The Drug Discovery Boom Is Coming, and Sooner Than You Think


In 1997 a remarkable event caught everybody’s attention – the then champion of the world Garry Kasparov lost a tournament to a supercomputer Deep Blue. It was called “a beginning of a new era of computers” by many and now it seems that time keeps justifying those loud statements…


Figure 1. ACM Chess Challenge 1997 (photo credit:

A human’s brain simulation

Being a sub-set of artificial intelligence (AI), machine learning involves algorithms allowing computers to autonomously learn from input data. A fundamental distinction from “usual” software programs, such as Photoshop or, say, Excel, is that in machine learning computers don’t have to be explicitly programmed but can change and improve their algorithms by themselves.

The history of machine learning goes back to the 1950th. The first learning program was created by Arthur Samuel in 1952 and it was the game of checkers. Implemented in an IBM computer, the program was able to improve itself the more it played studying winning moves and incorporating them into the next rounds.

Five years later Frank Rosenblatt designed the first neural network for computers – the perceptron, which simulated the thought processes of the human brain. And just a decade later the “nearest neighbor” algorithm was written, which was a conceptual step towards pattern recognition technology.

Today, machine learning algorithms enable computers to “see” and distinguish objects and text in images and videos, discover and categorize real-world things, “communicate” with humans, drive cars on auto-pilot, write and publish sport match reports, and … help discover new drugs.


Can computers suggest new drug candidates?

In a pursuit of decreasing the cost and time of drug discovery and more accurately predicting the structure-activity relationship of early drug candidates, scientists developed specialized mathematical models and computer programs able to conduct “in-silico” drug discovery. In this approach, the available structural information about target proteins is used to conduct a virtual screening of numerous chemical structures and identify hits which better fit the target in terms of energy of interaction and other calculated functions.


Figure 2. Protein-ligand docking (photo credit:

Although promising, standard in-silico methods are still limited and not accurate enough to substitute costly and time-consuming real-world experimental screening and trials, because of their explicitly pre-programmed nature and pre-determined models, used for calculations. Essentially, the models are limited to certain level of abstraction and they can not improve unless scientists update them manually.

This is where machine learning algorithms and new drug discovery startups come into play.


Startups bringing the new drug discovery boom

Atomwise-Logo-ColorAtomwise: Recently, a group of scientists at the University of Toronto created a machine learning algorithm that they hope will revolutionize the way pharmaceutical drugs are discovered.

They founded a health tech startup Atomwise offering a solution which can help researchers develop the next generation of drugs and do it faster and cheaper than ever before.


Figure 3. Atomwise artificial intelligence based drug discovery (photo credit: Atomwise).

The algorithm Atomwise developed is similar to the Deep Learning Neural Networks used by AI startup DeepMind, acquired by Google recently for $628 million. The algorithm teaches itself complex biochemical principles and the factors that are ultimately the most predictive when it comes to the effectiveness of a drug.

Our system takes into account not a dozen or two dozen, but thousands of factors at the same time and combines them in complicated and nonlinear ways. It’s like having a virtual super-intelligent brain that can analyze millions of small molecules and potential interactions in days instead of years,” said Alexander Levy, chief operating officer at Atomwise.

The company’s machine learning algorithm is acting similarly to how computers go about image recognition, which is a unique feature of this approach. Levy says their system has devised some unintuitive methods for understanding what small molecules will properly latch onto a biological target.

To date, Atomwise has raised $6 million to advance AI for drug discovery and launched more than a dozen projects to find cures for both common and orphan diseases. The company is collaborating with IBM to find a cure to Ebola and with Dalhousie University in Canada to search for a measles treatment. The startup studied 8.2 million small molecules to find potential cures for multiple sclerosis in a matter of days. Besides, Atomwise is already partnering with a pharmaceutical giant Merck to explore the frontiers of using AI for drug discovery.

twoXAR company logoTwoXAR: Palo Alto based TwoXAR, founded in 2014, has recently raised $3.4 million in a seed round led by a tech investor Andreessen Horowitz. TwoXAR’s solution is DUMATM Drug Discovery platform able to evaluate large public and proprietary datasets to identify and rank high probability drug-disease matches in minutes rather than years.

twoxar_artificial_intelligence platform

Figure 4. TwoXAR artificial intelligence based drug discovery platform (photo credit: TwoXAR).

The company has already tested their technology on more than twenty diseases and is now actively collaborating with academic researchers at the University of Chicago and Michigan State University to further develop the platform. Being a part of the elite Stanford-backed “StartX Med Program”, TwoXAR is collaborating with some unnamed biopharmaceutical organizations.

berg_health_logoBERG Health: A Berg company claims a very ambitious plans on using AI for revolutionizing cancer treatment and first promising results were reported recently by BBC News. BERG’s Interrogative Biology® platform is a unique combination of adaptive-omic biological data and advanced AI, machine learning algorithms. This allow them to stratify patient populations by phenotype to build predictive models.

artificial-intelligence-platform company

Figure 5. BERG’s Interrogative Biology® platform (photo credit: TwoXAR). Full-video is available here.


Concluding Notes

Recent news on using supercomputing and AI-based algorithms in drug discovery sparked in Spain, China, Great Britain and other regions. A clear rising trend is now developing and with the first FDA-approved example of “AI-born” drug, an investment stampede will follow.

Now, it is time for leading pharmaceutical companies and drug discovery CROs to start reflecting on how they can leverage new technologies and adjust to the coming boom in drug discovery. Those who will adopt earlier and better will dominate the future market.


andrii_buvailoAbout the guest author: Andrii Buvailo is a Ph.D in Chemistry and Head of E-commerce at Enamine Ltd – a leading supplier of building blocks and screening compounds for drug discovery research. Previously, he worked as a project manager and later as a director forYUNASKO – an energy storage startup, a developer and a licencor of supercapacitor technology. He was also involved in various research projects in Ukraine, United States, Germany and Belgium.

He is writing occasionally for science and technology blogs, including (a blog about drug discovery trends), (technology commercialization blog) and EnamineStore Blog (chemical products for drug discovery). This article was first published on by Andrii-Buvailo. All statements made, opinions expressed, etc. in his articles only reflect his personal opinion.

Credits for edits: Pankaj Mishra

Featured Image Credit: Bigstockphoto

Comments (2)

  1. While the possibility exists for computational screening for drug candidates (particularly drugs that fit into known pockets of a protein complex for their efficacy), there are still some troubling questions surrounding this avenue of investigation.

    It is of tremendous benefit if some well-crystallized drug-protein coordinates are available for a starting point for most analyses (i.e. obtaining a set of internal coordinates for the main “players” to be interrogated in the simulation- sometimes this includes water). This lends a certain amount of credibility to the simulation since the complex is demonstrated to exists under laboratory conditions. However, how well this crystallized structure represents in vivo conditions remains a question.

    Should such internal coordinates be unavailable, “”docking” programs can generate a likely set of internal coordinates for use in the simulation- presumably enabling the researcher to explore a set of energy minima and possibly find a significant orientation for the drug-target complex. Here again, questions remain about how physiologically significant such a complex might be.

    Under docking conditions, I have been able to find docking partners with proteins that have little significant interactions in vivo. This is a well-known phenomenon in vivo and anyone who has done an ELISA or any type of immunostain knows about “non-specific” binding. Sometimes such binding can present a real problem in these experiments. Proteins can almost always find regions in other proteins to bind, particularly if other species are present to foster such binding. So how can we tell real binding from non-specific, reansitory binding?

    One answer may well be binding thermodynamics: dG=dH-TdS. This is the Gibbs free energy of binding and it’s value is related to the equilibrium constant, a true measure of thermodynamic stability. Yet measuring dS (entropy change) on binding remains challenging- particularly for hydrated systems (i.e. almost every system of interest). How about enzyme-substrate binding? The enzyme does not bind the substrate, it binds the transition state. This still presents a significant barrier since most programs only accept standard bond lengths and angles that are QUITE UNLIKE the transition state. What if a drug is stabilized in the same way in some systems? I’m not aware of any simulation programs that can distinguish this effect. Indeed, it would be difficult to incorporate. Many such simulations are designed to find stability in terms of energy and complex structure. This may not be physiologically relevant in all cases.

    Nevertheless, this technology may still hold a significant degree of value in time. Transitory states, a better understanding of how kinetics and thermodynamics play a role in complexation as well program development that is more flexible in terms of quantum mechanical deviations from ground states (such as occur in transition states) will only help our understanding and augment our powers of prediction for valid target selection.

  2. your information is really new to me and thus i got more relevant information and interesting news from your blog , it is really awesome.

Leave a reply

Your email address will not be published. Required fields are marked *