AminoGrapes: Structure-Weighted Evolutionary Scoring for Viral Protein Fitness Prediction

Viruses are masters of change. To survive, they constantly rewrite their own genetic instructions, swapping out tiny building blocks called amino acids to escape our immune systems or to become better at infecting cells. Scientists want to predict which of these changes will help the virus and which will break it, a task that is crucial for designing vaccines and understanding how diseases spread. The challenge is that viruses evolve so quickly that there are often very few examples of them to study, making it hard to learn their rules. Traditionally, researchers have used two main ways to guess the outcome of a mutation. One method looks at history, scanning thousands of related viral sequences to see which parts have stayed the same over time; if a spot never changes, it is likely essential. The other method relies on the physical shape of the protein, knowing that changes to the tightly packed, hidden core of a molecule are more likely to cause damage than changes to its exposed surface.

A researcher named Muhammad Saad Khan, working at Afrium in Islamabad, has combined these two approaches into a new tool called AminoGrapes. The goal was to create a simple, fast way to predict how well a viral protein will function after a mutation, specifically for the difficult cases where there is not enough data for the most advanced computer models to work well. The method takes a viral protein sequence and does two things. First, it builds a family tree of similar viruses to calculate how much each position in the protein has been conserved by evolution. Second, it uses a computer-generated 3D model of the protein to determine how buried or exposed each position is. The tool then multiplies these two pieces of information together. If a position is both highly conserved by history and buried deep inside the protein's core, a mutation there is given a severe penalty. If a position is on the surface or has changed frequently in the past, the penalty is much lighter. This creates a single score that predicts whether a specific change will be tolerated by the virus.

The researchers tested this tool on a standard set of thirty-one different viral protein experiments, known as assays, which measure how well mutated proteins actually work in a lab. They compared their new tool against a powerful, widely used artificial intelligence model called ESM2, which is a large language model trained on millions of protein sequences. The results showed that AminoGrapes was significantly better at predicting the fitness of viral proteins than the standard AI model. On average, the new tool matched the real-world experimental results with a correlation of 0.430, while the AI model only reached 0.269. In twenty-five out of the thirty-one tests, AminoGrapes was the more accurate predictor. This improvement came from the fact that the new tool explicitly used the physical structure of the protein and the history of its family, whereas the AI model tried to guess based only on the sequence of letters in the genetic code.

However, the story is not one of total victory. When the researchers compared AminoGrapes to the very best methods currently available in the scientific community, it fell short. Other established tools, which often use more complex mathematics or larger datasets, achieved higher scores, with the top performer reaching 0.480. Notably, AminoGrapes scored significantly lower than the original RSALOR implementation, which achieved 0.480 on the same assays. The author was careful to note that their tool did not add any new value when combined with these top-tier methods; in fact, adding AminoGrapes to a group of the best existing models actually made the combined prediction slightly worse. This suggests that while the new tool is a solid, independent step forward for viral proteins, it does not yet surpass the state of the art. Furthermore, the researchers admitted that the tool was developed while looking at the results of these specific viral tests, which means the numbers might be slightly optimistic and need to be confirmed on new, unseen data in the future.

The study also explored what happens when you mix the new tool with the old AI model. They tried blending the two predictions together to see if they could get the best of both worlds. On non-viral proteins, where the AI model is very strong, the blend worked well. But on the viral proteins, the blend performed worse than the new tool alone. This happened because the AI model is generally weaker on viruses, and mixing its lower-quality predictions with the new tool's higher-quality ones dragged the average down. This finding highlights a specific weakness in current AI models when applied to fast-evolving viruses and suggests that for these specific biological problems, a simpler method that directly uses evolutionary history and 3D structure is currently more reliable than a massive, general-purpose language model.

Ultimately, this work demonstrates that for viral proteins, a straightforward calculation that respects both the deep history of the virus and the physical constraints of its shape can outperform sophisticated artificial intelligence. It does not claim to have solved the problem of predicting viral evolution, nor does it claim to be the most powerful tool available overall. Instead, it offers a clear, interpretable, and effective alternative for a specific, difficult corner of biology. The researchers emphasize that their tool is not a magic bullet, but a practical instrument that fills a gap where current AI struggles, providing a better way to understand which mutations might allow a virus to survive and which will cause it to fail.