4 minute read
Eugene Wang ENGR-E 434 Oct. 20, 2020
AI in Drug Discovery Process from company Atomwise
A long time ago, most drugs are discovered by either identifying the active ingredients used in traditional remedies, or accidental discovery by rando chance. In the modern era, the commonly used approach to discovery new drugs has been to use the knowledge and understanding of how diseases operates at the molecular level in the body and try to target the specific areas in the disease or body. The entire drug development process can be broadly split into non-human testing, human testing (clinical trials), then FDA review. The non-human part can also be characterized as the in-vitro and in-vivo testing. And of course, in-vitro/in-vivo testing comes before clinical trials. This part consists of discovery and development of a drug, and preclinical research about the safeness of the drug.[^1] Here we will see how AI can increase the speed and efficiency of the 1) discovery of potential drug candidates and 2) the in-vitro testing of those potential drug candidates.
AtomNet is a AI technology built by Atomwise. This technology is based on a type of deep neural network called deep convolutional neural network. This type of neural network is well suited to learn data that contains hierarchical structures, which is why it has been enormously successful in image recognition. Turns out, the strengths of convolutional models in localizing features and hierarchical composition can also be applied to the modeling of bioactivity and chemical interactions. AtomNet is trained on 3 dimensional images that digitally represents atoms like carbon, hydrogen, and oxygen, and the bonds formed between them. AtomNet, on its own, learns about the laws and rules governing molecular binding and the degree of affinity between molecules. AtomNet learns about things like how spatial arrangement, angle, and proximity affect the strength of repulsion or attraction between molecules. AtomNet has the ability to screen billions of molecules and pick out the subset of promising molecules with the desired effects.[^2]
Finding the molecules that can deliver the desired properties is incredibly time consuming and tedious. Because the amount of these tests is easily in the ranges of millions of tests. These tests include genetic, pharmacological, and chemical tests. These tests help identify active ingredients that affect molecules on the human body. Long time ago, scientists and pharmacists have to do them be hand, which takes a mindboggling amount of time, money, and effort. But nowadays, these tests are often automated with laboratory robots who do all the work with liquid handling, data gathering, and using sensors, etc. This step can often take years even with robotic automation.[^1] But with AI technology, AtomNet, this process can be massively accelerated and shortened into a matter of weeks. If any molecules with desired effects are found, they are further tested than optimized to increase potency and reduce side effects.[^2]
There are a total of around twenty thousand proteins in the human genome and only around 750 have FDA-approved drugs. And about four thousand proteins have evidence that they are linked to some kind of disease. So the remaining sixteen thousand (80%) of gene targets are quite unknown to us and not really studied. AtomNet can help scientists advance into these unchartered zones. AtomNet is able to screen fourteen thousand (out of total of twenty thousand) gene target even without complete structural data. AtomNet can also screen over sixteen billion synthesizable molecules for their reactions against biological compounds in the short time of only two days.[^3] This task in the past has been painstakingly labor and money intensive. This task is like finding the needle in a haystack; with the haystack representing the entire complex chemical space. With AtomNet, this task is now automated by machine learning and AI.
Drug discovery and development has always been expensive and time consuming. With heavy regulation to ensure safety and the amount of investment capital needed, drug discovery and development related industry hasn’t seen the explosive growth in industries like information technology. The good thing is, the growth and emergence of AI and machine learning technology in the IT industry has found great applications in the pharmaceutical industry. Here we review one of the problem, drug discovery and testing, and compared its traditional solution and the new, AI-accelerated solution that promises faster and cheaper drug discovery for the future.