Artificial intelligence can make effective antibodies against diseases faster than humans
In what used to be an old biscuit factory in London, mixers and industrial ovens have been replaced by robotic arms, incubators and DNA sequencing machines. Instead of biscuits, today artificial intelligence makes medical antibodies against diseases. James Field and his company LabGenius have made this possible by preparing a revolutionary approach to designing new antibodies, precisely based on AI.
In nature, antibodies are the body’s response to diseases and serve as the first defense of our immune system in such cases. These are strands of proteins that are specially designed to neutralize body invaders such as bacteria and viruses. Since the late 1980s, pharmaceutical companies have been making synthetic antibodies to treat diseases like cancer and reduce the chance that the body will reject transplanted organs.
However, when people make such antibodies it takes a very long time: protein designers have to go through millions of potential combinations of amino acids to find the ones that will “fold” properly. After that, scientists have to test them experimentally, tweaking some of the variables to improve the healing properties, hoping that those changes don’t make the human body worse in other ways.
„If you want to create a new therapeutic antibody, somewhere in this infinite space of potential molecules, there is a molecule that you want to use.“, says the owner of LabGenius.
Field founded the company in 2012, when he was finishing his PhD in synthetic biology and saw the costs of DNA sequencing, robotics and computing falling. His company uses all three mentioned methods to automate the antibody discovery process.
A machine learning algorithm designs antibodies to target specific diseases, then automated robotic systems build and grow them in the lab, run tests and then feed that data back to the algorithm, all with limited human supervision.
Scientists start by identifying a search space for potential antibodies to fight a particular disease, and for that they need proteins that can distinguish between healthy and diseased cells. They then stick to the diseased cells and recruit the immune cell to finish the job. However, those proteins could be anywhere in the vast search space of potential options. Field’s company has developed a machine learning model that can explore that space much faster and more efficiently than humans.
„The only task you give the system as a human is: here is an example of a healthy cell and here is an example of a sick cell” says Field. Then, you let the system explore different antibody designs that can distinguish these cells.
How artificial intelligence makes antibodies versus conventional human design
The AI model selects more than 700 initial options from a search space of one hundred thousand potential antibodies, then automatically designs, builds and tests them. All this with the aim of finding potentially fruitful areas for more detailed research.
The tests are almost fully automated, and include a range of advanced equipment involved in sample preparation and running them through the various stages of the testing process. Antibodies are bred based on their genetic sequence and then go to bioassays – samples of the diseased tissue they are designed to fight. Humans oversee this automated process, but their job is mostly to move samples from one machine to another.
The experimental results from the first set of 700 molecules are returned to the artificial intelligence model and it uses them to refine the understanding of the space in which it investigates. In other words, the algorithm begins to build a picture of how different antibody designs change treatment effectiveness. With each successive round of antibody design, the AI model gets better, carefully balancing the exploitation of potentially good antibody designs with the exploration of new areas.
The challenge with conventional protein making is that as soon as scientists find something that works a little, they tend to make a large number of such molecules to see if they can further refine it. Those extra “tweaks” can improve a property, for example, how easily an antibody can be made on a large scale. However, they have a disastrous effect on many other attributes such as selectivity, toxicity and others. This means that the traditional approach can mean you’re “barking up the wrong tree,” meaning you’re endlessly trying to optimize something that only works a little bit, while at the same time there may be much better options on the other side, Field explains to ..
In addition, scientists in such an approach are limited by the number of tests you can run, or as Field compares, you are limited by “the number of shots on goal.” This means that human protein engineers tend to look for things they know will work, but as a result we get an accumulation of dogma in medicine.
That role is taken by the LabGenius approach in synthetic antibodies, which brings unexpected solutions that people may not have even thought about, and at the same time finds them faster. It takes only six weeks from the problem setting to the completion of the first batch of research, all of which is guided by machine learning models, claims the owner and CEO of this company.
Field’s firm has raised $28 million from sponsors and is starting to collaborate with pharmaceutical companies, offering them services such as consulting. Field says this is changing modern medicine and that the automated approach could be applied to other forms of drug discovery, making the long process of drug discovery much simpler.
He also claims that his company’s process actually provides better than antibody treatments that are more effective or have fewer side effects than those designed by humans. Field bases his argument for such a claim in the explanation that AI finds molecules that humans would never find using conventional methods.
The molecules found by artificial intelligence are very different and often counterintuitive compared to the antibody designs that humans would come up with using conventional methods. This, says this scientist, should allow us to obtain molecules with better properties, which ultimately translates into better results for patients.