CANYA

What is CANYA?

CANYA (Convolution Attention Network for amYloid Aggregation) predicts amyloid nucleation propensity from protein sequence. The model was trained on aggregation measurements for more than 100,000 random peptides, then tested on independent sequences. It learns short sequence motifs and interactions between them through a compact convolution and attention architecture.

CANYA scores sequence propensity, not the stability of a folded protein or the rate of aggregation under a specific formulation. Because the training assay used surface-accessible 20-residue peptides fused to Sup35N, predictions for regions buried inside a folded protein still need structural and experimental context. Aggrescan3D complements CANYA when a three-dimensional structure is available.

How to use CANYA online

Run CANYA online by submitting one or more protein sequences as FASTA records or a two-column TSV, choosing the published model or 10-model ensemble, and selecting how 20-residue windows should be summarized. ProteinIQ returns a spreadsheet of nucleation scores and the original tab-delimited CANYA result file.

Inputs

Input format	Requirements
FASTA	One or more records with the first > at the very start of the file. Sequences must use uppercase one-letter amino acid codes without internal spaces.
Two-column TSV	One record per line, with sequence ID and amino acid sequence separated by exactly one tab. Do not include a header row.

CANYA predicts the 20 standard amino acids. A stop symbol (*) truncates the sequence at that position. Records containing X or Z are accepted because the CANYA program skips them, but every submission must contain at least one record CANYA can predict. ProteinIQ reports how many records were skipped, and skipped records are absent from the results.

Settings

Setting	Description
Mode	Default model runs the trained instance used for interpretation in the paper. 10-model ensemble averages the ten most interpretable trained instances and also reports their score standard deviation.
Summarize	For sequences longer than 20 residues, combines overlapping 20-residue window scores using Median (default), Mean, Maximum, or Minimum. Per-window scores keeps every window instead.

The published work recommends the median for longer sequences because it was the most stable summary across the authors' evaluations. Maximum can help locate a strong local hotspot, but it answers a different question from the overall median and is more sensitive to a single high-scoring window.

The 10-model ensemble always returns per-window predictions and standard deviations, regardless of the selected summary.

Results

The columns depend on the chosen mode and summary:

Column	Meaning
Sequence ID	FASTA header or identifier from the first TSV column.
CANYA nucleation score	Sequence-level score after applying the selected summary. Higher values indicate greater predicted amyloid nucleation propensity.
Window sequence	The 20-residue subsequence scored when per-window output is returned.
Position	Location of the scored window in the source sequence.
CANYA prediction	Mean prediction for a window in ensemble output.
Model standard deviation	Variation across the ten ensemble models, which reflects model uncertainty rather than experimental error.

The Files tab contains the native _canya.tsv output for downstream analysis. In workflows, score rows and the complete native TSV are available as separate outputs so you can route either representation to the next step.

How CANYA works

CANYA first converts a peptide into a one-hot encoded amino acid matrix. A convolutional layer with 100 filters learns short motifs, then a self-attention layer models positional effects and interactions between those motifs. A 64-unit dense layer feeds a sigmoid output. The complete model has 17,491 parameters.

The original model accepts up to 20 residues at once. ProteinIQ preserves CANYA's sliding-window behavior for longer proteins by scoring every overlapping 20-residue segment and applying the selected summary function.

Interpreting CANYA scores

CANYA scores rank sequences by the pattern learned from its aggregation assay. A higher score supports greater nucleation propensity within that learned sequence context, but the value is not a kinetic rate constant and should not be read as the percentage of protein that will aggregate.

Useful comparisons keep the following factors constant:

• Summary function: Median, maximum, and minimum scores are not interchangeable.
• Sequence context: A high-scoring motif may be solvent-accessible in one protein and buried in another.
• Construct design: Tags, truncations, mutations, and terminal context can change experimental aggregation.
• Ensemble uncertainty: A large Model standard deviation means the trained models disagree. Such windows deserve less confidence than equally scored windows with low disagreement.

For engineering work, per-window scores are usually the most actionable output because they identify the exact region driving a sequence-level result. Candidate substitutions can then be checked against protein solubility, structural aggregation patches from Aggrescan3D, and experimental expression or aggregation assays.