CPP: Identifying Physicochemical Signatures
The central algorithm of the AAanalysis framework is Comparative Physicochemical Profiling (CPP), a sequence-based
feature engineering algorithm for interpretable protein prediction [Breimann25]. CPP enables the identification
of physicochemical signatures underlying biological recognition processes. It thereby extends rational protein
biology beyond mere sequence motifs.
Provided by
In AAanalysis this is the CPP class, with
SequenceFeature for building parts and splits and
CPPPlot for the figures. The golden pipeline
find_features() automates the search. See the
API reference, the tutorials, and the
Evaluation Regimes chapter for how to read a
feature-selection score.
The core idea of CPP is its feature concept:
Scheme of CPP feature (Part-Split-Scale combination) with example of feature creation, from [Breimann25].
All possible parts are sub-parts or combinations of the Target Middle Domain (TMD), Juxta Middle Domain N-terminal (JMD-N), and Juxta Middle Domain C-terminal (JMD-C).
TMD: from Transmembrane Domain to Target Middle Domain
The letters come from CPP’s first application, to substrates of γ-secretase, where the span of interest is the transmembrane domain and the flanks are the juxtamembrane domains. When the method was generalized beyond membrane proteins the acronyms were deliberately kept and re-read: TMD as Target Middle Domain, JMD as Juxta Middle Domain. Same letters, same geometry, wider meaning.
So the vocabulary is a geometry - one target span flanked by two juxta spans - and it
already applies unchanged to a Pfam domain, a kinase domain, a cleavage-site window or a
whole chain. A feature id such as TMD-Segment(2,4)-ANDN920101 names the target span; it
describes a transmembrane helix only when the data are membrane proteins, as in the
γ-secretase case below. There is nothing to rename per data set.
Scheme of sequence Parts comprising three basic parts from which each other can be derived from ([Breimann25]): target middle domain (TMD), N- and C-terminal juxta middle domain (JMD-N and JMD-C).
These names were generalized from the first application of CPP on predicting substrates of γ-secretase, which is a pivotal intramembrane protease implicated in cancer and Alzheimer´s disease. γ-Secretases cleaves its substrates within their transmembrane domain (TMD) and their N- and C-terminal juxtamembrane domains (JMDs) are of high importance for recognition. The three different split types (Segment, Pattern, and PeriodicPattern) are exemplified for the two prominent γ-secretase substrates: the amyloid precursor protein (APP) and NOTCH1:
Scheme of part Splits, exemplifying the three split types ([Breimann25]): segments, patterns, and periodic patterns.
CPP uses by default 120 continuous (segments) and 210 discontinuous (patterns and periodic patterns) splits:
Overview of Split classification and their count using default CPP settings, from [Breimann25].
Scales can be chosen from AAontology, our two-level scale classification, based on their category or subcategory
classification. To then select a redundancy-reduced scale set, AAanalysis provides the AAclust clustering wrapper.
Once features are identified, the number a CPP or model score reports depends on how it was
produced. The default exploratory search ranks features but does not estimate generalization,
while an honest nested regime, a final refit, and an external test set each answer a different
question. The find_features() selection_scope knob moves feature
selection in and out of the cross-validation to switch between them. See the
Evaluation Regimes chapter for how to read each score.