Roche: Improving Drug Discovery and Development with Machine Learning

Machine learning has been touted as a potential cure-all for high drug prices. A number of leading biopharmaceutical companies like Roche have made some large bets in using artificial intelligence tools to improve drug development, but it remains to be seen if machine learning is living up to its hype.

As artificial intelligence tools have continued to progress, the healthcare sector, which makes up almost 20% of the U.S. GDP, has been highlighted as a field that could be well-served by the capabilities of machine learning [7]. Within the larger healthcare ecosystem, biopharmaceutical companies in particular have been criticized in recent years due to the rapidly increasing prices of prescription drugs. Machine learning has the potential to significantly improve drug discovery and development and, hopefully, reduce prices in the long-term [4].

Machine Learning Applications at Biopharma Companies

Biopharmaceutical companies attribute the high price tags of their leading therapies to the significant costs of research and development, which have recently been estimated to average around $2.6 billion per treatment [7]. Given that it takes up to 15 years to bring a new therapy to market and less than 12% of drugs that enter clinical trials end up being commercialized, there has been significant attention in recent years to the inefficiencies in the drug development process [10]. In traditional drug development, companies must conduct basic research to identify and validate a disease target of interest. To create a therapy for the identified target, large numbers of drug compounds must be screened before a lead candidate is identified. This lead candidate must then be refined further and tested preclinically before the drug can be studied on patients. An overview of the drug development process is provided below [10].

A number of prominent technology companies have focused some of their resources on tackling inefficiencies in this space. Pfizer, a pharmaceutical company, has partnered with IBM Watson to process thousands of scientific publications and identify more robust targets during the discovery phase, determine novel combinations of drugs for improved efficacy, and optimize the selection of patients for clinical trials [1,2,5]. A number of smaller companies, such as Insilico Medicine and Exscientia, are attempting to utilize genomics and artificial intelligence tools for computational design of new drug candidates [1].

Roche’s Machine Learning Strategy

Roche, a multinational biopharmaceutical firm, is tapping into developments in machine learning in the short-term through collaborations with smaller biotech companies. One such startup is GNS Healthcare, which recently partnered with Roche subsidiary Genentech to use large amounts of patient data to uncover “new pathways, novel targets, and diagnostic markers that are better matched to individual patients [3].” The partnership with GNS Healthcare is intended to augment Roche’s capabilities in precision medicine so that Roche can develop innovative therapies that are tailored to the genomes of individual patients. Roche has also made a few strategic acquisitions of larger, more established healthcare companies like Foundation Medicine and Flatiron Health to compete in the longer-term [8]. Foundation Medicine has developed a number of companion diagnostics that can identify specific tumor mutations and pair patients with the appropriate cancer therapy [9]. Flatiron Health has developed a specialized electronic medical record (EMR) for cancer patients that Roche can leverage in designing more targeted clinical trials and precision medicines [9].

Overall, I think Roche’s approach to capitalizing on the capabilities of machine learning is prudent. As machine learning tools continue to advance, the data generated by Foundation Medicine and Flatiron Health can be a key strength for Roche in designing and developing personalized cancer therapies. In addition, by operating Foundation Medicine and Flatiron Health as independent subsidiaries, Roche allows both companies to maintain their leads in their respective segments of the industry and continue to partner with other companies and improve upon their capabilities. In the short-term, partnerships with early-stage companies like GNS Healthcare allow Roche to follow developments in the exciting field of machine learning and build experience understanding how these technologies can best be used in the drug development process [2,5]. As the field continues to evolve, Roche will be better equipped for future partnerships (and acquisitions) that can improve the efficiency of the drug development process.

Given that Roche’s bets on machine learning have been bold so far, I would encourage Roche to look to more incremental means through which machine learning can provide value to their organization. One area where machine learning already has shown promise is in large-scale image analysis [6]. Currently, pathologists spend a significant portion of their time going through histopathology slides from pre-clinical studies in order to make diagnoses and help understand the impact of therapeutics. Machine learning could be a useful tool in this area by alleviating some of the mundane tasks involved with analyzing preclinical data, allow pathologists to focus their resources on more value-add activities, and increase the pace of drug development.

Key Questions Remain 

  • How will the biopharmaceutical industry’s business model have to adapt if machine learning does end up taking over the process of drug development and discovery?
  • How will companies with machine learning capabilities build a sustainable competitive advantage in the long-term?

(798 words)


[1] Buvailo, Andrii. “How Big Pharma Adopts AI To Boost Drug Discovery.” The Rise of R&D Outsourcing in Pharmaceutical Industry | BioPharmaTrend. Accessed November 13, 2018.

[2] Fleming, Nic. “How Artificial Intelligence Is Changing Drug Discovery.” Nature News. May 30, 2018. Accessed November 13, 2018.

[3] “GNS Healthcare Announces Collaboration to Power Cancer Drug Development with REFS™ Causal Machine Learning and Simulation AI Platform.” September 28, 2018. Accessed November 13, 2018.

[4] Gustafsson, Claes. “Building Biology with Machine Learning.” GEN. October 31, 2018. Accessed November 13, 2018.

[5] Lowe, Derek. “Machine Learning’s Awkward Era.” In the Pipeline. August 08, 2018. Accessed November 13, 2018.

[6] Lowe, Derek. “Images of Machine Learning.” In the Pipeline. March 12, 2018. Accessed November 13, 2018.

[7] Mullin, Rick. “Cost to Develop New Pharmaceutical Drug Now Exceeds $2.5B.” Scientific American. November 24, 2014. Accessed November 13, 2018.

[8] “Partnering in a Digital Era.” Roche. Accessed November 13, 2018.

[9] Ramsey, Lydia. “We Spoke to the CEO of Roche Pharmaceuticals about How the Pharma Giant Became a Deal Machine in 2018.” Business Insider. August 05, 2018. Accessed November 13, 2018.

[10] “The Biopharmaceutical Research and Development Process.” Phrma. February 26, 2015. Accessed November 13, 2018.


Machine-learning in insurance is a sweet deal for customers


How Birchbox Stays Competitive using Machine Learning

2 thoughts on “Roche: Improving Drug Discovery and Development with Machine Learning

  1. It is very frustrating to know how it can take 15 years for a life-saving therapy to get to market in a process that seems to have so many opportunities for acceleration. Is it partly because it is hard to actually find an adequate number of viable volunteers for each phase of clinical trials? If so, can machine learning really help Roche find volunteers without access to external databases with information on potential volunteers?

  2. Interesting reading! I do not have much knowledge about the industry, but if machine learning can shorten the lead time of drug development to market, that would be very good. If machine learning could be used for drug development and discovery, I believe the biopharmaceutical industry’s business model needs to be changed to speed up testing and getting approval from regulators. Time to test drugs may not be able to be reduced, but companies need to be creative while maintaining health issues. At the same time, educating and working closely with regulators to influence them is important as approval tends to be a lengthy, burdensome process. Overall, I am hoping that machine learning could disrupt the industry to reduce the lead time without sacrificing the quality of drug testing/development.

Leave a comment