AI's Breakthrough in Understanding Aging Cells - The Key to Eternal Youth?

Exploring huge piles of data on genes and cells, AI has found some amazing secrets. What exciting mysteries might they solve for us next? Now, to understand the sheer scale of what AI has achieved, we need to rewind a bit and look at the story behind it. So back in 1889, Francois Gilbert Viol, a French doctor, ventured down from the Andes, took some of his blood and looked at it under a microscope. He he found that the red blood cells crucial for carrying oxygen had increased by 42%. This was Voult's introduction to a fascinating aspect of human biology. Our bodies can produce these essential cells on demand when necessary. Early in the 20th century, scientists proposed that a hormone was behind this capability. They named it erythropoietin, or red maker in Greek. It took 70 years for scientists to isolate erythropoietin by filtering through 670 gallons of urine. Fast forward about 50 years and researchers in is identified a unique kidney cell responsible for producing this hormone when there's a shortage of oxygen. They named it the Norn cell, after the Norse gods thought to determine human destiny. It took humanity 134 years to uncover the Norn cells. Last summer, however, computers in California found them by themselves in just six weeks. This discovery happened when Stanford researchers set up computers to self learn biology. They used an AI program similar to ChatGPT, which learned language from billions of Internet techs. However, Stanford's team fed their AI raw data on millions of actual cells, including their chemical and genetic profiles. Without explaining the significance of these details or the differences among cells. The computers analyzed the data independently, organizing a model of cells based on their similarities in a complex multidimensional framework. The outcome was impressive. The AI could identify a previously unseen cell as one of over a thousand types, including the Norn cell. It's incredible because the AI model discovered the Norn cell in the kidney without being informed of its existence, mentioned Jure Leskovec, a computer scientist at Stanford behind the project. This software is among various new AI enabled tools dubbed foundation models, aiming to understand biology's basics. These models aren't just organizing biological data, they're uncovering new insights into gene functions and cell development. As these models grow, incorporating more lab data and computing power, experts believe they'll lead to even more significant findings to potentially unveiling mysteries about cancer and other diseases or discovering methods to transform cell types. Discovering something about biology that biologists haven't been able to would be a landmark moment, and I believe it's coming, said Dr. Eric Topol, head of the Scripps Research Translational Institute. The extent of these models capabilities is up for debate. While some are skeptical, others are hopeful that foundation models will eventually address one of biology's greatest puzzles, the distinction between life and non life. Now, for years, biologists have been intrigued by how our body's different cells utilize genes to perform the myriad tasks necessary for survival. Roughly 10 years ago, large scale experiments began aimed at identifying genetic information from individual cells. The findings were cataloged in extensive databases or cell atlases that grew to contain billions of data points. Dr. Christina Theodorus, while a medical resident at Boston Children's Hospital, learned about a new AI model developed by Google in 2017 for translating languages. This model was trained on millions of English sentences and their translations in German and French, gaining the ability to translate previously unseen sentences. Dr. Theodorus speculated on the possibility of a similar model being able to decipher the information contained within cell atlases. In 2021, she faced challenges in finding a lab willing to explore this idea. Due to widespread skepticism about its feasibility, Shirley Liu, a computational biologist at the Dana Farber Cancer Institute in Boston, decided to give her a chance. Dr. Theodorus gathered data from 106 human studies encompassing 30 million cells and fed this into her newly developed program Geneformer. The model acquired a profound comprehension of gene behavior across various cells. For instance, it predicted that deactivating a gene named TED4 in a specific heart cell type would cause significant disruption. When this prediction was tested on real heart cells, known as cardiomyocytes, their ability to beat diminished. In another experiment, Dr. Theodorus and her team introduced gene former to heart cells from both individuals with abnormal heart rhythms and healthy individuals. They tasked the model with identifying changes needed to restore health to the disease cells. Geneformer suggested diminishing the activity of four genes previously unassociated with heart disease. Following the model's guidance, Dr. Theodorus team attempted to suppress these genes. In two of the four attempts, the treatments enhanced the cell's functionality. The Stanford team later joined the field of foundation models after contributing to the creation of CELX gene, one of the largest cell databases globally. They trained their AI on 33 million cells from this database, focusing on messenger RNA and protein structures, both products of genes. The resulting model, dubbed Universal cell Embedding, learned to categorize over a thousand cell types by observing gene activation patterns. It arranged 36 million cells into clusters based on gene usage, echoing discoveries made by generations of biologists. UCE also deduced significant insights into cellular development from a single fertilized egg, essentially redefining developmental biology, it understood that body cells could be classified by their origin from one of the early embryo's three layers. Furthermore, UCE proved capable of applying its knowledge to unfamiliar species. When given genetic data from an unknown animal like a naked mole rat, it could accurately identify many of its cell types. This model can handle any new organism chicken, frog, fish, you name it and still produce meaningful output, Dr. Leskovec explained. Upon identifying the norn cells, the team speculated, based on their database, that these cells might exist beyond the kidneys, possibly throughout the body. Dr. Catalyn Sustak, a researcher studying norn cells, expressed both interest and skepticism about this finding. While doubting the presence of erythropoietin producing norn cells outside the kidneys, she acknowledged the potential for these newly identified cells to sense oxygen similarly. In essence, UCE might have uncovered a new cell type before traditional biological methods could alright now. Just as ChatGPT can make errors, so can biological AI models. Kazia Kadzierska, a computational biologist at Oxford University, along with her team put geneformer and another foundational AI model, szgpt, through rigorous testing. They challenged these models with unseen cell atlases, asking them to categorize cells into types. While the AI models excelled at certain tasks, they sometimes underperformed compared to simpler software. Dr. Kzyrska remains optimistic about the potential of these models, but cautions against using them blindly due to their current limitations. Dr. Leskovecz pointed out that as more data becomes available for training, these models are getting better. However, he mentioned that the volume of data in cell atlases is relatively small compared to the vastness of information ChatGPT was trained on, expressing a wish for an Internet of cells the future looks promising with larger cell atlases being developed in a wider variety of cell data data being collected. This includes detailing the molecules attached to genes and capturing detailed images of cells to identify the precise locations of proteins. This wealth of information will enhance the foundation model's ability to understand cell functionality deeply. Efforts are also being made to merge the self learned insights of these models with the vast knowledge already documented by biologists. This includes integrating discoveries from thousands of scientific papers with cell measurement databases, aiming for a comprehensive mathematical model of a cell. Bo Wang, the creator of SDGPT and a computational biologist at the University of Toronto, believe such a virtual cell could revolutionize biology. It would enable scientists to simulate experiments digitally predicting cell behavior in any given scenario without needing a physical laboratory. Dr. Quake is intrigued by the possibility that foundation models might uncover not only the second cell types known to exist, but also those that could potentially exist adhering to the biochemical rules of life. He imagines creating a map of life's boundaries, exploring beyond which life cannot sustain. Such a map might even lead to the creation of novel cells not currently found in nature, with foundation models devising chemical blueprints to convert regular cells into ones with unique capabilities. These futuristic cells could perform tasks like clearing plaque from blood vessels or examining diseased organs from the inside. Dr. Quake acknowledges the speculative nature of these ideas, likening them to the science fiction narrative of Fantastic Voyage. Yet he remains open to the boundless possibilities the future may hold. Now let's talk about some of the potential new risks. If the foundation models achieve what Dr. Quake envisions, they could introduce various new challenges. Recently, over 80 biologists and AI specialists have called for regulations on this technology to prevent its potential misuse, such as the development of novel biological weapons derived from artificially created cells. Concerns about privacy could emerge even more swiftly. There's hope to develop foundation models tailored to individual genomes, providing insights into how specific genetic variations influence cellular functions. This level of personalization could lead to groundbreaking medical discoveries, but might also expose sensitive genetic information of those who contribute their DNA and cell data for research. Amid these advancements and concerns, some experts question the ultimate capabilities of foundation models. Their effectiveness hinges on the quality and scope of the data they're trained on. Uncovering significant new truths about life may require data we don't yet know how to gather or may not even realize is necessary. Sarah Walker, a physicist at Arizona State University who explores the origins of life, acknowledges that while these models could lead to interesting findings, their potential for fundamental breakthroughs is limited by the current scope of data. Nevertheless, the success of foundation models has prompted a reevaluation of the role of human biologists. Traditional biology has valued the creativity and labor involved in explaining experimental research that uncovers life's mysteries. However, computers, by analyzing vast quantities of cell data, might reveal complex patterns and insights in a fraction of the time, challenging our notions of scientific creativity and the future role of biologists in research. Dr. Quake suggests this shift could necessitate a significant rethinking of creativity's nature in the scientific process, hinting at a future where the academic stage status quo, particularly for professors accustomed to traditional research methodologies, may be upended. All right, don't forget to hit that subscribe button for more updates. Thanks for tuning in, and we'll catch you in the next one.