CANYA Is Now Available on Vecura — Predict Amyloid Nucleation Propensity from Protein Sequence

What is CANYA?

CANYA (Convolution-Attention Neural Network for Amyloid prediction) is a deep learning model that predicts amyloid nucleation propensity — the likelihood that a protein or peptide segment will seed amyloid aggregation — directly from amino-acid sequence. It combines a convolutional layer with a multi-head attention mechanism to capture local sequence patterns and positional dependencies within 20-residue windows.

It helps users rapidly screen protein and peptide sequences for aggregation-prone regions, generating a per-residue nucleation propensity profile along the length of any input protein. It is especially useful for triaging candidate peptide therapeutics, assessing aggregation liability in engineered protein designs, and identifying which regions of a longer protein are most likely to drive amyloid nucleation — tasks that would otherwise require slow, low-throughput experimental assays.

What can users do with CANYA on Vecura?

With CANYA on Vecura, users can:

• Screen protein and peptide sequences for amyloid nucleation risk — Upload sequences in FASTA or tab-delimited format and receive per-sequence nucleation scores, enabling rapid triage of candidate designs for aggregation liability.

• Generate residue-level nucleation propensity profiles — For proteins longer than 20 amino acids, CANYA tiles overlapping 20-mer windows across the full sequence, producing a positional profile that pinpoints exactly which regions are driving aggregation risk.

• Run ensemble inference for uncertainty-aware predictions — Switch to ensemble mode to score every window with all ten bundled model checkpoints, obtaining both a mean prediction and an across-model standard deviation as an epistemic uncertainty estimate.

• Summarize results at the sequence level — Choose how to collapse per-window scores into a single score per sequence using median, mean, max, or min aggregation, making it easy to rank and compare large libraries of sequences.

What the output means

The output provides a nucleation propensity score (sigmoid output between 0 and 1, where higher values indicate greater predicted amyloid nucleation propensity) for each input sequence and each 20-residue window. In ensemble mode, an additional column reports the across-model standard deviation as a measure of prediction confidence. Results are delivered as structured tables and a downloadable TSV file for downstream analysis.

This output should be used to support scientific decision making. It does not replace experimental validation. CANYA scores reflect predictions from a model trained on a specific in-vivo amyloid nucleation assay and capture nucleation propensity within a 20-residue context window — they do not account for long-range structural effects, full-protein folding, or environmental conditions.

Why this matters

Amyloid nucleation is the critical first step in protein misfolding, a process underlying neurodegenerative diseases such as Alzheimer's, Parkinson's, and ALS, as well as a growing concern in the development of peptide therapeutics and engineered proteins. Experimentally measuring nucleation propensity has traditionally been slow and low-throughput, creating a bottleneck in both disease research and protein design pipelines. Earlier computational aggregation predictors were limited by the size and bias of their training data, often performing inconsistently across diverse sequence contexts.

CANYA was trained on a massively parallel in-vivo assay that quantified amyloid nucleation across more than 100,000 20-amino-acid peptides — a dataset large and unbiased enough to expose the performance limitations of prior tools. By combining convolutional feature extraction with attention-based positional modeling, CANYA achieves stable performance across diverse prediction tasks. Its compact architecture (113 KB per checkpoint) runs efficiently on a single CPU core, scoring approximately 100,000 sequences in under a minute — making high-throughput nucleation screening accessible to any research lab without specialized hardware.

• Developed by: Lehner Lab (CRG — Centre for Genomic Regulation, Barcelona)

• Source: GitHub Repository (lehner-lab/canya)

• Reference: Thompson et al., "Massive experimental quantification of amyloid nucleation allows interpretable deep learning of protein aggregation," Science Advances 11, eadt5111 (2025). DOI: 10.1126/sciadv.adt5111 | bioRxiv preprint