An important aspect of computational immunology is modeling the properties of peptides (short strings of amino acids). Peptides can arise as substrings cut out of a larger protein, naturally occurring small proteins, or be synthesized for therapeutic purposes. To make useful clinical and research predictions (i.e. "which peptides should go in this vaccine?") we need to partition the combinatorial space of peptides into classes such as T-cell epitopes or MHC ligands. One way to capture such distinctions is to collect large volumes of data about peptides and use that data to build statistical models of their immune properties. This library helps you build such models by providing simple Python/NumPy/Pandas interfaces to commonly used immunology and bioinformatics datasets.
Data Sources
iedb
: Immune Epitope Database, large collection of epitope assay results for MHC binding as well as T-cell/B-cell responsestcga
: Variant peptide substrings extracted from TCGA mutations across all cancer typesreference
: Peptide substrings from the human reference protein sequenceimma2
: IMMA2 epitope immunogenic vs. non-immunogenic data set used by Tung et al. for evaluating the POPISK immunogenicity predictorcalis
: Two datasets used in Calis et al.'s Properties of MHC Class I Presented Peptides That Enhance Immunogenicityhpv
: Human Papillomavirus T cell Antigen Databasetoxin
: Toxic protein sequences from Animal Toxin Databsedanafarber
: Dana Farber Repository for Machine Learning in Immunologytantigen
: Tumor T-cell Antigen Databasehiv_frahm
: Reactions to HIV epitopes across different ethnicities (from LANL HIV Databases)cri_tumor_antigens
: Tumor associated peptides from Cancer Immunityfritsch_neoepitopes
: Mutated and wildtype tumor epitopes from Fritsch et al. HLA-binding properties of tumor neoepitopes in humans
Planned:
bcipep
: B-cell epitopes
Dataset API
When a dataset consists only of an unlabeled list of epitopes, then it only needs two functions:
load_wuzzle
: Returns set of amino acid stringsload_wuzzle_ngrams
: Array whose rows are amino acids transformed into n-gram vector space.
If the dataset contains additional data about the epitopes (such as HLA type u or source protein):
load_wuzzle
: Returns data frame with epitope strings and additional propertiesload_wuzzle_set
: Set of epitope amino acid stringsload_wuzzle_ngrams
: Array whose rows are amino acids transformed into n-gram vector space.
If the dataset is labeled (contains positive and negative assay results), then the following functions should be available:
load_wuzzle
: Load all available data from the "wuzzle" dataset (filtered by options such asmhc_class
).load_wuzzle_values
: Group the dataset by epitope string and associate each epitope with the positive and negative counts, along with percentage of positive results (in a column called "value").load_wuzzle_classes
: Split the epitopes into positive and negative classes, return a set of strings for each.load_wuzzle_ngrams
: Transform the amino acid string representation (or some reduced alphabet) into vectors of n-gram frequencies, return a sklearn-compatible(samples, labels)
pair of arrays.
Amino Acid Properties
The amino_acid
module contains a variety of physical/chemical properties for both single amino residues and interactions between pairs of residues.
Single residue feature tables are parsed into StringTransformer
objects, which can be treated as dictionaries or will vectorize a string when you call their method transform_string
.
Examples of single residue features:
hydropathy
volume
polarity
pK_side_chain
prct_exposed_residues
hydrophilicity
accessible_surface_area
refractivity
local_flexibility
accessible_surface_area_folded
alpha_helix_score
(Chou-Fasman)beta_sheet_score
(Chou-Fasman)turn_score
(Chou-Fasman)
Pairwise interaction tables are parsed into nested dictionaries, so that the interaction between amino acids x
and y
can be determined from d[x][y]
.
Pairwise interaction dictionaries:
strand_vs_coil
(and its transposecoil_vs_strand
)helix_vs_strand
(and its transposestrand_vs_helix
)helix_vs_coil
(and its transposecoil_vs_helix
)blosum30
blosum50
blosum62
There is also a function to parse the coefficients of the PMBEC similarity matrix, though this currently lives in the separate pmbec
module.