AI/ML is a broad group of diverse disciplines, not a singular item. AI/ML includes advanced statistics, Natural Language Processing (NLP), deep learning, neural networks, and many other mathematical and computer science specialities. The event I attended was a Linguamatics-hosted dinner. The panel of five experts covered a broad span of expertise across different AI/ML disciplines, and ranged from academia to small and large pharma:
- John Brimacombe, Senior Director, Linguamatics an IQVIA company; a serial entrepreneur in experienced investor, focused on technologies concerned with NLP, knowledge representation and machine reasoning
- Matt Crawford, Senior Manager, Biomedical information services at Pfizer; with a passion for high-throughput analysis of critical data from text for drug discovery
- Stuart Murray, Research Fellow, Informatics; Agios Pharmaceuticals; plays a key role in developing knowledge-driven integrated data analytics for research and clinical programs
- Jimeng Sun, Associate Professor of College of Computing at Georgia Institute of Technology; his research focuses on health analytics and data mining especially deep learning methods and large-scale predictive modelling systems
- Cao (Danica) Xiao, Director of Machine Learning at Analytics Center of Excellence of IQVIA; develops ML and DL models to solve real world healthcare challenges
The attendees came from biotech and pharma, from different areas within the drug discovery and development pipeline, including commercial and medical affairs. The breadth of experience allowed a wide-ranging discussion to evolve, from the likelihood of general intelligence to specific use cases showing value of NLP, an AI technology, in pharma research and development.
From the expert panel, the general sense was that we are still very early in our exploration to realize value of AI/ML. One discussion point was on trust: how can researchers, clinicians and patients know whether they can trust the conclusion provided by an AI/ML model? Jimeng Sun pointed out that the quality of the data is critical. People can trust algorithms if these are supported by sufficient data, and if that data is clean and reliable. Today, the key is providing the best data for clinical decision support. Danica’s team is working on machine learning models for some specific disease diagnoses; she explained that AI/ML may be able to replace some doctors one day, but that is years away. And there are still many barriers to overcome. Good ML models need training data, annotated data with both positive and negative outcomes.
Natural Language Processing in Drug Discovery Research
Stuart Murray and Matt Crawford both gave some specific examples of the value of NLP, an AI technology, in pharma research.
Matt discussed using NLP to find emerging patterns from literature and patents, and gave us a new definition: Deep Language. The process he discussed requires very flexible querying tools to iterate and refine queries, and involves:
- Trusting that scientists are going to describe what you’re looking for in a semi-consistent manner
- Examining enough samples to identify the key phrases and grammatical constructs
- Building queries using a set of test documents in an iterative agile manner
- Using entity normalization to group results and find trends
Matt showed an example, looking for novel discoveries. He used NLP to pull out patterns around CAR-T cell therapy, as a training example to find all the ways people describe this potentially promising therapy. The results were used these to create a “discovery flag” query, combining adjectives such as promising, breakthrough, significant, future, novel, with nouns such as treatment, paradigm, therapy, frontier, development. Using this query pattern, he was able to pull out other trends of potential future interest, including circular RNA, adeno-associated virus (AAV) vectors (for gene therapy), and some long non-coding RNAs.
Stuart touched briefly on a number of use cases for Linguamatics NLP and broader AI/ML at Agios, from discovery, development (e.g. clinical safety) and post-approval. Stuart talked in more depth about using AI/ML in Agios’ target strategy: “One of the things we love doing at Agios is disrupting the way the industry does things. Years ago, we began thinking: ‘Are there faster ways of getting compounds into a pipeline?’ And what we started looking at was chemical genetic screens.” Rather than looking at chemical genetics in the traditional way Stuart took a much more focused approach, only looking at metabolic genes and pulling out compounds with high-quality evidence (IC50, activity, cell potency, etc) using Linguamatics NLP to search large textual data sources. One candidate from this research is now in the clinic; Stuart estimates this saved Agios at least 3 years of work from initial target to Investigational New Drug application.
Our journey with AI/ML continues
These examples, and the overall panel presentation triggered many questions from the audience, generating an exciting discussion. Overall, as one attendee said, when you are presenting AI/ML outcomes to the business, no-one cares about technology, they care if they can trust the results. And, it’s important to ensure that AI/ML is addressing the correct problems to solve; too often AI/ML is brought into an organization to tick a box, rather than because these technologies are the best way to solve a challenge. However, as the capabilities and scope of AI/ML technologies become more broadly understood, these issues are likely to decline. One final note, the panel agreed that while the wave of augmenting human intelligence has taken off in the last 5 years, we are still a long, long way from true general intelligence. Right now, we are on an exciting journey, realizing many potentials across healthcare for value from AI/ML.