Proteins are not always neat, tidy molecules. Some fold into stable shapes, but a large portion of the human proteome contains stretches that stay deliberately unstructured. These intrinsically disordered regions behave in ways that can be useful inside a healthy cell, or damaging when something goes wrong. Two particularly important behaviors are amyloid formation, where proteins stack into rigid, sticky fibers, and liquid-liquid phase separation, where proteins cluster into droplet-like compartments that can influence how the cell reads its own genetic instructions.
Researchers publishing in the Proceedings of the National Academy of Sciences describe a new computational framework containing two predictive tools, called amyloid-predict and LLPS-predict. Together, these tools can rapidly scan every disordered protein region in the human proteome and flag which ones are most likely to undergo either of those two behaviors. The work represents a significant step in understanding protein biology at a systems-wide scale, and the authors suggest it could accelerate disease research as well as the rational design of peptide-based therapeutics.
Two distinct but related protein behaviors
Amyloid formation happens when short segments of a protein stack on top of one another in a very regular, repeating pattern. The resulting fiber is extremely stable and difficult for the cell to break down. Many well-studied conditions, including certain neurodegenerative diseases, involve the buildup of amyloid fibers in tissues where they cause damage over time.
Liquid-liquid phase separation is a different process. Here, proteins concentrate themselves into tiny, droplet-like zones inside the cell, similar in concept to how oil separates from water. These condensates are not solid, they are dynamic and reversible under normal conditions. They play roles in organizing genetic activity, managing stress responses, and coordinating how the cell splices its RNA. However, when phase separation becomes uncontrolled or the droplets solidify, the result can contribute to pathological states.
Both phenomena are most common in proteins that contain disordered regions, and both can occur in the same protein. The researchers noted that several known amyloid-forming and prion-like proteins showed high propensity scores on both predictors simultaneously, suggesting a mechanistic link between the two behaviors worth studying further.
How the predictors work
Both tools are built on protein language model embeddings. A protein language model is a type of artificial intelligence trained on enormous libraries of protein sequences. Much like a text-based language model learns the statistical patterns of words in sentences, a protein language model learns the statistical patterns of amino acids in protein chains. The embeddings it produces are numerical representations that capture rich information about context, not just which amino acids are present but how they are arranged relative to one another.
The researchers trained classifiers on top of these embeddings to score any given peptide or protein region for its amyloid propensity or LLPS propensity. One key finding is that amyloid-predict is sensitive to subtle single-amino-acid mutations and responds to sequence patterning and context rather than amino acid composition alone. That distinction matters because older, physics-based tools tend to rely on compositional rules, which can miss cases where the specific order of amino acids drives behavior.
On a standard hexapeptide benchmark, amyloid-predict outperformed both existing AI-based tools and physics-based competitors, while also running substantially faster. Speed is important when the goal is scanning millions of protein segments across an entire proteome.
A map of the disordered human proteome
The team applied both classifiers to all intrinsically disordered regions across the human proteome, producing what the authors describe as side-by-side landscapes of amyloid potential and LLPS potential. Several protein categories stood out with notably elevated scores.
Signaling receptors, carbohydrate-binding proteins, and calcium-binding proteins showed enrichment in aggregation propensity. That finding is interesting because signaling receptors are central to how cells communicate and respond to the environment, and the authors suggest elevated amyloid propensity in their disordered regions may reflect functional properties that have not been fully characterized.
On the LLPS side, mRNA-binding proteins, ribonucleoprotein complexes, and nuclear matrix proteins showed the strongest enrichment. This aligns with known biology, since many of these protein families are already associated with the formation of membraneless organelles like stress granules and splicing bodies. The computational results provide a quantitative confirmation and extend the picture to protein categories that had not previously been examined systematically.
Relevance to disease mechanism research
One of the main motivations stated by the researchers is understanding how protein misbehavior contributes to disease. Amyloid accumulation and dysregulated phase separation both appear in the literature as contributors to conditions ranging from neurodegeneration to certain metabolic disorders. Having a fast, accurate tool to flag which proteins carry high intrinsic risk for either behavior could help researchers prioritize which targets to study in cell models or animal studies.
The tools are particularly notable for their sensitivity to mutation effects. Many disease-relevant variants involve a single amino acid change that shifts a protein from a safe state toward a dangerous one. Because amyloid-predict captures sequence context rather than just composition, the researchers suggest it can detect these subtle shifts in a way that older tools cannot. Early data points at this being relevant for studying both inherited and sporadic disease mechanisms.
Implications for peptide therapeutic design
The authors specifically mention rational design of peptide therapeutics as one application area for the framework. Peptides intended for therapeutic research need to avoid unintended aggregation, since a peptide that forms amyloid-like structures may lose activity or become problematic in other ways. The ability to screen candidate sequences rapidly against both predictors could help researchers identify and engineer out aggregation-prone segments early in the design process.
On the positive side, researchers studying peptides that interact with phase-separated condensates could use LLPS-predict to understand whether a given sequence might be recruited into those droplets or whether it would remain outside them. The literature on condensate biology is growing quickly, and tools that make it easier to predict condensate interactions from sequence alone could prove valuable across multiple research disciplines.
The framework is described as enabling proteome-wide annotation of individual peptides and residues, meaning the resolution is fine enough to pinpoint specific regions within a larger protein rather than just flagging the whole protein. That level of detail supports the design of targeted interventions that address specific problematic segments.
What comes next in this research area
The published work focuses on prediction and annotation rather than experimental validation at scale, which is a natural limitation of any computational study. The researchers note the value for disease-mechanism studies and drug design, but experimental follow-up will be needed to confirm whether the highest-scoring proteins identified in the proteome scan do indeed form amyloid or phase-separate under biologically relevant conditions.
The observation that certain proteins show high propensity on both predictors simultaneously is highlighted as a particularly interesting finding. The relationship between amyloid formation and phase separation is an active area of research, and understanding how one behavior might seed or reinforce the other could clarify mechanisms that have been difficult to study with existing tools.
As protein language models continue to improve in accuracy and coverage, frameworks like this one are likely to become more precise. The approach of layering task-specific classifiers on top of general-purpose language model embeddings is increasingly common in computational biology, and the results described in this study suggest it is a productive direction for studying disordered protein behavior at scale.
