‘We don’t know how to analyze data, so let’s make a dashboard’ – Leonardo Rodrigues, BERG
The pharmaceutical industry faces serious challenges. It takes 15 years and 2 billion dollars on average to develop a new medicine, and failures are costly. If the development time of new drugs could be reduced with modern data-driven approaches, one can reduce the cost of drug development and help patients faster. Machine learning has revolutionized the modern digital economy through industries ranging from e-Commerce to the promise of self-driving cars. Machine learning is not new to the Pharma industry, but its application to drug discovery has been limited. Every step of the drug development pipeline could benefit from novel Machine learning approaches. It’s been shown that genetic evidence nearly doubles the probability of success for a new drug. It takes on average 5,000 molecular assays to find a lead drug candidate, so ‘lead optimization’ by reducing the number of experimental assays could enable early drug discovery to be more agile, save cost, and focus more time on finding novel targets. Finally, drug candidates that make it to clinical trials often don’t make it to market due to efficacy failures, and these clinical trial failures are extremely costly.
It’s with the above context that ‘Artificial Intelligence in Pharma’ conferences are starting to appear all over the place, including the 2018 ‘AI Pharma Innovation in Drug Discovery’ conference in San Francisco that I attended. ‘Artificial Intelligence’ is a term that I believe is being overused now, and everyone defines ‘AI’ differently. ‘AI’ in this conference is predominately used as a synonym for Deep Learning. Here are my top five takeaways from the conference:
1) It’s all about the data
‘If the data is not easily accessible, all AI initiatives are destined to fail’ remarked Leonardo Rodrigues of BERG on the opening day of the conference. Modern deep learning approaches are data-hungry, and Pharma isn’t great historically at making data across the discovery pipeline easily accessible to R&D researchers, encouraging data silos and analysis results (in the form of excel spreadsheets and powerpoints!) that disappear into the ether of mapped network drives. It was strangely comforting to see that these problems are still endemic across Big Pharma, and there’s universal agreement that data needs to be widely accessible and reliable for Deep Learning initiatives to be successful. Hugo Cuelemans of Janssen commented during his Virtual Drug Screening talk that although Janssen has found Deep Learning to be extremely powerful for predicting compound activities, the actual approach and specific neural network architecture was less important than having as much historical, clean assay data as possible. It’s a difficult problem, since the large variety of pharmaceutical data is significantly more complex than other industries, such as data from e-commerce sites.
2) The Pharma Machine Learning startup field is getting crowded
The amount of new startups focused on using machine learning and deep learning to solve specific drug discovery problems is staggering. If you didn’t know better, one would think AI drug discovery is completely a solved problem! Imaging provides the lowest hanging fruit in applying deep learning technology; as such many of the early Pharma Deep Learning initiatives are specific to imaging applications. This was reflected in the startup world for a while (Arterys, for example) as well. Today, there are a plethora of new startups applying deep learning methods to a wide variety of applications including immunotherapy modeling (Kadmon), pharmacokinetics (Numerate), biological network mapping (e-Therapeutics), among other applications.
3) Pharma is beginning to appreciate the value of Transfer learning
One of the pleasant surprises from the conference is that there is a general understanding of the value and power of Transfer Learning. Promising approaches in machine learning publications are normally trained and validated with public data sets, and therefore some transfer learning approach would be necessary to test such approaches on proprietary pharmaceutical data. Janssen went even further, suggesting that Pharma companies could form consortiums to share deep learning models, but not proprietary enterprise data. This could be made possible through transfer learning, and might be a great strategy to help accelerate the adoption of machine learning in drug discovery.
4) There’s anxiety that Deep Learning is over-hyped and simply a fad
Despite the general excitement surrounding Deep Learning at the conference, there is still a lot of anxiety and skepticism surrounding the ability of the field to make a significant impact in drug discovery. This is due to several reasons. One reason deals with recent high-profile failures in the field, such as IBM Watson Health’s partnership with MD Anderson, the DeepMind collaboration with the NHS in the UK, and a few negative recent publications such as this medium article (https://medium.com/the-ai-lab/artificial-intelligence-in-drug-discovery-is-overhyped-examples-from-astrazeneca-harvard-315d69a7f863) that highlights examples from AstraZeneca, Harvard, and InSilico Medicine. The other reason is that many veterans in the industry are skeptical of Deep Learning because these same veterans have seen Big Pharma adopt cutting edge technologies such as high throughput screening and massively parallel genome sequencing early, yet the costly and lengthy drug development pipelines have only worsened. The anxiety is understandable, but representatives from companies such as BERG, Genentech, and Roche mentioned that Deep Learning is here to stay and Pharma has to modernize due to the ‘disruption’ the field has caused in other industries such as self driving vehicles.
5) No one addressed deep learning model deployment
The conference addressed a wide array of topics which were broadly classified by the conference organizers into the following categories:
- The fundamentals of setting up a successful machine learning approach in drug discovery
- Matching AI & Machine Learning to Biopharma’s challenges
- Enhancing drug discovery productivity with AI & Machine Learning
- Practical applications of AI & Machine Learning in drug discovery
- What is the future of AI & Machine Learning in drug discovery
None of the talks addressed the practical considerations of how to deploy deep learning models to production. I saw many approaches and initiatives but I didn’t see any success stories; meaning I didn’t see any examples of how deep learning insights were used to make specific decisions in drug discovery. I did not witness any ideas of how to turn deep learning models to data ‘products’ that help scientists and managers to make recommendations. Although it is the early days of machine learning adoption in Pharma, there are valuable lessons learned from other industries in building machine learning models, (one of my personal favorites is from a Netflix alum https://medium.com/@xamat/10-more-lessons-learned-from-building-real-life-machine-learning-systems-part-ii-93fe7008fa9) many of which are only sensible and obvious with experience. I am emphasizing this point because I firmly believe practical deployment considerations are a bottleneck for the success of Deep Learning initiatives in the Pharma industry. (Would Amazon be Amazon if their product recommendations were flyers they mailed once a month rather than easy to use and consume recommendations from your personalized home page?)
Philadelphia Physicist 