StructureDCA

The structuredca Python package implements Structure-Informed Direct Coupling Analysis (StructureDCA) to predict the effects of missense mutations on proteins.

Standard DCA methods use Multiple Sequence Alignments (MSAs) to build a statistical evolutionary model of homologous protein families. They rely on single-site fields h and pairwise couplings J that capture co-evolution between residue positions. StructureDCA extends this framework by incorporating the residue–residue contact map derived from the protein 3D structure to infer a sparse DCA model, in which couplings between spatially distant residue pairs are removed. This approach leverages the observation that functionally relevant, co-evolving residues are most often structurally in contact.

The package includes a pseudolikelihood-maximization DCA solver capable of inferring sparse DCA models, where selected coupling coefficients Jij are constrained to zero. StructureDCA combines a flexible, user-friendly Python interface with the high computational efficiency of its C++ backend. This model was initially developed to improve classical DCA methods for predicting the effects of missense mutations in proteins. However, StructureDCA can be applied to any DCA-based analysis (except for contact predictions...) and provides the full functionality of both standard and sparse DCA models.

Please cite:

• Matsvei Tsishyn, Hugo Talibart, Marianne Rooman, Fabrizio Pucci. Structure-informed direct coupling analysis improves protein mutational landscape predictions. BioRxiv.

Installation and Usage

Colab Notebook

You can instantly try StructureDCA in this Colab Notebook. This notebook acts as a user-friendly web server / graphical interface, offering helpers to automatically fetch or generate the MSA and 3D structure for your target protein. You can then visualize your mutational landscape predictions as a DMS heatmap or mapped to the 3D structure.

Installation

Installation with pip:

pip install structuredca

CLI usage

Use StructureDCA with a Command Line Interface (CLI). For example, from the directory ./test_data/, run:

structuredca ./6acv_A_29-94.fasta ./6acv_A_29-94.pdb A -o ./6acv_A_29-94_structuredca.csv

To show CLI usage and optional arguments, run:

structuredca --help

Python usage

Make sure the first sequence in your MSA file is the target sequence to mutate (otherwise have a look at tutorial 1).
From directory ./test_data/ execute the following Python code:

# Import
from structuredca import StructureDCA

# Log basic usage and arguments
StructureDCA.help()

# Initialize StructureDCA model
sdca = StructureDCA(
    msa_path='./6acv_A_29-94.fasta', # in .fasta, .a2m or .a3m format
    pdb_path='./6acv_A_29-94.pdb', # in .pdb, .ent or .cif format
    chains='A',
    use_contacts_plddt_filter=False, # use only if 3D structure is an AlphaFold model (or similar) to remove low pLDDT regions from contacts
)

# Evaluate the evolutionary energy difference (ΔE) of mutations
# scores can be reweighted by Relative Solvent Accessibility-complement (RSAc) -> advised to predict stability changes (ΔΔG)
# * dE = 0 means neutral mutation
# * dE >> 0 means destabilizing / deleterious mutation
dE_mut1 = sdca.eval_mutation('K13H', reweight_by_rsa=True)
dE_mut2 = sdca.eval_mutation('K13H:K12G', reweight_by_rsa=True)

# Evaluate ΔE of all single mutations and save results to a file
dE_all = sdca.eval_mutations_table(
    save_path='./6acv_A_29-94_structuredca.csv',
    log_output_sample=True,
)

# Evaluate absolute evolutionary energy (E) of a sequence
seq_to_evaluate = 'A' * sdca.msa_length # arbitrary example: AAAAA...
E_only_alanine = sdca.eval_sequence(seq_to_evaluate, reweight_by_rsa=True)

# Evaluate relative probabilities for the 20 Amino Acids at this position given a background sequence 
fasta_position = 10 # as in FASTA index system (starts at 1)
array_position = fasta_position - 1 # As in a Python array (starts at 0)
amino_acid_probabilities = sdca.position_probabilities(array_position) # P(a) = e^{-dE(wt→a)} / ∑_b e^{-dE(wt→b)}

Tutorials and Advanced Usage

In the ./tutorials/ directory, we provide a series of Jupyter notebooks that illustrate different ways to using StructureDCA:

1. Basics and arguments (1_sdca-basics.ipynb): basics, evaluating effects of mutations with StructureDCA, using optional arguments (like distance_cutoff or lambda_h / lambda_J), evaluate mutations with an alternative background sequence.

2. Access properties (2_sdca-properties.ipynb): access StructureDCA coefficients and properties (like fields h, couplings J, Frobenius norms, residue-residue distance matrix, contact map, ...).

3. Standard DCA (3_sdca-standard-dca.ipynb): solve standard (fully connected) DCA models and run without protein 3D structure.

4. Protein–Protein Interactions (4_sdca-ppis.ipynb): working with protein–protein interactions (PPIs).
   Compute RSA from the biologically relevant conformation, include inter-chain contacts arising from homomers, and build a StructureDCA model from a concatenated MSA of a heteromer PPI of highly coevoling proteins.

5. Custom contacts (5_sdca-custom-contacts.ipynb): build a StructureDCA model with custom contact map (instead of the default distance criteria) and custom weights for StructureDCA[RSA] (instead of default RSA-based weights) to derive any possible sparse DCA model.

Build and Installation Notes

Requirements

• Python 3.9 or later
• Python packages numpy and biopython (version 1.75 or later)
• A C++ compiler that supports C++17 (such as GCC, LLVM or MSVC).

Releases

See the release notes for version history and changes between versions.

Credits

• For inferring the DCA coefficients, StructureDCA uses a gradient descent solver: L-BFGS by Naoaki Okazaki (which is included in this repo).
• The part of the code that makes the bridge between Python and C++ is inspired from the plmDCA implementation 'pycofitness' by Mehari B. Zerihun, Fabrizio Pucci.