Tutorials

Tutorials teach the AAanalysis tools — what each one does, its parameters, and the outputs it returns. They cover the mechanics; for how to combine tools into a valid end-to-end analysis, see the Protocols, which link back here for the mechanics instead of repeating them — so the two stay distinct with no overlap. New to AAanalysis? Begin with Getting Started for your first result, then return here to go deeper on each tool.

The tutorials at a glance — each tile is the headline figure of one tutorial; click it to open that tutorial. The full, section-ordered list follows below.

Load datasets

Load scales

AAclust

SequenceFeature

CPP

Data representations

Embeddings & AlphaFold

dPULearn

ShapModel

Evaluation

Protein engineering

Upstream bridge

Data Handling

Learn how to load protein benchmarking datasets and amino acid scale sets in the Data Loader and Scale Loader tutorials.

• Data loading
• Scale loading

Feature Engineering

Explore interpretable feature engineering, the core of AAanalysis, with the AAclust, SequenceFeature, and CPP tutorials, then see how CPP turns different data representations (scales, embeddings, structure) into features. The Embeddings & AlphaFold tutorial goes deeper on the two heavy representations: the recommended way to feed protein language model embeddings (EmbeddingPreprocessor) and AlphaFold-derived channels (StructurePreprocessor) into run_num(), fuse them, and paint the signature onto the 3D model with CPPStructurePlot. Because feature_matrix() returns a plain numeric matrix, these features drop directly into a stock scikit-learn Pipeline — the prediction protocol demonstrates this end to end.

• AAclust: Selecting redundancy-reduced scale sets
• SequenceFeature: Creation of CPP feature components
• CPP: Identification of physicochemical signatures
• CPP across data representations: sequence, structure, embeddings, annotation
• CPP with protein language model embeddings and AlphaFold structures

PU Learning

Start positive-Unlabeled (PU) learning to tackle unbalanced and small data through our dPULearn tutorial.

• dPULearn: Learning from unbalanced data

Explainable AI

Explaining sample level predictions at single-residue resolution is introduced in our ShapModel tutorial.

• ShapModel: Explaining with single-residue resolution

Evaluation & Comparison

Learn the evaluation tools — CPPGrid configuration sweeps, per-protein site-localization metrics, and fair ranking under cross-validation — in the CPPGrid tutorial. These are the mechanics that the P10: Validation protocol puts to work end to end.

• CPPGrid: Sweeping, evaluating, and ranking configurations

Protein Engineering

Optimize an existing sequence with SeqOpt — machine-learning-guided directed evolution — and read the results with SeqOptPlot. This is protein engineering (mutating a known protein), distinct from de novo protein design (generating new proteins, e.g. RFdiffusion → ProteinMPNN → AlphaFold). The SeqOpt tutorial walks a complete case study: training a substrate classifier, engineering a “super substrate” for gamma-secretase, and visualizing the Pareto front, convergence, mutation map and lineage.

• SeqOpt: Optimizing sequences by directed evolution

Interoperability

Most analyses start with another tool’s output. The Upstream bridge tutorial is the recipe that carries a plain FASTA file, the one hand-off every tool can produce, all the way to a fitted model: read_fasta() reads it, get_df_parts() adapts it to the part geometry CPP needs, and SequenceFeatureTransformer drops into a stock scikit-learn Pipeline so CPP feature selection runs inside cross-validation instead of before it. This is the counterpart to the fixed-feature route, where feature_matrix() hands a plain numeric matrix to any estimator. The two heavy representations, language model embeddings and AlphaFold channels, have their own bridge in the Embeddings & AlphaFold tutorial above.

• Upstream bridge: from another tool’s FASTA to a scikit-learn pipeline