At the intersection of language and life, researchers at Mount Sinai have taught an artificial intelligence to read the grammar of the genome — learning how genes shift meaning depending on their cellular context, much as words shift meaning depending on their sentence. The system, trained on millions of gene groupings drawn from decades of published science, can now anticipate biological relationships before a single experiment is run. In doing so, it offers medicine something rare: a map of what we do not yet fully know, drawn from the patterns of what we do.
Mount Sinai researchers unveil AI model that decodes how genes interact in human cells
A gene can play different roles in different settings, like a word in different sentences.
Mark
So the model learns from millions of gene sets. How does it actually know what it's learned? What's the evidence it understands anything real?
Mimi
They tested it by training on older publications, then asking it to predict discoveries in newer papers. It successfully identified gene relationships that scientists later confirmed experimentally. That's a real test of whether the patterns it found were meaningful.
Luke
But that's one validation approach. We should ask: how many predictions did it make? How many were correct? The reporting doesn't give us the accuracy rate or false positive rate. We know it showed "strong performance," but strong compared to what, exactly?
Mark
Fair point. So what makes this different from other AI models that look at genes?
Mimi
Most other biological AI models rely on gene expression data—basically, measuring which genes are turned on or off. This one is trained on gene sets, which are curated groupings of genes that work together. It's a different kind of information entirely.
Luke
And that's genuinely novel. But I want to know: how diverse is the training data really? They say it integrates data from many diseases and methods, but which diseases? Which methods? Are there blind spots in what the model learned?
Mark
What can scientists actually do with this right now?
Mimi
The immediate use is improving gene set enrichment analysis, which is already standard in molecular biology labs. It could help researchers interpret their data faster and more accurately. They could also use it to make educated guesses about what unknown genes do without running expensive experiments first.
Luke
That's useful, but it's still a tool for interpretation. It's not replacing experiments; it's suggesting which experiments to run. That's an important distinction for readers to hold.
Mark
What's the real long-term play here?
Mimi
They want to combine this with language models so it can explain gene functions in plain English, and eventually link it with drug-focused AI to predict how drugs interact with cells. That could accelerate drug discovery.
Luke
Those are future directions, though. The paper is about the foundation model itself. We should be clear about what exists now versus what's planned.
Il Polso
- Genes rarely act alone, yet science has lacked a unified system to track how they form and reform their molecular alliances across thousands of biological conditions — until now.
- Mount Sinai's GSFM was trained like a language model, tasked with predicting missing genes from partial sets until it internalized the deep logic of how genomes organize themselves.
- The model demonstrated it could anticipate gene-function relationships that had not yet been experimentally confirmed, validated by correctly predicting discoveries published after its training cutoff.
- Unlike tools that rely solely on gene expression data, GSFM draws on gene sets — an underused information type — allowing it to integrate findings across diseases, methods, and research conditions into a single coherent map.
- Researchers now aim to fuse GSFM with language models and drug-focused AI, pointing toward a future where the system can explain gene behavior in plain language and forecast how drugs will interact with living cells.
At the intersection of language and life, researchers at Mount Sinai have taught an artificial intelligence to read the grammar of the genome — learning how genes shift meaning depending on their cellular context, much as words shift meaning depending on their sentence. The system, trained on millions of gene groupings drawn from decades of published science, can now anticipate biological relationships before a single experiment is run. In doing so, it offers medicine something rare: a map of what we do not yet fully know, drawn from the patterns of what we do.
Researchers at the Icahn School of Medicine at Mount Sinai have built an AI system that learns how genes work together inside human cells — a development that could reshape how scientists understand disease and develop new treatments.
The model, called a gene set foundation model or GSFM, was trained on millions of gene groupings drawn from published studies and gene expression datasets representing hundreds of thousands of independent research efforts. Its design takes inspiration from large language models: just as a word shifts meaning depending on its sentence, a single gene can behave differently depending on where and when it activates. The team asked whether AI could learn the "meaning" of genes the same way — by observing patterns across vast data.
To train the system, researchers used a puzzle-solving method, presenting partial gene sets and asking the model to predict what was missing. Over time, it internalized the underlying logic of how genes group and interact. Crucially, the team validated its predictive power by training on publications up to a set date, then testing whether it could anticipate findings reported in studies published afterward — which it could.
What distinguishes GSFM is its reliance on gene sets rather than gene expression data alone, an underused category of biological information. This lets the model integrate findings across many diseases, methods, and conditions into a unified map of gene relationships. Practical applications include identifying the function of poorly understood genes, flagging disease-relevant genes, and suggesting drug targets — all without immediate laboratory work.
Looking ahead, the Mount Sinai team plans to combine GSFM with language-based AI to generate plain-language explanations of gene function, and with drug-focused models to predict how therapeutics interact with cells — extending the system's reach from biological understanding toward the design of new medicines.
Researchers at the Icahn School of Medicine at Mount Sinai have built an artificial intelligence system that learns how genes work together inside human cells—a breakthrough that could reshape how scientists understand disease and develop new treatments.
The model, called a gene set foundation model, or GSFM, was trained on millions of gene groupings pulled from published scientific studies and gene expression datasets, representing hundreds of thousands of independent research efforts. The work, published in the journal Patterns, takes inspiration from large language models like ChatGPT, which learn how words shift meaning depending on context. The Mount Sinai team applied the same logic to genes: a single gene can behave differently depending on where and when it activates in a cell, much as a word can mean different things in different sentences.
Avi Ma'ayan, a professor of pharmacological sciences and director of the Mount Sinai Center for Bioinformatics, explained the core insight: genes rarely work in isolation. Instead, they participate in multiple biological processes, forming different molecular groupings depending on their cellular environment. The question the researchers asked was whether an AI system could learn the "meaning" of genes the way modern language models learn the meaning of words—by observing patterns across vast amounts of data.
To train the model, the team used a puzzle-solving approach: they gave the system part of a gene set and asked it to predict the missing pieces. Over time, the model learned underlying patterns describing how genes are grouped and interact. When tested against other approaches, the GSFM showed strong performance, including the ability to identify gene-gene and gene-function relationships before they were confirmed experimentally. The researchers validated this by training the model on gene sets from publications up to a specific date, then testing whether it could predict discoveries reported in studies published after that cutoff.
What sets this approach apart is that it relies on gene sets rather than gene expression data alone—a largely underused type of biological information. This allows the model to integrate diverse data from many diseases, experimental methods, and research conditions, creating a unified map of gene relationships across biology. The system can help identify the function of poorly understood genes without immediate laboratory work, highlight genes involved in disease, suggest potential drug targets and biomarkers, and provide a reusable knowledge system for many types of biomedical research tasks.
One immediate application is in gene set enrichment analysis, a widely used method in molecular biology research. By improving how scientists interpret gene groupings, the model may help uncover new biological insights from both existing and future datasets. The Mount Sinai team plans to expand the system further by combining GSFM with other AI foundation models. One goal is to integrate it with language-based models to generate natural-language explanations of gene functions. Another direction involves combining GSFM with drug-focused AI models, with the long-term aim of predicting how drugs interact with cells and supporting the design of new therapeutics.
Citazioni salienti
Genes rarely act alone. Instead, they participate in multiple biological processes, forming different molecular groupings depending on where and when they are active in the cell.— Avi Ma'ayan, Professor of Pharmacological Sciences, Mount Sinai
Unlike previous biological AI models that primarily rely on gene expression data, our GSFM is uniquely trained on gene sets, a different and largely underused type of biological information.— Avi Ma'ayan
Vuoi la storia completa?
Leggi l'originale su Phys.org ↗
Technology