Embedding Atlas

Explore how amino-acid representations relate before applying a sequence transform. This atlas compares individual property channels and the complete residue geometry of embedding matrices. It includes BLOSUM100, the ten Kidera factors, and valid AAindex1/2/3 properties and matrix profiles.

Hover to inspect identifiers and descriptions, drag to zoom, double-click to reset, and click legend entries to hide or show families. BLOSUM100 appears as red stars and Kidera as purple X markers, both with thin black outlines. The figures below are exports of the current Explore the residue-embedding atlas notebook at its default settings; changing hyperparameters requires rerunning that notebook. No code cells are included here.

How the analysis was performed

  • Each individual channel is centered across the 20 canonical amino acids and scaled to unit L2 norm. Default sign-invariant comparison uses squared Pearson correlation, so opposite-sign profiles have the same shape. Duplicate shapes share a point; hover text retains their aliases. This is normalization over amino acids, with no peptide-data fitting or outcome labels.

  • For whole representations, each channel is centered and normalized separately to give it equal weight. The residue Gram matrix is then Frobenius-normalized; its inner product with another normalized Gram gives linear CKA. This compares residue geometry even when channel counts differ; it does not whiten the embedding. Constant channels do not contribute.

  • PCA is fitted to vector projectors or normalized Gram features. UMAP uses their original-space distances, sqrt(2 * (1 - similarity)), rather than PCA coordinates. UMAP uses 15 neighbors, minimum distance 0.15, seed 0, and 300 epochs for channels or 500 for whole representations. PCA fits ten components and displays the first two; axis labels report explained variance.

  • Hierarchical clustering uses complete linkage on the original-space distances. Similarity, clustering, and heatmap values are independent of the 2D layouts. Distances on PCA/UMAP plots can distort the full geometry and do not measure predictive performance.

Individual property channels

PCA

Each point is a unique centered property shape across amino acids. PCA uses sign-invariant projector features, so properties with strong positive or negative correlation are close in the comparison space. Family labels identify a preferred source for shared shapes, rather than exclusive ownership of that property.

UMAP

UMAP emphasizes local neighborhoods in the same sign-invariant original space. Hover on Kidera points to see the property descriptions, such as hydrophobicity, alongside stable factor identifiers. Islands and visual gaps depend on the UMAP settings; use original-space similarities for quantitative comparisons.

Whole representations and matrix geometry

CKA-space PCA

Each point now represents a complete embedding configuration. AAindex2/3 matrices supply canonical row or column profiles; asymmetric orientations can therefore appear separately. Kidera has ten channels, BLOSUM100 has twenty, and the complete AAindex1 property collection is another configuration. CKA compares the residue relationships encoded by these different channel sets.

CKA-space UMAP

The same configurations are placed using the original CKA-derived distances. This view emphasizes local geometric similarity without requiring equal numbers of channels or a shared channel basis.

Original-space CKA heatmap and dendrograms

Colors show the actual pairwise CKA values, not distances inferred from a projection. Both axes follow the same complete-linkage leaf order; the top and left dendrograms show the corresponding original-space linkage distances. Hover to inspect both representation identifiers and their similarity. Zooming the heatmap keeps the corresponding dendrogram leaf axes aligned.

Inspect the embedding matrix

This default notebook view shows BLOSUM100 with sequence residues as rows and reference channels as columns. The colors are the raw signed scores, without channel normalization. This inspection panel is distinct from the normalized geometry comparisons above; the notebook allows selecting other representations.

Signed aligned matrix-score PCA

Only square matrices with matching canonical residue axes enter this comparison. Each matrix is flattened after global mean removal and unit-norm scaling. This preserves channel labels, signs, and relative score patterns, unlike the CKA comparison of residue geometry. Kidera and the complete scalar-index collection are excluded because their channel axes do not match the square matrices.

Sources and reproducibility

The AAindex files are explicitly downloaded from the official database and SHA256-checked before analysis. Raw database files are not bundled in this site. These are derived exploratory plots, not a supervised feature-selection result. Kidera labels describe dominant properties; factors 5–10 combine multiple properties. See Residue representations, AAindex1 representations, and AAindex2 and AAindex3 matrix profiles for sources and interpretation.

The Explore the residue-embedding atlas page links the editable notebook and explains its controls. The plot export manifest records source, data-snapshot, and figure hashes. To explore sequence spectral features after choosing channels, continue with Compare the four transforms.