Limits and potential of combined folding and docking using PconsDock

TL;DR: A fold-and-dock method, PconsDock, based on predicted residue-residue distances with trRosetta, that can simultaneously predict the tertiary and quaternary structure of a protein pair, even when the structures of the monomers are not known.

Abstract: In the last decade, de novo protein structure prediction accuracy for individual proteins has improved significantly by utilizing deep learning (DL) methods for harvesting the co-evolution information from large multiple sequence alignments (MSA). In CASP14, the best method could predict the structure of most proteins with impressive accuracy. The same approach can, in principle, also be used to extract information about evolutionary-based contacts across protein-protein interfaces. However, most of the earlier studies have not used the latest DL methods for inter-chain contact distance predictions. In this paper, we showed for the first time that using one of the best DL-based residue-residue contact prediction methods (trRosetta), it is possible to simultaneously predict both the tertiary and quaternary structures of some protein pairs, even when the structures of the monomers are not known. Straightforward application of this method to a standard dataset for protein-protein docking yielded limited success, however, using alternative methods for MSA generating allowed us to dock accurately significantly more proteins. We also introduced a novel scoring function, PconsDock, that accurately separates 98% of correctly and incorrectly folded and docked proteins and thus this function can be used to evaluate the quality of the resulting docking models. The average performance of the method is comparable to the use of traditional, template-based or ab initio shape-complementarity-only docking methods, however, no a priori structural information for the individual proteins is needed. Moreover, the results of traditional and fold-and-dock approaches are complementary and thus a combined docking pipeline should increase overall docking success significantly. The dock-and-fold pipeline helped us to generate the best model for one of the CASP14 oligomeric targets, H1065.

Summary

Summary

• Protein structure is crucial for their understanding of biological function.
• At a depth of 100 sequences, the average TM-score is over 0.6, indicating that about 100 effective sequences are in most cases sufficient to obtain the fold of a protein.
• The default (N3) performance is compared withpyconsFold (uses the pyconsFold program instead of Rosetta), RaptorX (uses inter-chain contacts predicted by RaptorX instead of distances from trRosetta), RaptorX and N3-pdb use the intra-chain distances from the native structures, and N3-merged uses intra-chain distances predicted by the full alignments for each chain independently.
• First, it can be seen that the successful dockings tend to have a multiple sequence alignment of one hundred or more residues, see Figure 5A.
• There are a few targets whose performance increases significantly.
• First, the authors compared it to one shape complementarity method, Gramm, and one template-based docking method, TMdock (see Figure 9.
• In some cases, only specific alignment gives correct folding and docking based on the intrinsic evolutionary characteristic of the proteins and their interaction.
• Here, it should be noted that a dockQ score over 0.23 roughly corresponds to an “acceptable” model in CAPRI [45], and the authors will therefore call all models with dockQ >0.23 as correct and all others as incorrect.
• The distances were then used in Rosetta as described in the original trRosetta protocol.
• Morcos F, Pagnini A, Lunt B, Bertolino A, Marks D, Sander C, et al. Estimation of Residue-Residue Coevolution using Direct Coupling Analysis Identifies Many Native Contacts Across a Large Number of Domain Families.

Figures

Figure 3: Results of PconsDock using different alignments. A) dockQ scores for all models using six different alignments (see Table 1). B) The dockQ scores for the 15 proteins where at

Figure 8. A) Prediction qualities for all models (using the methods to produce an alignment marked with a star in Table 1). B) Predictions qualities for all models using different alignments and sequence databases (see Table 1 for details). C) Comparison of dockQ scores for individual models compared to the N3 models.

Table 1: Overview of alignment methods used in this study (top rows) as input to the fold and dock protocol or alternative docking methods and their performance (Number of correct docked proteins, average dockQ score, and average TM-score.

Table 2: Overview of consensus methods used to rank models.

Figure 6. Distance maps for predicted (lower left) and native (upper right) distances for three proteins with artefacts 3qlu (A,B), 3pv6 (C,D) and 4yoc(E,F). Left columns (A,C,E) with N3 alignments, right (B,D,F) with reciprocal best hits.

Full text

Limits and potential of combined folding
and docking using PconsDock.
Gabriele Pozzati
1*
, Wensi Zhu
1*
, Claudio Bassot
1*
, John Lamb
1
, Petras
Kundrotas
1,2
, Arne Elofsson
1
1
Science for Life Laboratory and Dep of Biochemistry and Biophysics,
Stockholm University, Box 1031, 171 21 Solna, Sweden
2
Center for Computational Biology, The University of Kansas,
Lawrence, KS 66047, USA
*
=contributed equally.
Abstract
In the last decade, de novo protein structure prediction accuracy for individual proteins has
improved significantly by utilising deep learning (DL) methods for harvesting the co-evolution
information from large multiple sequence alignments (MSA). In CASP14, the best groups
predicted the structure of most proteins with impressive accuracy. The same approach can, in
principle, also be used to extract information about evolutionary-based contacts across
protein-protein interfaces. However, most of the earlier studies have not used the latest DL
methods for inter-chain contact distance prediction. This paper introduces a fold-and-dock
method, PconsDock, based on predicted residue-residue distances with trRosetta. PconsDock
can simultaneously predict the tertiary and quaternary structure of a protein pair, even when the
structures of the monomers are not known. The straightforward application of this method to a
standard dataset for protein-protein docking yielded limited success. However, using alternative
methods for MSA generating allowed us to dock accurately significantly more proteins. We also
introduced a novel scoring function, PconsDock, that accurately separates 98% of correctly and
incorrectly folded and docked proteins. The average performance of the method is comparable
to the use of traditional, template-based or ab initio shape-complementarity-only docking
methods. However, no a priori structural information for the individual proteins is needed.
Moreover, the results of conventional and fold-and-dock approaches are complementary, and
thus a combined docking pipeline could increase overall docking success significantly.
PconsDocck contributed to the best model for one of the CASP14 oligomeric targets, H1065.
.CC-BY 4.0 International licenseavailable under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (whichthis version posted June 7, 2021. ; https://doi.org/10.1101/2021.06.04.446442doi: bioRxiv preprint

Introduction
Protein structure is crucial for our understanding of biological function. However, experimentally
determining the structure of a protein is still time-consuming and expensive. Therefore,
computational methods will be the only method to determine the structure of most proteins in
the foreseeable future. Until recently, the only method to reliably predict the structure of a
protein was to model it using a homologous template. However, reliable templates are not
available for close to half the residues in the human proteome [1].
For several decades the prediction of protein structure directly from sequence information has
been an unachievable dream. However, that changed about a decade ago when improved
methods using co-evolution achieved sufficient residue contact information to predict the
structure of many proteins [2,3]. Later, deep learning [4,5] and prediction of residue-residue
distances provided further improvements [6,7]. Today this means that for many, if not most,
individual proteins, it is possible to accurately predict the structure of its folded domains [8].
Recently, Deepmind demonstrated at CASP14 that using an end-to-end learnable approach,
high-quality prediction of almost all protein domains is already feasible today (although not
generally available).
In principle, the same type of methods used for predicting the structure of a single protein can
predict the interaction between two proteins [9,10]. However, there is one fundamental
difference: it is necessary to create paired alignments to identify the interaction between two
proteins, i.e. identifying what pairs of proteins interact in the same manner. The identification of
interacting pairs is assumed to be relatively easy for pairs of proteins that both only contain a
single homolog in a set of genomes, but when multiple paralogs exist - the exact pairing is
difficult [11].
Proteins do, however, not act alone. They function by interacting with other proteins and other
molecules. Protein interaction can vary in nature from stable interaction present in small and
large protein complexes to transient interactions often used for regulation. Experimentally the
study of stable protein interactions can be done using various techniques. Structural
determination methods, including crystallography and Cryo-EM electron microscopy, can solve
the structure of protein complexes, while other methods can be used to identify that two proteins
interact without obtaining detailed structural information.
Prediction of protein interactions has been an even more significant challenge than predicting
the structure of individual proteins. Many different techniques have been developed, but in short,
they can be divided into four categories: (i) docking primarily based on shape complementarity
[12], (ii) template-based modelling [13], and (iii) flexible docking [14,15]. Various energy
functions have also been used to improve the identification of correct docking poses [16]. In
addition, co-evolution-based methods have also been used to predict the structure of complexes
[9,17].
.CC-BY 4.0 International licenseavailable under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (whichthis version posted June 7, 2021. ; https://doi.org/10.1101/2021.06.04.446442doi: bioRxiv preprint

Benchmarks have been developed to elucidate the advantages and disadvantages of different
docking methods [18]. Shape complementarity works excellently on native complexes, but the
accuracy drops fast when using the structures of unbound complexes and even further if models
of the proteins are used [19,20]. Template-based modelling works excellently if a complex with
significant sequence identity exists in PDB but does not work for novel complexes[21,22].
Successful DCA based methods to predict protein-protein interactions preceded the large-scale
prediction of single proteins by predicting the bacterial two-component signalling in 2009 [17].
These methods were then extended to a handful of other complexes by several groups [9,10].
However, it is still unclear how generally applicable these methods are, but the potential to
vastly increase the space of known protein-protein interactions should lie in using some type of
co-evolution based methods. The computational cost limits flexible docking, but a fold-and-dock
protocol [23] based on coevolution does not require an exact structure of the two individual
proteins.
In addition to determining the structure of a protein complex, it is also crucial to determine which
proteins interact. However, protein-protein interaction is not an easily defined entity. It might
include anything from proteins regulating the expression of genes to proteins strongly bound to
each other in a large molecular machine. Several interaction databases exist [24,25], and
methods, including co-evolution based methods [26], to predict interactions have been
developed.
Here, we examine if it is possible to simultaneously fold and dock [23] two proteins by using
coevolutionary information and not only dock them. In addition, we use one of the best methods
(trRosetta) instead of DCA[2] based methods to predict intra- and inter-chain distances. One
advantage of a fold-and-dock methodology is that it is not dependent on the availability of
individual structures and should therefore be less sensitive to structural rearrangements upon
binding. The disadvantage is that obviously, there are many more degrees of freedom in the
system. We find that for several cases, it is possible to fold and dock the dimer simultaneously
accurately. Although the success rate is low (<10%), this is comparable to the accuracy of other
docking methods, which utilises the structure of both individual proteins. In addition, the
methods are complementary.
Results
The protocol used here starts from two multiple sequence alignments, created by searching with
jackhmmer [27] against all complete proteomes from UniProt [28]. After that, a combined
multiple sequence alignment is created by including the top paired hit from each proteome. It
should be noted that the depth of the combined multiple sequence alignment is often
significantly smaller than for the individual proteins. In addition, a few alternative methods both
for generating the alignments and selecting the sequences were tried. These are discussed
below. Next, twenty Glycine residues were added to separate the two sequences in the
.CC-BY 4.0 International licenseavailable under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (whichthis version posted June 7, 2021. ; https://doi.org/10.1101/2021.06.04.446442doi: bioRxiv preprint

combined multiple sequence alignment. The combined alignment can be created in two different
orientations, A-B vs B-A, and we have tried to use both combinations.
Next, the combined multiple sequence alignment has been adopted to predict distances and
angles with trRosetta [29]. These are then used as input to Rosetta or CNS [30] to fold and dock
the two proteins.
Below, we will discuss when this methodology works, when it fails, compare the performance of
different alignments, and compare the performance with other docking techniques, and finally
introduce a score, PconsDock, which accurately can be used to distinguish successful and
unsuccessful docking attempts.
Example of successful fold and dock.
Figure 1: A) Predicted (lower triangle) and actual (upper triangle) distance map of the
protein 4gmj. The two blue stripes represent the poly-G linker between the two chains.
The title shows that 287 interchain contacts are predicted and that 48.4% of these are
correct. B) Real (dark colours) and modelled (light colours) structure of the protein
1vrs. The accuracy of the models is good, dockQ score 0.42, and the TM-scores for
the two chains are 0.82 and 0.85, respectively.
First, we demonstrate that the algorithm can accurately fold and dock a pair of proteins in at
least one case. Figure 1 presents one successful example of the fold-and-dock protocol for the
human protein complex between NOT1 MIF4G and CAF1 (PDB: 4gmj)[31]. The prediction is
built on an alignment containing 1189 sequences (Meff=523) created by three iterations of
jackhmmer[27] and an E-value cutoff of 10
-3
against all reference proteomes in UniProt[28].
Visually, it can be seen that the intra-chain distance maps are similar and most intra-chains
.CC-BY 4.0 International licenseavailable under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (whichthis version posted June 7, 2021. ; https://doi.org/10.1101/2021.06.04.446442doi: bioRxiv preprint

contacts are predicted accurately (PPV>0.90 for both chains), resulting in well-folded models of
both chains (TM-score >0.8 for both). In total, 139 out of 287 inter-chain contacts are accurately
predicted (287 contacts predicted with a PPV of 49%). The final docked model is also accurate
(dockQ score 0.42). However, as we will show below, unfortunately, many models are not as
easy to model as 4gmj. To test the performance of the algorithm, we have, therefore, used 222
heterodimeric protein pairs from dockground 4.3 [18,28].
Modelling accuracy depends on the size of the MSA; docking
performance does not.
Figure 2: Performance of the fold-and-dock methodology versus the size of the joint
alignments. Average TM-score of the two chains (A) and dockQ scores (B) plotted
against the size of the multiple sequence alignment used to predict the contacts.
The Dockground heterodimeric dataset was used to test the performance of the fold-and-dock
methodology. First, we examined the dependence of the size of the multiple sequence
alignment on the performance. It can be seen that the average TM-score for both chains is
increasing with the size of the combined alignment, Figure 1. At a depth of 100 sequences, the
average TM-score is over 0.6, indicating that about 100 effective sequences are in most cases
sufficient to obtain the fold of a protein.
Next, we examined the quality of the predicted dimers, Figure 1B. A few models are docked
correctly (dockQ score >0.23). However, most protein pairs are not accurately docked (dockQ
.CC-BY 4.0 International licenseavailable under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (whichthis version posted June 7, 2021. ; https://doi.org/10.1101/2021.06.04.446442doi: bioRxiv preprint

Frequently asked questions

Q1. What contributions have the authors mentioned in the paper "Limits and potential of combined folding and docking using pconsdock" ?

This paper introduces a fold-and-dock method, PconsDock, based on predicted residue-residue distances with trRosetta. The authors also introduced a novel scoring function, PconsDock, that accurately separates 98 % of correctly and incorrectly folded and docked proteins. CC-BY 4. 0 International license available under a was not certified by peer review ) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.