pyconsFold: A fast and easy tool for modelling and docking using distance predictions

John Lamb, Arne ElofssonbioRxiv 2021 · DOI

Abstract

Contact predictions within a protein have recently become a viable method for accurate prediction of protein structure. Using predicted distance distributions has been shown in many cases to be superior to only using a binary contact annotation. Using predicted inter-protein distances has also been shown to be able to dock some protein dimers. Here we present pyconsFold. Using CNS as its underlying folding mechanism and predicted contact distance, it outperforms regular contact prediction-based modelling on our dataset of 210 proteins. It performs marginally worse than the state-of-the-art pyRosetta folding pipeline but is on average about 20 times faster per model. More importantly, pyconsFold can also be used as a fold-and-dock protocol by using predicted inter-protein contacts to simultaneously fold and dock two protein chains.

Introduction

Predicting how a protein folds into its three-dimensional shape is a cornerstone of structural biology. Traditionally, researchers relied on "contact maps"—binary grids indicating whether two amino acids are close to each other. However, recent breakthroughs, notably in CASP13, have demonstrated that predicting actual distances between residues provides significantly richer information for constructing accurate models.

Despite these advancements, existing tools for distance-based modeling are often slow, difficult to install, or lack the capability to handle "docking"—the prediction of how separate protein chains fit together. pyconsFold addresses these challenges by providing an open-source, lightweight Python tool that leverages distance-based constraints for both individual protein folding and simultaneous dimer docking.

Methods

pyconsFold is implemented as a Python 3 wrapper around the Crystallography & NMR System (CNS), a well-established engine for molecular structure determination. The tool translates predicted distance distributions and their associated error estimates into geometric constraints that CNS can process.

A key innovation is the "Fold-and-Dock" protocol. To model two protein chains simultaneously, the tool concatenates the Multiple Sequence Alignments (MSAs) of the chains, separated by a "poly-Glycine" (poly-G) linker. This linker acts as a physical buffer, preventing the software from making false proximity predictions between the end of one chain and the start of the next. Models are automatically ranked using CNS internal energy and the pcons quality assessment score.

Results and Discussion

The researchers evaluated pyconsFold using a dataset of 210 proteins, comparing it against binary contact-based modeling and the high-performance pyRosetta pipeline.

Accuracy and Efficiency

pyconsFold distance-based models significantly outperformed traditional contact-based methods. While pyRosetta remains the gold standard for absolute accuracy, pyconsFold achieved comparable results while being approximately 20 times faster per model. This speedup is a critical practical advantage for large-scale research where thousands of models must be generated.

Comparison of pyconsFold performance (Distance TM score) against contact-based modeling and the pyRosetta pipeline.

Simultaneous Docking

The "fold-and-dock" approach simplifies the study of protein-protein interactions by solving the folding and orientation of a complex in a single step. In a test on the 1GPW protein dimer, pyconsFold achieved a TM-score of over 0.8 for each chain, with a total Root Mean Square Deviation (RMSD) of only 3.19 Å, representing a high-quality prediction of the complex.

Simultaneous folding and docking of a protein dimer using the pyconsFold protocol.

Limitations and Future Work

The quality of pyconsFold models is heavily dependent on the quality of the input Multiple Sequence Alignment (MSA). Poor evolutionary data leads to inaccurate distance predictions and, consequently, lower-quality 3D models. Additionally, while the current study focused on protein dimers, future iterations of the "fold-and-dock" protocol could be expanded to handle larger multi-protein assemblies. The tool's design emphasizes minimalism and few dependencies, ensuring ease of installation via pip and long-term maintainability.

Availability: pyconsFold is available as a pip package and on GitHub under the GPLv3 license (https://github.com/johnlamb/pyconsfold).

Summary updated

Abstract

Abstract Motivation Contact predictions within a protein has recently become a viable method for accurate prediction of protein structure. Using predicted distance distributions has been shown in many cases to be superior to only using a binary contact annotation. Using predicted inter-protein distances has also been shown to be able to dock some protein dimers. Results Here we present pyconsFold. Using CNS as its underlying folding mechanism and predicted contact distance it outperforms regular contact prediction based modelling on our dataset of 210 proteins. It performs marginally worse than the state of the art pyRosetta folding pipeline but is on average about 20 times faster per model. More importantly pyconsFold can also be used as a fold-and-dock protocol by using predicted inter-protein contacts to simultaneously fold and dock two protein chains. Availability and implementation pyconsFold is implemented in Python 3 with a strong focus on using as few dependencies as possible for longevity. It is available both as a pip package in Python 3 and as source code on GitHub and is published under the GPLv3 license. Contact [email protected] Supplemental material Install instructions, examples and parameters can be found in the supplemental notes. Availability of data The data underlying this article together with source code are available on github, at https://github.com/johnlamb/pyconsfold .
BibTeX
@misc{lamb2021pyconsfold,
  author = {John Lamb and Arne Elofsson},
  title = {{pyconsFold: A fast and easy tool for modelling and docking using distance predictions}},
  year = {2021},
  doi = {10.1101/2021.02.08.430195},
  howpublished = {bioRxiv},
}
Figure