On a dark night, far from the city light, the stars of the Milky Way appear to be incalculable. But from any point, no more than 4500 stars are visible to the naked eye. In our galaxy, there are 100-400 billion of them, there are even more galaxies in the Universe. It turns out that there are not many stars in the night sky. However, even this number opens before us a deep insight … drugs and drugs. The fact is that the number of possible organic compounds with medicinal properties exceeds the number of stars in the Universe by more than 30 orders of magnitude. And the chemical configurations that scientists create from existing medicines are akin to the stars that we might see downtown at night.
Finding all possible drugs is an overwhelming task for humans, as is the study of the entire physical space, and even if we could, most of what was discovered would not correspond to our goals. However, the idea that miraculous drugs might lurk amid abundance is too tempting to ignore.
That's why we should use artificial intelligence that can work harder and accelerate discovery. So says Alex Zhavoronkov, who spoke at Exponential Medicine in San Diego last week. This application could be the largest for AI in medicine.
Dogs, diagnosis and medications
Zhavoronkov - CEO of Insilico Medicine and CSO Biogerontology Research Foundation. Insilico is one of many startups developing AI that can accelerate the discovery of new drugs and drugs.
In recent years, Zhavoronkov said, the famous machine learning technique - deep learning - has made progress on several fronts. Algorithms capable of learning to play video games - like AlphaGo Zero or poker player Carnegie Mellon - are of the greatest interest. But pattern recognition is what gave a powerful boost to deep learning when machine learning algorithms finally began to distinguish cats from dogs and do it quickly and accurately.
In medicine, deep learning algorithms trained on databases of medical images can detect life-threatening diseases with equal or greater accuracy than human specialists. There is even speculation that AI, if we learn to trust it, could be invaluable in diagnosing disease. And as Zhavoronkov noted, more applications are coming and the track record will only grow.
Promotional video:
“Tesla is already taking cars out onto the street,” Zhavoronkov says. “Three- and four-year technology is already transporting passengers from point A to point B at a speed of 200 kilometers per hour; one mistake and you're dead. But people trust their lives to this technology."
"Why not do the same in pharmaceuticals?"
Try and fail, over and over
In pharmaceutical research, AI won't have to drive a car. He will become an assistant who, paired with a chemist or two, can speed up drug discovery by scrolling through more options in search of better candidates.
The space for optimization and efficiency improvement is enormous, Zhavoronkov said.
Finding drugs is a painstaking and costly undertaking. Chemists are sifting through tens of thousands of possible compounds, looking for the most promising. Of these, only a few go for further study, and even fewer will be tested on humans, and of these, in general, crumbs will be approved for further use.
This entire process can take many years and cost hundreds of millions of dollars.
This is a big data problem, and deep learning excels at big data. Early applications showed that AI systems based on deep learning were able to find subtle patterns in gigantic data samples. Although drug manufacturers already use software to sieve compounds, such software requires clear rules written by chemists. The advantages of AI in this matter are its ability to learn and improve on its own.
“There are two strategies for AI innovation in pharmaceuticals that will provide you with better molecules and quicker approval,” Zhavoronkov says. "One looks for a needle in a haystack, and the other creates a new needle."
To find a needle in a haystack, algorithms are trained on a large database of molecules. Then they look for molecules with suitable properties. But create a new needle? This opportunity is provided by the generative adversarial networks that Zhavoronkov specializes in.
Such algorithms pit two neural networks against each other. One generates a meaningful result, and the other decides whether this result is true or false, Zhavoronkov says. Collectively, these networks generate new objects such as text, images, or, in this case, molecular structures.
“We started using this particular technology to make deep neural networks imagine new molecules to make it perfect from the start. We need perfect needles,”Zhavoronkov says. "You can turn to this generative adversarial network and ask it to create molecules that inhibit protein X at a concentration of Y, with the highest viability, desired characteristics, and minimal side effects."
Zhavoronkov believes that AI can find or make more needles from a multitude of molecular possibilities, freeing human chemists to focus on synthesizing only the most promising ones. If it works, he hopes, we can increase the number of hits, minimize misses, and generally speed up the process.
In the bag
Insilico is not alone in exploring new avenues for drug discovery, and this is not a new area of interest. Last year, a Harvard group published a paper on AI, which similarly selects candidates from drugs. The software trained on 250,000 drug molecules and used its expertise to create new molecules that mixed existing drugs and made suggestions based on desired properties. However, as noted by the MIT Technology Review, the results obtained are not always meaningful or easily synthesized in the laboratory, and the quality of these results, as always, is as high as the quality of the data provided initially.
Stanford Chemistry professor Vijay Pande says images, speech, and text - which are deep learning interests at this point - have good and clean data. But chemistry data, on the other hand, is still optimized for deep learning. In addition, while public databases exist, much of the data still lives behind the closed doors of private companies.
To overcome all obstacles, Zhavoronkov's company is focused on technology validation. But this year, skepticism in the pharmaceutical industry seems to be giving way to interest and investment. Even Google can break into the race.
As AI and hardware advances, the greatest potential still needs to be unlocked. Perhaps one day, all 1060 molecules in the drug domain will be at our disposal.
Ilya Khel