AmyloCore-ML: AI/Machine Learning Enabled Identification of Amyloid Fibril Core Regions Using Protein Language Models

Proteins are the workhorses of life, folding into intricate shapes to perform essential tasks within our cells. However, when these proteins malfunction, they can misfold and clump together into rigid, rope-like structures called amyloid fibrils. This process is at the heart of several devastating neurodegenerative diseases, including Alzheimer's and Parkinson's. For decades, scientists have understood that these fibrils are not random piles of protein; they possess a highly ordered, stable core where the protein strands lock together in a specific pattern. Yet, a critical puzzle remained: while a protein might be long and complex, only a specific segment of it actually becomes part of this rigid, ordered core. The rest of the protein often remains disordered or hangs loosely on the outside. Knowing exactly which amino acids form this solid core is vital for understanding how these diseases develop and for designing drugs that can stop the process, but identifying these specific regions from the protein's genetic code alone has proven to be a difficult challenge.

Researchers have long tried to predict these dangerous clumping regions using computer programs that look for simple chemical patterns, such as stretches of amino acids that naturally like to stick together. However, a new study suggests that these traditional methods miss the mark because they confuse a protein's general tendency to clump with the specific reality of what actually gets locked into the final, rigid structure. To solve this, a team of scientists developed a new tool called AmyloCore-ML. Instead of just looking for sticky patches, this tool was trained to recognize the precise "grammar" of the sequence that ends up inside the solid core of a fibril, distinguishing it from the parts of the same protein that do not. By learning from high-resolution images of real fibrils, the system learned to spot the subtle differences between the ordered core and the disordered remainder, offering a much sharper view of where the danger lies.

The team began by gathering a massive collection of experimentally solved fibril structures from a public database known as the Amyloid Atlas. They carefully selected 69 different proteins, examining the detailed maps of their fibril forms to see exactly which residues were locked in the ordered core and which were left outside. They treated the locked residues as the "core" and the rest of the protein from the same molecule as the "non-core" control. This approach was crucial because it allowed the computer to learn the difference between two parts of the same protein under the same conditions, rather than comparing unrelated proteins. The researchers then broke these protein sequences into overlapping windows of varying lengths and fed them into a machine learning system. They tested two different ways of describing the protein: one using standard chemical properties and another using advanced language models that understand the context of amino acids, much like a reader understands how a word's meaning changes based on the sentence it is in.

The results showed that the advanced language models were far superior to the traditional chemical descriptions. The system learned that the identity of the core is not just about a short, sticky sequence, but depends heavily on the surrounding neighborhood of amino acids. When the researchers tested their best models, they found they could correctly distinguish core regions from non-core regions with high accuracy, achieving a score of roughly 0.88 on a scale where 1.0 is perfect. More importantly, when they scanned entire, full-length proteins to see if they could locate the core, the tool successfully pinpointed the experimental core regions. The best version of the model could locate the center of the core within an average distance of just 12.4 residues, a significant improvement over existing tools that often missed the mark by much larger margins.

To ensure their findings were robust, the team compared their new predictor against two other well-known tools designed to find aggregation-prone regions. The new tool consistently outperformed them, not only in ranking which residues were most likely to be in the core but also in physically locating the correct segment of the protein. The researchers noted that while other tools might identify a region as generally prone to clumping, their model specifically identified the part that becomes structurally incorporated into the mature fibril. This distinction is vital because a region might be capable of starting a clump but not necessarily end up in the final, stable structure. The study also confirmed that using longer windows of amino acids provided better results, suggesting that the signal for forming a core is spread out over a broader area rather than being confined to a tiny, isolated spot.

The team validated their approach by testing it on proteins that were completely new to the system, including horse myoglobin, P02070, Oat 11S globulin, Oat 12S seed storage globulin 2, and human Integral membrane protein 2B. Without any retraining or adjustment, the tool successfully identified high-probability core segments that overlapped with the experimentally resolved structures of these unseen proteins. This demonstrated that the system had learned a general rule about how proteins fold into fibrils, rather than just memorizing the specific examples it was trained on. The researchers made their tool freely available through an interactive online interface, allowing other scientists to input protein sequences and instantly see a map of where the amyloid core is likely to form. By shifting the focus from general stickiness to the specific architecture of the fibril core, this work provides a clearer, more precise way to understand the molecular mechanics of these diseases, potentially guiding the design of future therapies that target the exact surfaces where these harmful structures assemble.