Based on network theory, AI helps to explore complex puzzles of biology.

In the early days of modern genetics research, almost no one foresaw the enormous complexity of disease biology. When the human genome map was originally completed, people thought that they had obtained a manual on how the human body works. Based on the genomic map, we can find the gene that explains a particular disease and help find a cure for the disease.

To a certain extent, these studies have indeed brought us the gem of supreme. For example, geneticist Dr. Nancy Wexler studied the family history of patients with Huntington's disease in Venezuela and finally found that mutations in a single gene can predict whether a person will get the disease.

However, scientists soon discovered that the link between genes and disease is not always that simple, and that complex diseases like cancer and Alzheimer's disease are not caused by a mutation in a gene. Today, Dr. Cohen and other insightful people believe that there is an important link between the “simplification of research” and the decline in the efficiency of drug development. This decline in efficiency has resulted in a FDA-approved success rate of only 10% for a new treatment, and the cost of drug development has risen rapidly.

In recent years, scientists have begun to solve the problem of biological complexity with the help of network theory. The famous scientist of network theory, Dr. Albert-László Barabási of Northeastern University, believes that disease is like a bad signal spread from gene to protein through the network, and then spread to cells and tissues until all the disturbance to the network is finally expressed as the symptoms of the disease that we are usually familiar with.

Complex diseases are a combined result of countless effects, because gene pleiotropic means that any protein may play a role in different parts of the body. Startups like Pharnext assume that drugs can also be pleiotropic, interacting with multiple proteins and producing multiple effects in the body. In order to find a combination of drugs that can solve complex diseases, we must combine the important ability of machine learning to discover the law from massive data, and organically combine with the structural mechanism of disease occurrence.

And this requires the evolution of partnerships between computer scientists and biologists. A new generation of machine learning tools can absorb a lot of data and discover insights that transcend relevance. However, harnessing these “deep learning” neural networks to enable them to produce predictive capabilities still requires the construction of sophisticated algorithmic systems.

Colin Hill, founder and CEO of GNS Healthcare, is one of the engineers who built these algorithmic systems. His company in Cambridge, Massachusetts, has spent 18 years developing a computer system called REFS. GNS has raised $38 million venture capital from Celgene and Amgen's as well as other investors, to build and debug computer models of disease. In a series of recent studies published, GNS details the potential of the REFS system to mimic complex diseases such as Parkinson's disease.

Parkinson's disease is a very complex neurodegenerative disease, and its complexity and pleiotropic factors that cause disease make the efficacy of existing therapies very inconsistent. However, for Parkinson's disease, a series of network interactions caused by genetic defects have specific characteristics, and the destruction of exercise capacity is the most reliable indicator of disease progression. By introducing genetic information from patients with Parkinson's disease and healthy controls into the REFS system, it can help GNS generate more than 100 computer models to predict mechanisms that contribute to motor function deterioration. These models can help identify previously unknown genetic mutations that may accelerate the rate of disease progression.

This is just the first application of this model. Using these findings, GNS enables computers to simulate 5,000 different randomly controlled clinical trials, each of which is used to predict what disease progression will result from different treatments. This rapid detection is much faster than using real human clinical controlled trials to achieve the same results. GNS has partnered with other pharmaceutical companies to apply similar techniques to screen for potential therapies for diseases such as diabetes, ALS, multiple myeloma, and breast cancer. They now have the ability to create alternative models of human patients and diseases on computers, and these models can be used to test each drug and predict which treatments will be effective for those patients.

This kind of simulation is no longer just about discovering relevance. It began to answer the question of causality. What happens if we give drug A, not the drug B, to a specific patient? The ability to simulate and answer such hypothetical questions is a recent development in the AI field. According to GNS's technical consultant, Dr. Judea Pearl, a computer professor at UCLA and a senior AI researcher at UC, the real intelligence needs to go further from the level of discovery to the level of analyzing what will happen based on these rules to infer hypothetical situations. The data itself cannot provide any real insight if it is disconnected with any idea related to the mechanism.

From 2000 to 3 lead compounts, AI redefines "drug discovery"

Returning to the Pharnext example, Dr. Cohen is very optimistic about the prospects of Pharnext. At the same time, he also clearly recognized the limitations of AI technology. Google's artificial intelligence, AlphaZero, can defeat the world's top human players in Go games without any human chess. However, Dr. Cohen pointed out that the rules of Go are not complicated and AlphaZero is able to fully grasp these rules. However, in the field of biology, because of the multi-effects, we still don't understand and may never know all the rules, not to mention AI.

However, well-designed AI systems allow Pharnext to build models based on known rules and rely on them to make choices. Of the 10,000 known drugs, the drug development model selected 2,000 drugs that are already on the market and whose patents have expired. These drugs have been considered effective and safe by regulatory agencies.

To develop a therapy for CMT, Pharnext spent a year building a network model of the disease. Similar to the GPS Parkinson's disease model, this network model can show how genetic mutations can cause neurological and muscular disorders through various cascades. Based on this model, the computer calculates 57 candidate drugs that target different nodes in the cascade reaction. Pharnext then tested the drugs in an in vitro test, screened 22 drugs for animal testing, and finally found a combination of the three drugs into clinical trials. The results of a recent positive phase 3 clinical trial confirmed that the combination therapy of PXT3003 did play a role in multiple nodes of the cascade reaction.

Pharnext only took 3 years to develop preclinical development of PXT3003. Without the help of AI model, pre-clinical testing will take much longer. Dr. Cohen said that 2000 drugs can constitute one billion combinations, if in vitro testing is used. These combinations will bring countless false positive results and failures.

The progress of Pharnext and GNS shows that AI technology is growing and it is also driving the growth of pharmacology. An important demarcation point in the development of artificial intelligence is the ability to infer causality and use it to explore the answers to hypothetical questions. The computer models of these companies are moving in this direction.

Today, with the cost of new drug research and development hundreds of millions of dollars, AI-driven "new use of old drugs" may help pharmaceutical companies to extract more value from drugs that have been spent hundreds of billions of dollars. "You don't necessarily need to design a new drug," Cohen asserts. "My feeling is that I need only 50 different combinations of drugs to treat all the diseases." This gives totally new definition to "drug discovery".

Author's Bio: 

BOC Sciences