Revolutionary AI Unveils the Secrets of Cellular Mechanisms

By employing a new artificial intelligence technique, researchers at Columbia University Vagelos College of Physicians and Surgeons can accurately anticipate gene activity in any human cell, essentially unveiling the internal workings of the cell. The system, outlined in the current edition of Nature, has the potential to revolutionize how scientists strive to comprehend everything from cancer to genetic disorders.

“Predictive generalizable computational models enable the rapid and precise discovery of biological processes. These techniques can efficiently carry out large-scale computational experiments, enhancing and steering traditional experimental methods,” states Raul Rabadan, professor of systems biology and senior author of the recent publication.

Conventional research techniques in biology excel at demonstrating how cells perform their roles or respond to disruptions. However, they lack the capability to make predictions about cellular operations or how cells will adjust to alterations, such as a mutation that induces cancer.

“Possessing the ability to forecast a cell’s actions would revolutionize our grasp of essential biological processes,” Rabadan remarks. “It would shift biology from a discipline that describes seemingly haphazard processes to one that can forecast the fundamental systems that influence cellular behavior.”

In recent years, the accumulation of extensive data from cells, along with more advanced AI models, is beginning to shift biology towards a more predictive discipline. The 2024 Nobel Prize in Chemistry recognized researchers for their pioneering work in utilizing AI to predict protein structures. Yet, applying AI methodologies to forecast the behavior of genes and proteins within cells has proven to be more challenging.

New AI method predicts gene expression in any cell

In this new research, Rabadan and his associates attempted to utilize AI to determine which genes are active within specific cells. Such insights into gene expression can inform researchers about the cell’s identity and its functional processes.

“Previous models were trained on data from particular cell types, typically cancer cell lines or something that bears little resemblance to healthy cells,” Rabadan indicates. Xi Fu, a graduate student in Rabadan’s laboratory, opted for a distinct approach, training a machine learning model on gene expression data gathered from millions of cells derived from normal human tissues. The inputs included genome sequences and information indicating which parts of the genome are accessible and expressed.

This overall method resembles how ChatGPT and other renowned “foundation” models operate. These systems use a body of training data to identify underlying principles, the grammar of language, and then apply those inferred principles to new contexts. “Here, it’s precisely the same: we learn the grammar in various cellular states, and then we enter a specific condition — whether it’s a diseased cell or a standard cell type — and we attempt to see how well we can predict patterns from this data,” Rabadan explains.

Fu and Rabadan soon brought together a team of collaborators, including co-first authors Alejandro Buendia, currently a Stanford PhD student who previously worked in Rabadan’s lab, and Shentong Mo from Carnegie Mellon, to train and validate the new model.

After training with data from over 1.3 million human cells, the system achieved sufficient accuracy to predict gene expression in cell types it had never encountered, yielding results that closely matched experimental data.

New AI methods reveal drivers of a pediatric cancer

Next, the researchers demonstrated the potential of their AI system when they tasked it with uncovering previously hidden biology in diseased cells, particularly in a hereditary form of pediatric leukemia.

“These children inherit a mutated gene, and it was not entirely clear what these mutations were affecting,” Rabadan explains, who also co-directs the cancer genomics and epigenomics research program at Columbia’s Herbert Irving Comprehensive Cancer Center.

Through AI, the team predicted that the mutations interfere with the interaction between two distinct transcription factors that govern the fate of leukemic cells. Laboratory experiments validated AI’s prediction. Understanding these mutations’ impact reveals specific mechanisms that propel this disease.

AI could reveal “dark matter” in genome

The new computational techniques should enable researchers to commence investigating the role of the genome’s “dark matter” — a term adapted from cosmology, referring to the vast majority of the genome that does not code for known genes — in cancer and other ailments.

“The vast majority of mutations discovered in cancer patients occur in so-called dark regions of the genome. These mutations do not influence protein functions and have largely remained unexamined,” Rabadan states. “The aim is that by employing these models, we can explore mutations and illuminate that segment of the genome.”

Rabadan is already collaborating with researchers at Columbia and other institutions, probing various cancers from brain to blood cancers, deciphering the grammar of regulation in healthy cells, and observing how cells alter during cancer progression.

The research also opens new possibilities for comprehending numerous diseases beyond cancer and potentially identifying targets for novel treatments. By presenting new mutations to the computational model, researchers can now derive profound insights and predictions about how those mutations influence cell behavior.

Following closely on the heels of other recent advancements in artificial intelligence for biology, Rabadan perceives this work as part of a substantial trend: “It marks the dawn of a new era in biology that is extraordinarily exhilarating; transforming biology into a predictive discipline.”