KANG2P enables interpretable genotype-to-phenotype prediction through genetically anchored molecular-state learning
KANG2P is a scalable, interpretable framework that enhances genotype-to-phenotype prediction accuracy and mechanistic insight by integrating genetically predicted molecular profiles via modality-specific autoencoders and Kolmogorov–Arnold Networks, successfully identifying novel disease genes validated through cross-species functional analysis.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Genetics has long promised to explain how the code of life, written in DNA, translates into the living, breathing reality of an organism. This translation, known as mapping genotype to phenotype, is the central challenge of modern biology. It is the difference between reading a list of ingredients and understanding how they combine to create a cake. While scientists have become very good at reading the ingredients—the specific variations in DNA that make one person susceptible to a disease or one plant taller than another—understanding the complex, invisible steps in between remains difficult. These steps involve a cascade of molecular events, where genes turn on and off to produce proteins and other molecules that ultimately shape physical traits. The gap between the static DNA code and the dynamic, complex outcome is vast, and bridging it has required tools that can handle immense complexity without losing the ability to explain why a prediction was made.
A new framework called KANG2P offers a fresh approach to this problem, designed to predict physical traits and disease risks while keeping the biological story clear. Instead of treating DNA as a simple list of switches, this system acts as a translator that reconstructs the hidden molecular steps. It starts with genetic data and uses information learned from large reference groups to imagine what the molecular landscape looks like for a specific individual. It then combines this imagined molecular state with the original genetic code to predict the final outcome, whether that is the risk of a neurodegenerative disease in a human or the yield of a crop in a field. What makes this approach distinct is that it does not just guess the answer; it reveals the path taken to get there, identifying which specific genes and molecular pathways are driving the result.
The researchers behind KANG2P built their system to solve a specific bottleneck in biological research. Traditionally, to predict a complex trait accurately, scientists needed to measure the DNA, the RNA, and the proteins of the exact same individuals. This is like trying to understand a car by measuring the engine, the wheels, and the fuel at the same time, but for millions of people or plants. This is often impossible because taking tissue samples from the human brain or harvesting specific plant parts is too expensive, too invasive, or simply not feasible for large groups. KANG2P bypasses this hurdle by using a two-step strategy. First, it learns the rules of how DNA influences molecular activity from large, existing datasets where both types of information are available. Then, it applies these rules to new groups where only DNA is available, effectively reconstructing the missing molecular picture from the genetic code alone.
Once the system has reconstructed these molecular profiles, it feeds them into a sophisticated mathematical engine known as a Kolmogorov–Arnold Network. Unlike many modern artificial intelligence models that act as "black boxes," making predictions without revealing their internal logic, this engine is designed to be transparent. It learns the complex, non-linear relationships between genes, molecules, and traits in a way that allows scientists to trace the influence of every single gene back to the final prediction. The researchers tested this framework on two very different types of problems: predicting the risk of amyotrophic lateral sclerosis (ALS) and Alzheimer's disease in humans, and forecasting agricultural traits like grain yield and plant height in rice and maize.
In the human disease studies, the system proved remarkably effective. When predicting the risk of ALS, KANG2P outperformed existing methods, including other advanced artificial intelligence models. More importantly, it successfully identified genes already known to be involved in the disease, such as SOD1, while also highlighting new candidates that had not been previously prioritized. One of these new candidates, a gene called SLC1A2, was selected for further investigation. The researchers turned to the tiny roundworm C. elegans, which shares a version of this gene, to test its function. They found that worms with a broken version of this gene struggled to move and were significantly more sensitive to oxidative stress, a type of cellular damage linked to ALS. This experimental validation confirmed that the computer's prediction was not just a statistical fluke but pointed to a real biological mechanism. The same success was seen with Alzheimer's disease, where the system integrated genetic data with predicted protein levels to improve risk prediction accuracy.
The framework also demonstrated its power in agriculture, where predicting crop traits is essential for food security. The researchers applied KANG2P to large populations of rice and maize. In these tests, the system consistently predicted traits like grain yield and plant height more accurately than traditional statistical models and other deep-learning approaches. A key finding was that the system's ability to reconstruct the missing molecular layers helped even the older, simpler models. When the researchers fed the reconstructed molecular data into these established models, their performance improved, suggesting that the molecular information captured a layer of biological reality that DNA alone could not fully express. This implies that the value of the approach extends beyond the specific new model; it provides a way to enrich any genetic prediction with biological context.
The strength of KANG2P lies in its ability to turn a massive, complex dataset into a clear, actionable story. By anchoring its predictions in genetically regulated molecular states, it bridges the gap between the static code of DNA and the dynamic reality of life. It does not require the impossible task of measuring every molecule in every individual, yet it still manages to see the molecular picture. This makes it a scalable tool for understanding complex traits across species, from the human brain to the rice paddy. The work suggests that by combining the power of large reference datasets with transparent, interpretable machine learning, scientists can move beyond simple prediction to genuine biological discovery, identifying the specific genes and pathways that drive health and disease.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.