Knowledge Graph-Based Hypothesis Generation and Experimental Validation in Yeast

Science often feels like trying to solve a massive jigsaw puzzle where the picture on the box is missing, and the pieces are scattered across thousands of different tables. In biology, these pieces are facts about how living things work: which genes control which proteins, how cells react to chemicals, and what happens when a specific part of a cell breaks down. For decades, scientists have gathered these facts into vast digital libraries, but the sheer volume of information makes it nearly impossible for a human to look at every piece and guess which ones fit together next. This is where the idea of a "knowledge graph" comes in. Imagine a giant, interconnected web where every fact is a dot, and every relationship between facts is a line connecting them. By using computer programs to study the shape of this web, researchers can spot patterns and predict which missing connections are most likely to be real, effectively suggesting new experiments before a single test tube is touched.

This approach is particularly vital for understanding how organisms survive stress. Yeast, a tiny single-celled fungus, is one of the most studied life forms on Earth, serving as a simple model for how more complex cells, including our own, function. Yeast cells face constant threats from their environment, such as toxic chemicals or extreme temperatures. When a yeast cell encounters a stressor like vinegar or alcohol, it relies on specific genes to help it survive. If a gene that is crucial for this defense is missing or broken, the cell will struggle to grow. The challenge for scientists is figuring out exactly which genes protect the cell against which specific chemicals. With thousands of genes and hundreds of chemicals, the number of possible combinations is astronomical. Testing them all by hand would take a lifetime.

To tackle this problem, a team of researchers at Chalmers University of Technology and Gothenburg University in Sweden built a system that combines a digital map of yeast biology with automated laboratory robots. Their goal was to predict which genes help yeast resist six specific chemical stressors: hydrogen peroxide, tunicamycin, caffeine, sodium chloride, acetic acid, and ethanol. They started by feeding a computer program all the known facts about yeast genes, proteins, and chemical reactions into a knowledge graph. This graph acts like a massive database where every piece of information is linked to others. The computer then used a sophisticated type of artificial intelligence, known as a graph neural network, to look at this web of connections. The network didn't just memorize the facts; it learned to understand the structure of the relationships, using the known connections to guess which unknown connections were most likely to exist. Specifically, it looked for genes that, if removed, would make the yeast more vulnerable to the six chemicals.

The computer generated a list of the most promising guesses, ranking them by how likely they were to be true. The researchers selected the top thirty-five predictions to test in the real world. This included the five best guesses for each of the six chemicals, plus five additional guesses involving genes that were poorly understood, essentially unknown variables in the yeast genome. To test these ideas, they used an automated laboratory platform called Eve. This robot system handled the entire process with precision, growing thousands of yeast cultures in tiny wells. The robot mixed the yeast with the specific chemicals, monitored their growth over twenty-four hours, and recorded the data. By comparing how fast the yeast grew with and without the chemicals, and comparing the mutant yeast (missing a specific gene) to the normal yeast, the team could determine if their computer predictions were correct.

The results were a clear success for this method. Out of the thirty-five hypotheses the computer generated, ten were confirmed by the experiments. This means that for ten specific gene-chemical pairs, the computer correctly predicted that removing the gene would make the yeast significantly weaker when exposed to that chemical. The success rate varied depending on the chemical. For hydrogen peroxide, a reactive chemical that damages cells, four out of the five top predictions were correct. For caffeine and acetic acid, two out of five were confirmed. For sodium chloride and tunicamycin, one out of five was confirmed. Surprisingly, none of the five predictions for ethanol were supported by the data. One of the most interesting findings involved a gene that was previously a mystery, with no clear function assigned to it. The computer predicted that this unknown gene was important for resisting tunicamycin, and the experiment proved it right, showing a significant drop in growth when the gene was missing. This suggests the computer can find new biological insights even for genes that scientists know very little about.

The study also revealed something about how yeast handles stress. The genes that the computer successfully identified were often involved in building or maintaining the cell's outer layers and its internal transport systems. For example, one of the strongest findings involved a gene called GET2, which helps insert proteins into the cell's internal membrane. When this gene was removed, the yeast suffered a massive 83.4% drop in its ability to grow under hydrogen peroxide stress. Another successful prediction involved a gene called CAX4, which helps process sugars attached to proteins. When this gene was missing, the yeast struggled significantly under caffeine stress, showing a 44% drop in fitness. These results suggest that the cell's ability to manage its internal structure and transport materials is critical for surviving chemical attacks. However, the researchers noted that the computer's confidence scores did not perfectly match the real-world results. While the computer gave high scores to many predictions, only about 30% of the top guesses for the target chemicals turned out to be correct. This indicates that while the system is good at narrowing down the possibilities, it cannot yet predict the outcome with absolute certainty.

To ensure that this new knowledge is preserved and useful for future scientists, the team developed a way to record the entire process, including both the successes and the failures. They extended a digital framework called RIMBO, which is designed to track how scientific models change over time. In this system, every experiment is logged with its evidence. If a hypothesis was supported, the new connection is added to the digital knowledge graph. If a hypothesis was not supported, it is not thrown away; instead, it is recorded as a failed attempt with all the data attached. This creates a transparent history of discovery, allowing future researchers to see not just what is known, but what has been tested and ruled out. This approach turns the scientific process into a continuous loop: the computer makes a guess, the robot tests it, the results are recorded, and the updated knowledge is fed back into the system to make better guesses next time.

The study demonstrates that combining artificial intelligence with automated experimentation is a powerful way to navigate the complexity of biology. By using a knowledge graph to prioritize which experiments to run, scientists can focus their efforts on the most promising leads, saving time and resources. While the system is not perfect and still requires human oversight and further investigation, it successfully identified new relationships between genes and chemicals that were previously unknown. The work highlights that even in a field as complex as biology, structured data and machine learning can guide discovery, turning a chaotic pile of facts into a navigable map of life's inner workings. The researchers plan to use these findings to explore the supported interactions further, potentially using other types of data to understand exactly how these genes protect the cell. For now, the study stands as a proof that computers can help scientists ask the right questions, even if the answers still require the careful touch of a laboratory.