• Sustainability
  • DE&I
  • Pandemic
  • Finance
  • Legal
  • Technology
  • Regulatory
  • Global
  • Pricing
  • Strategy
  • R&D/Clinical Trials
  • Opinion
  • Executive Roundtable
  • Sales & Marketing
  • Executive Profiles
  • Leadership
  • Market Access
  • Patient Engagement
  • Supply Chain
  • Industry Trends

Insights: How Biopharma is Applying AI and Machine Learning

Article

Pharmaceutical Executive

Gabe Musso, Chief Scientific Officer of BioSymetrics, talks about insights around how AI and ML are being used in biopharma.

Gabe Musso

Artificial intelligence (AI) and machine learning (ML) have dominated many industries, and the biopharmaceutical industry is taking a keen interest in applying novel technologies towards improving the efficiency of R&D and decision-making. In this interview, Gabe Musso, Chief Scientific Officer of BioSymetrics, will discuss insights around how AI and ML (machine learning) are being used in the biopharmaceutical industry.

Moe Alsumidaie: What are some of the most common challenges that biopharmaceutical executives face?

Gabe Musso: We primarily see challenges in estimating and evaluating ROI when engaging in data science and machine learning projects. Every executive in the pharmaceutical industry has seen the landscape change to the point where they know that AI and ML approaches are going to impact their future business. However, there is still much noise; identifying what is likely to be a successful implementation of AI/ML and how it will impact your business practices is undeniably tricky. Most executives approach it by starting with small projects, either on the commercialization side or research side, trying to understand, in a very measured way, how AI or ML can impact the products they are seeking to get to market. However, while this is an approach that seems sensible, it can often mask the true value of machine learning, which really isn't seen until applied on much larger and more encompassing datasets. It does require multiple data sets and cross-disciplinary engagements to see the benefit in ML.

You cannot avoid the fact that data are messy in any aspect of healthcare. For any team, there is a lot of manual processing. 60-70% of the time is spent cleaning data, massaging data, filling in the blanks, and just getting your data to the point where you can then start interpreting it. People do not realize the decisions you make in terms of cleaning data and preparing that data impacts how the machines interpret it. In other words, what you choose to put in the curriculum affects how the student learns. That is the “Ah-ha! Moment” we are experiencing in applying AI to biopharma today.

Alsumidaie: What sort of data are we looking at here?Musso: There is no dataset that cannot be subject to a learning process-anything from appointment or scheduling information to diagnostic and prescription data on the operational side. R&D data typically include medical imaging, chemical compounds, experimental results, genomics, medical history, etc. All of these are very different data types and typically managed or maintained with different levels of rigor or standardization. The goal is arriving at a level where machines can start to interpret multiple data sets in combination. The increasing richness of the data landscape presents an opportunity in terms of improving care and operations but also poses a substantial technical challenge.

Alsumidaie: Where would AI be implemented, at what stage, and how can that inform the decision making for therapeutics?Musso: Pre-clinical stages appear to provide the best business value. The resources required to screen through a large library of compounds or diagnostics get compounded if you get it wrong based on an informed hunch. ML provides the opportunity to remove our human perspective and allow a more unbiased decision-making process.

Alsumidaie: What types of biopharma companies are implementing this type of technology in their R&D?Musso: Big pharma companies seem willing to take chances on pilots. There is a lot of noise in the space, and pilot projects help companies evaluate what is truly useful to implement within their organizations. In our experience, pharma and CROs will provide us with a data set with a goal very similar to a problem that they are working on, and then evaluate if the ML process can confirm or build confidence in their work. The scale is achieved when the process arrives at the same outcome but in one-third the time. In many instances, executives are seeking a ‘fail faster’ approach so that resources can be allocated to projects with higher confidence of success.

Alsumidaie: Talk about some of the success stories and failures you have experienced using AI.Musso: On the pharmaceutical side, we have had success implementing ML for prioritizing small molecules, for integrating different data, for streamlining the discovery and translation processes, and reducing the time for data science.

In terms of failures, the entire BioSymetrics framework was motivated by the failure of an early ML project building a diagnostic model for autism. We had access to hundreds of different MRIs and FMRIs for over 1,000 different autistic children and control patients; we wanted to build a more objective classifier for autistic patients. Initial confidence in our model began to erode as we dug in and realized the results were still very dependent on the MRI machinery and the technician who took the MRI. We assumed our models had treated those as confounders. Going back to the drawing board, we found that a lot of the decisions we made during data processing were influencing the quality of the model. It was not until we changed our data ingestion and normalization processes that the process became more reliable. Addressing this issue ultimately became the basis for our “Contingent AI” framework-AI that is contingent on decisions made upstream.

Alsumidaie: So ML and AI can essentially enhance the productivity of discovery. Can companies rely solely on AI and ML? Do they also have to incorporate the human component?Musso: There will always be a human element, particularly around data collection and patient care. Healthcare is an intimately human experience; we cannot discount human intervention and oversight. The next few years are going to be about finding that balance between how much we want the AI to play a role in the decision-making process and what remains human regulated. Just as the automotive industry is evolving from adaptive cruise control and lane assist to the prospect of self-driving cars, the healthcare community will need to have a similar progression. However, we sense that there will always be a driver behind the wheel.

Related Content