Explore the residue-embedding atlas¶
The atlas compares embeddings before sequence-specific transforms. It helps identify redundant residue-property channels and similar whole representations; it does not directly measure antimicrobial performance or spectral equivalence.
See also
Download the editable atlas notebook.
Work from the checkout so its helpers and reference results are available.
Two views, two meanings¶
Individual vectors. Align channels to the same 20 residues, center and normalize, and optionally identify sign reversals. The sign-invariant comparison uses vector outer products, with squared correlation as similarity. Complete linkage cuts give explicit pairwise similarity guarantees. PCA and UMAP are visualizations; clustering and neighbors use original-space distances.
Whole representations. Compare centered residue Gram matrices with linear centered kernel alignment, explicitly choosing whether to equalize channel norms. This handles unequal channel counts. A separate aligned-matrix-score view keeps signed 20-by-20 substitution/contact scores. Geometry similarity does not imply equal per-channel Fourier magnitudes: rotating a channel basis can preserve a Gram matrix while changing channel coefficients.
Inputs and controls¶
Place official AAindex1/2/3 snapshots in ignored .aaindex/. Downloads happen
only after the notebook’s explicit DOWNLOAD_MISSING=True choice. Source
checksum verification is on by default. Raw databases are not distributed.
Install the research requirements documented in
research/embedding_atlas/README.md; UMAP and Plotly are not package runtime
requirements or part of the small offline tutorial notebooks.
Edit family/orientation filters, sign conventions, channel weighting, PCA axes, UMAP neighbors/minimum distance/seed/epochs, clustering linkage/cuts, and display settings. Refit affected stages after changing inputs. Plot-only edits do not require a new UMAP fit. Fixed-seed layouts are cached by input/settings. Exports reject stale fitted stages and preserve a new run’s settings/provenance.
Interpretation limits¶
Uniform residue comparisons ignore composition, order, task labels, and frequency-dependent relevance. Apparent UMAP islands are not quantitative clusters. Raw-offset equivalence also depends on centering and padding conventions. Treat atlas-based reduction as a candidate-search heuristic and evaluate the reduced channels in a training-only pipeline before claiming predictive equivalence.