AI is coming for big pharma

If there’s one thing we can all agree on, it’s that the captains of industry of the 21st century are working to embed artificial intelligence in every corner of our world. But for all the ways that AI gets shoved in our faces and isn’t very successful, it may actually have at least one useful purpose. For example, by dramatically speeding up the often decades-long process of designing, discovering and testing new drugs.

Risk mitigation isn’t a sexy concept, but it’s worth understanding how common it is for a new drug project to fail. To set the scene, consider that each drug project takes between three and five years to generate a hypothesis strong enough to start testing in the lab. A 2022 study by Professor Duxin Sun With each project costing more than $2 billion, 90 percent of clinical drug development fails. And that number doesn’t even include compounds that don’t work in the preclinical phase. Simply put, every successful drug must add to at least $18 billion in waste generated by its failed siblings, which rarely guarantees that less profitable treatments don’t get the attention they deserve.

Dr. Nicola Richmond is Vice President of AI Good, a biotech company that uses artificial intelligence in the drug discovery process. He explained to researchers classic systems tasks, such as finding a misbehaving protein—the cause of a disease—and then finding a molecule that can make it work. Once they find it, they need to put that molecule into a form that the patient can take, and then test whether it’s both safe and effective. The journey to clinical trials on a live human patient takes years, and often alone after researchers learn what works theory does not work in experience.

The current process “requires more than a decade and billions of dollars in research investment for each approved drug,” he said. Recursion, another company in the AI ​​drug discovery space. A big skill for artificial intelligence, he said, could be avoiding misses and helping researchers avoid running down blind alleys for too long. A software platform that can go through hundreds of options at once can, in Gibson’s words, “fail faster and faster so you can move on to other targets.”

Image of Human HT29 Cells highlighted in Cell Profiler, the Carpenter-Singh software platform used to examine cell images.
CellProfiler / Carpenter-Singh lab at the Broad Institute

Dr. Founder Anne E. Carpenter Carpenter-Singh Laboratory at the Broad Institute of MIT and Harvard. For more than a decade, he developed techniques in Cell Drawing to highlight elements in cells with dyes and make them readable by computers. He is also a developer Cell Profiler, is a platform that allows researchers to use artificial intelligence to extract large images of these stained cells. Together, this work makes it easier for a machine to see how cells change when affected by the presence of a disease or treatment. Looking at every part of the cell as a whole – a “discipline called”omics” – there are greater opportunities to create the kind of connection that AI systems excel at.

Using images to identify potential treatments seems a little left field, because how things look doesn’t always mean how things really are, right? People have always made subconscious assumptions about medical conditions, Carpenter said. He explained that most people can tell if someone might have a chromosomal problem just by looking at their face. And professional clinicians can only visually identify a number of disorders as a result of their own experience. If you take a picture of everyone’s face in a given population, a computer can identify patterns and sort them by common characteristics, he added.

This logic applies to images of cells where a digital pathologist can compare images from healthy and diseased samples. If a human can do this, it should be faster and easier to use a computer to detect these differences in scale, as long as it is accurate. “You’re letting this data gather itself into groups, and now [you’re] we start to see patterns,” he explained, “when we treat [cells] With 100,000 different compounds, we can say one by one that there are two chemicals. really look alike.’” And it’s really similar, it’s not just a coincidence, it shows how they behave.

In one example, Carpenter showed that two different compounds can produce similar effects in a cell and, by extension, can be used to treat the same condition. If so, then one of the two, which is not intended for this purpose, may have less harmful side effects. Then there’s the potential benefit of being able to identify something we didn’t know was affected by the disease. “It allows us to say, ‘hey, there’s this group of six genes, five of which are very well known to be part of this pathway, but the sixth one, we didn’t know what it did, but now we have a gene. a strong hint that it’s involved in the same biological process.” “Maybe those other five genes, for whatever reason, aren’t great direct targets themselves, maybe the chemicals don’t bind,” he said, “but the sixth gene. [could be] it’s really great for that.”

A 30-year-old man of Indian origin researches a vaccine against COVID-19 in a scientific laboratory.
FatCamera via Getty Images

In this context, startups using artificial intelligence in drug discovery processes hope to find hidden diamonds in plain sight. Dr. Richmond said that Goodwill’s approach is to pick a disease of interest to the team and then build a biological question around it. So at the start of a project, a team might wonder if there are ways to treat ALS by enhancing or correcting the cell’s own housekeeping system. (To be clear, this is a purely hypothetical example provided by Dr. Richmond.)

This question is then fed through Benevolent’s AI models, which aggregate data from various sources. They then develop a ranked list of potential answers to the question, which may include new compounds or existing drugs that can be adapted. The data then goes to the researcher, who can examine what, if anything, contributed to his findings. Dr. Richmond added that even if the model choices are out of left field, it should provide evidence from existing literature or sources to support its findings. And it’s always the person who has the final say on what the consequences of it should be and how strongly it should be implemented.

It’s a similar situation in Recursion, Dr. Gibson claims that his model is now able to predict “how any drug will interact with any disease without physically testing it.” Based on the data it has assimilated and simulated, the model has already generated nearly three trillion predictions that connect potential problems with their potential solutions. Gibson said the process at the company now resembles an Internet search: Researchers sit at a terminal and “type in a gene associated with breast cancer and [the system] accommodates all other genes and combinations [it believes are] related.”

Dr Gibson said: “The interesting thing is when [we] See a gene on the list that no one has heard of, that feels like new biology because the world doesn’t know it exists.” Once the target is identified and the findings are human-verified, the data will be transferred to Recursion’s in-house science lab, where researchers will conduct preliminary tests to see if what is found in the simulation can be replicated in the real world. Recursion’s wet lab, which uses large-scale automation, is capable of running more than two million experiments in a single work week, Dr. Gibson said.

“After about six weeks, with very little human intervention, we will have results,” said Dr. Gibson, and if it’s successful, then the team will “really start investing.” Because, up to this point, the short approval process has cost the company “very little money and time to buy.” The promise is that rather than a three-year preclinical phase, this entire process can be broken down into a few database searches, some oversight, and then a few weeks of ex vivo testing to confirm whether the system’s predictions are worth the real effort. to question. Dr. Gibson says it’s a “year-old animal model job and [compressed] it’s up to two months in many cases.”

Of course, there is no concrete success story yet, no surprising treatment that any company in this field can show as validation of the approach. But Recursion can show a real-world example of how close its platform has come to the success of a critical study. Recursion was launched in April 2020 Sequence COVID-19 through the system to see potential treatments. He researched both FDA-approved drugs and candidates in late-stage clinical trials. The system generated a list of nine potential candidates that needed further analysis, eight of which would later be proven correct. He also said that Hydroxychloroquine and Ivermectin, which were heavily influenced in the early days of the pandemic, would fail.

There are currently AI-informed drugs undergoing real-world clinical trials. Does recursion pointing five projects are currently completing phase one (tests in healthy patients) or entering phase two (tests in people with the rare diseases in question). It started well First phase of BEN-8744treatment for ulcerative colitis may help other inflammatory bowel diseases. And BEN-8744 targets an inhibitor unrelated to current research, which, if successful, would lend weight to the idea that AIs might be able to spot connections that humans miss. Of course, we can’t draw any conclusions until the results of those preliminary tests are released, at least early next year.

DNA molecular structure with sequence data of human genome analysis on black background.
Yuichiro Chino via Getty Images

There are many unanswered questions, including how much we should trust AI as the sole arbiter of the drug discovery pipeline. There are also questions around the quality of the training data and biases in the wider sources. Dr. Richmond highlighted issues related to biases in genetic data sources, both in terms of homogeneity of cell cultures and how these tests are performed. Similarly, Dr. Carpenter made public the results of his latest project JUMP-Cell Painting project was based on cells from a single participant. “We chose it for good reason, but it’s still a human and a cell type from a human.” In an ideal world it would have a wider range of participants and cell types, but currently the issues are funding and time, or rather the lack thereof.

But for now, all we can do is wait for the results of these first tests and hope they bear fruit. As with every potential application of AI, its value lies largely in its ability to improve the quality of work, or more likely, the bottom line of the business in question. If AI can make savings attractive enough, then there may be a chance for ills in the current system to not recoup investment demands. It could all collapse into hype, or it could give real hope to families fighting for help while battling a rare disease.

This article was originally published on Engadget at

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *