In a similar manner to how ChatGPT comprehends human language, an innovative AI model created by computational biologists is decoding the intricate communication of cells, enabling precise predictions of their actions.
In the same way that ChatGPT understands human language, a new AI model developed by Columbia computational biologists captures the language of cells to accurately predict their activities.
Employing a groundbreaking artificial intelligence technique, researchers from Columbia University Vagelos College of Physicians and Surgeons are capable of accurately forecasting gene activity in any human cell, essentially unveiling the cell’s internal workings. This system, detailed in the latest edition of Nature, has the potential to revolutionize how scientists investigate a wide range of issues, including cancer and genetic disorders.
“Predictive and generalizable computational models enable rapid and precise revelations of biological processes. These methodologies efficiently conduct extensive computational experiments, thus enhancing and directing traditional experimental techniques,” states Raul Rabadan, a professor of systems biology and the senior author of the study.
Conventional biological research methods excel at demonstrating how cells execute their functions or respond to alterations. However, they fall short in forecasting cellular behavior or how cells might respond to changes such as a mutation that leads to cancer.
“Being able to accurately predict a cell’s functions would reshape our comprehension of essential biological processes,” Rabadan notes. “It would shift biology from merely describing what appears to be random occurrences to one that can anticipate the fundamental systems that influence cell behavior.”
Recent years have seen the gathering of vast cellular data alongside the emergence of more advanced AI models, beginning to shift biology toward a more predictive science. For instance, the 2024 Nobel Prize in Chemistry honored scientists for their pioneering efforts in utilizing AI to anticipate protein structures. Nevertheless, applying AI techniques to forecast gene and protein activities within cells has proven to be much more challenging.
Revolutionary AI method for predicting gene expression in various cells
In their new research, Rabadan and his team aimed to leverage AI to determine which genes are activated within specific cells. Insights into gene expression can inform researchers about the cell’s identity and its functional capabilities.
“Earlier models were predominantly trained on data from specific cell types, often cancer cell lines or other cells that bear little resemblance to normal cells,” explains Rabadan. Graduate student Xi Fu, part of Rabadan’s team, opted for a different strategy: training a machine learning model using gene expression data from millions of normal human tissues. The data set included genome sequences along with information on which parts of the genome are accessible and expressed.
This approach is comparable to how ChatGPT and other foundational models operate. These systems utilize a training dataset to uncover fundamental rules, much like the grammar of a language, and then apply those learned rules to new scenarios. “In this case, we are doing something similar: we learn the grammar from various cellular states, and then analyze a specific condition — whether it is a diseased or normal cell type — to see how accurately we can predict patterns based on this information,” Rabadan elaborates.
Fu and Rabadan brought together a team of collaborators, including co-first authors Alejandro Buendia, now a PhD student at Stanford, and Shentong Mo from Carnegie Mellon, to train and test the novel AI model.
After processing data from over 1.3 million human cells, the system achieved sufficient accuracy to predict gene expression in cell types it had never encountered before, producing results that closely matched experimental data.
Innovative AI methods uncover the causes of a childhood cancer
The power of their AI system was next demonstrated when researchers tasked it with revealing previously hidden biological insights regarding diseased cells, particularly an inherited type of pediatric leukemia.
“Children with this condition inherit a mutated gene, and the exact implications of these mutations were unclear,” explains Rabadan, who also co-leads the cancer genomics and epigenomics research initiative at Columbia’s Herbert Irving Comprehensive Cancer Center.
By utilizing AI, the researchers predicted that these mutations interfere with the interaction between two transcription factors that influence the fate of leukemic cells. Subsequent laboratory experiments validated the AI’s prediction. Understanding how these mutations operate sheds light on specific mechanisms responsible for this disease.
AI’s potential to explore the “dark matter” of the genome
The new computational strategies are expected to empower researchers to investigate the role of the genome’s “dark matter” — a term borrowed from cosmology referring to the majority of the genome that does not encode known genes — in relation to cancer and other illnesses.
“Most mutations found in cancer patients occur in what are termed ‘dark regions’ of the genome. These mutations do not impact protein function and have largely gone unexplored,” says Rabadan. “The idea is to utilize these models to analyze mutations and shed light on that portion of the genome.”
Rabadan is already collaborating with researchers at Columbia and other institutions to study various types of cancers, spanning from brain to blood cancers, and to learn about the regulatory grammar in normal cells and how they change during cancer progression.
This research also opens doors to understanding numerous diseases beyond cancer and may help identify targets for new treatment options. By presenting novel mutations to the AI model, researchers can gain profound insights and predictions regarding the specific effects of those mutations on a cell.
Following other recent breakthroughs in AI’s application to biology, Rabadan views this research as part of a significant transformation: “We are entering a new era of biology that is incredibly exciting; it is transforming biology into a predictive science.”
