Scientists Develop a Machine Learning Method to Help Design Better Antibody Drugs

2024-03-08

According to foreign media reports, antibodies are not only produced by our immune cells to fight against viruses and other pathogens in the body. For decades, the medical community has also been using antibodies produced through biotechnology as drugs. Thisbecause antibodies are very good at specifically binding to molecular structures based on the "lock key principle". Their uses include learning from tumors to treat autoimmune diseases and neurodegenerative diseases.

However, developing this antibody drug is not easy. The basic requirement is that the antibody binds to its target molecule in an optimal manner. Meanwhile, an antibody drug must meet a series of additional criteria. For example, it should not trigger an immune response in the body, it should be produced effectively using biotechnology, and it should remain stable for a long period of time.

Once scientists find antibodies that bind to the desired molecular target structure, the development process is far from over. On the contrary, this marks the beginning of a phase in which researchers use biotechnology to attempt to improve the characteristics of antibodies. Scientists led by Professor Sai Reddy from the Department of Biosystems Science and Engineering at ETH Zurich have now developed a machine learning approach to support this optimization phase and help develop more effective antibody drugs.

Robots cannot manage over a few thousand

When researchers optimize the entire antibody molecule in therapeutic form (i.e. not just a fragment of the antibody), the past started with a candidate antibody that reasonably binds to the desired target structure. Then the researchers randomly mutated genes carrying antibody blueprints to generate thousands of related candidate antibodies in the laboratory. The next step is to search between them to find antibodies that bind well to the target structure. "Through automated processes, you can test thousands of candidate therapeutic drugs in the laboratory. However, it is not feasible to screen out more than that," Reddy said. Usually, a dozen or so better antibodies in this screening process will proceed to the next step and their degree of meeting additional criteria will be tested. He said, "In the end, this method can help you find good antibodies from a group of thousands of people.".

Massively increasing candidate pool through machine learning

Reddy and his colleagues are currently using machine learning to increase the initial antibody set to be tested to several million. Reddy said, "The more candidate drugs there are to choose from, the greater the chance of finding all the criteria that truly meet the requirements for drug development."

ETH researchers have provided a proof of concept for their new method using Roche's antibody cancer drug Hesetin, which has been on the market for 20 years. "But we don't want to make suggestions on how to improve it - you can't just retrospectively change an approved drug," Reddy explained. "We chose this antibody because it is well-known in the scientific community and its structure has been published in open databases."

Computer prediction

Starting from the DNA sequence of Herceptin antibodies, ETH researchers created approximately 40000 related antibodies using the CRISPR mutation method they developed a few years ago. Experiments have shown that 10000 antibodies can effectively bind to a specific cell surface protein, which is the target protein. Scientists use the DNA sequences of these 40000 antibodies to train a machine learning algorithm.

Then they used the trained algorithm to search a database containing 70 million potential antibody DNA sequences. For these 70 million candidate antibodies, the algorithm predicted the degree of binding between the corresponding antibodies and the target protein, resulting in a list of millions of sequences expected to bind.

Using further computer models, scientists predicted how these millions of sequences would meet additional criteria for drug development (tolerance, production, physical properties). This reduces the number of candidate sequences to 8000.

Scientists selected 55 sequences from their computer's optimized candidate sequence list, produced antibodies in the laboratory, and described their characteristics. Subsequent experiments showed that several of these sequences bind to the target protein even better than Herceptin itself, and are easier to produce and more stable than Herceptin itself. Reddy said, "A new variant may even have better tolerance in vivo than Herceptin. It is well known that Herceptin triggers a weak immune response, but in this case, it is usually not a problem." However, for many other antibodies, thisa problem and must be prevented for drug development. "

ETH scientists are now applying their artificial intelligence methods to optimize antibody drugs under clinical development. For this reason, they recently established the ETH derivative company DeepCDR Biology, which collaborates with early and mature biotechnology and pharmaceutical companies for antibody drug development.

Driven by the trend of industrial digitization, intelligent building technology is flourishing