BenevolentAI: From Machine Learning Startup towards New Drug Producer

BenevolentAI is using machine learning to gather insights from the body of pharmaceutical literature in order to produce new potential drugs, but can it keep top talent?

Machine learning is not a new entry to the pharmacological field. Scholarly articles explaining the use of machine learning techniques for pharmaceutical data analysis may be found dating back decades. The traditional paradigm has been to use a few techniques on a particular dataset and publish a paper comparing results to prove the efficacy of a particular technique [1]. BenevolentAI turns this paradigm on its head by using the vast and growing volume of pharmacological literature as the basis for datasets on which to use a variety of machine learning techniques to discover new drugs [2].

To produce a new drug, tens of thousands of candidate drugs are needed. The Pharmaceutical Research and Manufacturers of America (PhRMA) has created a brief video which explains the process:

Most pharmacological research does not result in breakthrough discoveries. PhRMA illustrates the rarity of a new drug discovered from the magnitude of potential drugs here [3]:

Not shown in the figure above is a depiction of how much research is required preceding drug discovery, but that is where BenevolentAI operates today [4]. Instead of discarding research which does not immediately lend itself to the further development of a particular drug, BenevolentAI harnesses both clinical trials data and academic papers to produce new potential medicines [5]. The final products are potential new drugs. Machine learning is not just important for BenevolentAI, it is essential.

Natural prerequisites for the development of such technology are corporate knowledge of the pharmacological literature as well as the talent required for developing new machine learning techniques which can digest this knowledge. Talent for developing machine learning algorithms is scarce and in high demand, as evidenced by the compensation machine learning scientists may receive [6]. In 2016 BenevolentAI recruited Dr. Jerome Pesenti, the former chief scientist of IBM Watson [2]. In the current environment of fierce competition for data scientists, retaining talent like Dr. Pesenti will be challenging.

Since BenevolentAI is still considered a startup, its future direction is not completely clear, though some reporting does give us a hint of where the firm is headed. In an interview with Business Insider, BenevolentAI’s founder, Ken Mulvany “predicted BenevolentAI would be selling its own drugs ‘in the next four years’” [4]. To develop its own capabilities as a platform for developing drugs and for analyzing previous research, BenevolentAI will inevitably have to expand its operations to perform these new functions. Both the pharmaceutical industry and the emerging machine learning landscape are complex. Maintaining harmony between the two aspects of this business will require sustained effort.

The company already has made progress towards this goal. It has produced “24 drug candidates in just four years” and has sold drug targets to a US company [4]. Navigating the regulatory environment of pharmacological development in both the US and UK is increasingly challenging [7]. One recommendation may be to use the machine learning talent embedded in BenevolentAI’s workforce to ensure compliance with increasing regulatory burdens.

As BenevolentAI works towards new drug solutions and continues to refine its own processes, it faces competition from larger industries and similar startups. In the same year Jerome Pesenti left IBM, his previous employer teamed up with Pfizer to work towards drug discovery [3]. Exscientia, another UK based startup, also uses machine learning for drug discovery and has partnered with Sanofi [3]. In response to these challenges, BenevolentAI should capitalize on its strength of harnessing the power of untapped information available in the literature.

As we have learned that the power of machine learning grows with the quality and quantity of information on which it is trained, BenevolentAI should consider expanding the base of literature on which it trains beyond articles published in English. An emerging technique known as neural machine translation can be used to translate written language [8]. For example, if BenevolentAI were to add Chinese research into its corpus, it could discover new insights at the intersection of research previously disconnected by the language barrier.

Key questions moving forward may include the following. How can a firm continue to recruit top talent when many of the professors that would teach these techniques have been drawn away from academia by extraordinary salaries at AI-based firms? Additionally, since drug development can take years, how can a firm retain top talent as it expands to perform new functions? Aside from neural machine translation, what other techniques might BenevolentAI consider incorporating into its machine learning arsenal?

(735 words)


[1] “Drug design by machine learning: support vector machines for pharmaceutical data analysis.” R. Burbidge, M. Trotter, B. Buxton, S. Holden. Computers & Chemistry. Volume 26, Issue 1, December 2001, Pages 5-14.

[2] “This AI unicorn is disrupting the pharma industry in a big way” Mederios, J. 2018. WIRED Health.

[3] Machine Learning Drug Discovery Applications – Pfizer, Roche, GSK, and More. 2018.

[4] A British tech unicorn is trying to cure Alzheimer’s and ALS with artificial intelligence. Gosh, S. 2018. Business Insider.

[5] Benevolent AI. 2018.

[6] Tech Giants Are Paying Huge Salaries for Scarce A.I. Talent. Metz, C. 2017. The New York Times.

[7] Richards, Natalie and Ian Hudson. “UK medicines regulation: responding to current challenges” British journal of clinical pharmacology vol. 82,6 (2016): 1471-1476.

[8] OpenNMT FAQ. 2018.

[9] BenevolentAI: Revolutionizing drug discovery using Artificial Intelligence. 2018. Patyal, S.


Betting on machine learning to save lives


Volition: Crowdsourcing Innovation in the Beauty Industry

Student comments on BenevolentAI: From Machine Learning Startup towards New Drug Producer

  1. Thanks for sharing this interesting article on machine learning in an important industry. It sounds like BenevolentAI offers a novel technology that could dramatically speed up the drug discovery process in a cost-effective manner. I wonder if they would benefit more by leveraging their technical expertise to partner with large pharmaceutical companies who have the know-how and infrastructure to turn a computer-generated drug candidate into a reality. I also wonder if they are using their machine learning technology to develop completely novel drug candidates or to repurpose previously used medications to treat new conditions, which has recently become a hot topic in medicine. Ideally, they would be able to do both. My concern, however, is that they are limiting the possibilities for drug candidates if their predictions are based only on clinical trials data and academic papers, both of which would mostly include drugs that have produced some finding worth publishing. They would have to incorporate pharmacologic and biochemical principles and constraints into their algorithm for it to be most effective.

  2. I think the question you posed around talent retention for these very technical roles that are in demand does seem important. Especially because some of these technical problems they would be working on are not simple and would probably require continuity. Because the job demand in AI applications is exploding, if they want to retain their top talent they have to incentivize them to stay beyond giving them a “difficult problem to solve” because it seems they could find a different problem to solve that is just as difficult for more pay at a different company. For pharmaceutical companies, I think they could work at aligning the technical work to the larger mission of creating drugs that presumably solve some medical problem and reduce suffering in the world. This could be a compelling reason for certain candidates. In addition to that, it seems like their machine learning work is a bridge between academic life and corporate life. For researchers who enjoy their academic work but simply want to move towards the application of their theories and higher pay, this job role would be a good fit. I think the problem of good talent changing jobs may actually be a broader trend in our generation. In the AI community it is more pronounced because the pool is smaller than other fields. From that lens, I think that companies really need to define their mission in a way that is compelling because when someone has to choose where to work I think they would rather pick somewhere that is actually trying to make a difference in the world vs just trying to make a profit.

Leave a comment