Autumn is Here: Changes to the Drug Discovery AI Landscape

The month of September brought several big announcements for AI in drug discovery and October is shaping up to be more of the same as AI startup Insilico Medicine just announced a deal worth up to $200 million with their collaborator Jiangsu Chia Tai Fenghai Pharmaceutical.

Early in September a landmark paper was published in Nature Biotechnology by Insilico Medicine. The paper, which included corresponding code, detailed how researchers used deep learning to discover inhibitors of DDR1 in 21 days, leveraging Insilico’s proprietary platform. Their approach, called GENTRL, coupled a predictive model with a generative one to discover novel chemistry with high efficacy. Nearly a week later Insilico secured it’s Series B round with $37 million in funding.

Around the same time another startup, Atomwise, whose partners include Merck, Abbvie, Lilly and Bayer, announced a deal with the Jiangsu Hansoh Pharmaceutical Group potentially worth up to $1.5 billion. Atomwise’s technology, termed AtomNet, offers a more structure-based drug design approach using convolutional neural nets.

Not to be outdone, industry veteran software company Schrodinger also announced a deal with Astrazeneca, although financials were not disclosed. Schrodinger has been partnering with biotech companies and building itself a stake in therapeutic assets over the past few years. Schrodinger has combined their already strong computational chemistry offering with machine learning techniques.

Does all this mean that the hype around AI and machine learning in discovery is real?

Direct experience over the past few years has made me believe the hype is real, but the lack of publication with accompanying laboratory results has tempered expectation throughout the industry.  Most papers are academic in nature and describe methods for the featurization of molecules (converting structures to appropriate input…think chemical fingerprint generation) or adapt popular machine learning approaches to drug discovery applications.  While these publications are advancing our understanding of capabilities and limitations, many of them operate on public datasets such as ChEMBL, Tox21 and SIDER. The GENTRL paper described above takes us from theory to results and I believe it is the first of many forthcoming papers.

As in most scientific and technological advancements, there tends to be a stark division between the champions and the skeptics, with the truth lying somewhere in between.  As a community we need to appropriately identify when AI methods can be most effective in the discovery of novel small molecules. Success will depend on the chosen application of the technology and in the next blogpost I will be exploring where we should look.