Deep learning-driven insights into super protein complexes for outer membrane protein biogenesis in bacteria
Curation statements for this article:-
Curated by eLife
eLife assessment
In this valuable contribution, the authors apply an artificial intelligence method to predict the three-dimensional structure of complexes of outer membrane proteins of the Gram-negative bacterium E. coli. Some of the cases presented are compelling, as they explain previously published biochemical data and/or reproduce existing structural data.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Molecular Biology (eLife)
Abstract
To reach their final destinations, outer membrane proteins (OMPs) of gram-negative bacteria undertake an eventful journey beginning in the cytosol. Multiple molecular machines, chaperones, proteases, and other enzymes facilitate the translocation and assembly of OMPs. These helpers usually associate, often transiently, forming large protein assemblies. They are not well understood due to experimental challenges in capturing and characterizing protein-protein interactions (PPIs), especially transient ones. Using AF2Complex, we introduce a high-throughput, deep learning pipeline to identify PPIs within the Escherichia coli cell envelope and apply it to several proteins from an OMP biogenesis pathway. Among the top confident hits obtained from screening ~1500 envelope proteins, we find not only expected interactions but also unexpected ones with profound implications. Subsequently, we predict atomic structures for these protein complexes. These structures, typically of high confidence, explain experimental observations and lead to mechanistic hypotheses for how a chaperone assists a nascent, precursor OMP emerging from a translocon, how another chaperone prevents it from aggregating and docks to a β-barrel assembly port, and how a protease performs quality control. This work presents a general strategy for investigating biological pathways by using structural insights gained from deep learning-based predictions.
Article activity feed
-
-
Author Response
Reviewer #1 (Public Review):
“Even though the methodology was already introduced, it should be described in some detail. Most importantly, AlphAfold's measures of accuracy have been part of the loss function during training/testing. What about the measure of protein-protein interaction accuracy? Was it also in the loss function?”
We thank the reviewer for this insightful comment. The metrics used for evaluating predicted structure quality, such as the predicted local distance difference test (pLDDT) score and predicted TM score (pTM), both proposed in the AlphaFold 2 publication (Ref. 27), and the interface score (iScore) proposed in the AF2Complex publication (Ref. 23), are not explicitly employed as the loss function in training the main deep learning model for structure prediction. Instead, the main loss function …
Author Response
Reviewer #1 (Public Review):
“Even though the methodology was already introduced, it should be described in some detail. Most importantly, AlphAfold's measures of accuracy have been part of the loss function during training/testing. What about the measure of protein-protein interaction accuracy? Was it also in the loss function?”
We thank the reviewer for this insightful comment. The metrics used for evaluating predicted structure quality, such as the predicted local distance difference test (pLDDT) score and predicted TM score (pTM), both proposed in the AlphaFold 2 publication (Ref. 27), and the interface score (iScore) proposed in the AF2Complex publication (Ref. 23), are not explicitly employed as the loss function in training the main deep learning model for structure prediction. Instead, the main loss function of AF2 is the Frame Aligned Point Error (FAPE) loss, which measures the errors in the predicted atomic coordinates within local coordinate frames spanned by vectors connecting backbone heavy atoms of individual protein residues. However, this FAPE loss function is very much relevant to predicting TM-scores or iScores; both are derived from an additional module that predicts alignment errors (PAEs) viewed from each residue’s local frame. The training of this PAE module was done separately as described in the AF2 publication (Ref. 27). According to DeepMind, the training of the deep learning models for AlphaFold-Multimer (Ref. 25, AF version 2.2.0 and above) has relatively minor changes in the loss function; changes were made mainly to reduce severe clashes, which were not uncommon in modeling large complexes by earlier versions of AF2.
We added in the Methods section, line 337,
“The iScore metric was derived from the predicted alignment errors that gives an estimated distance for interface residue j from its position in the experimental structure, as viewed from a local frame of residue interface residue i [23,27]. To better estimate confidence, the contribution of each interface residue to the interface score is calculated using local frames not located within the same protein chain, i.e., residue i and j belonging to different chains.”
“Figure 1a (upper panel, PpiD) includes quite a few promising hits but only the first, third, and 12th were considered. How were these chosen? For example, why not consider the second? The lower panel (YfgM) also shows many promising hits but only the first was chosen. Why not more? Likewise, only two of the top hits in Figure 4 are considered. What about the rest? For example, why taking into account the second best hit while skipping the first?”
These are important questions about similar issues raised by all three reviewers, i.e., R2.1 by reviewer 2 and R3.2 by reviewer 3. We emphasize that our approach predicts physical interactions between proteins, not the biological consequence of such interactions. However, since the most interesting predictions are the ones relevant to biological functions, about which the computational method cannot make a judgement, given the space limitations of the manuscript, we opted to select from the top predictions those that likely provide mechanistic insights into biological function, for example, those that might inspire new hypotheses about molecular mechanisms. In practice, our selection process was guided by existing literature and experimental evidence. Since such information is limited, we can only focus on the very few ones with both strong computational and experimental evidence. Most top predictions, including the ones the reviewers questioned, were not pursued further because we cannot at present say anything about the functional consequences of these predicted interactions, even though they may interact physically. One main contribution of this computational screening approach is to provide short lists that accelerate the search for functionally important protein-protein interactions. Thus, in this contribution, we provide some examples found in the top 20 hits ranked from ~1500 possible pairs for a given query protein.
In this revision, we added from line 85,
“Note that our computational predictions are about physical interactions between a pair of proteins subjected to screening, not about their biological roles even if they are predicted to interact physically. Moreover, the predicted physical interactions may not be relevant in the cellular environment due to various factors not considered in modeling, e.g., competition from other proteins with stronger binding affinities, post-translational modifications, etc. Thus, it is possible that many protein-protein interactions predicted by this pipeline do not necessarily have biological relevance. Nevertheless, since cognate protein-protein interactions required by their functions are more likely to be detected than randomly selected proteins, biologically interesting protein-protein interactions are enriched at the top of the screening results ranked by iScore. Thus, the screening procedure may provide valuable even critical clues for subsequent investigation. In this study, assisted by existing experimental evidence, we select from high confidence computational predictions those most likely to have significant biological implications, and then predict the structures of larger complexes if more than two proteins are involved according to our predictions or based on literature information. The interactions that we ignored are either of unknown biological significance, physically interacting but biologically irrelevant, or simply false positives.”
“Authors argue that the unstructured part of OmpA, which wraps around SurA, is to be trusted, which may be the case. But a more likely explanation is that it is an artefact, in agreement with the very low confidence assigned by AlphaFold.”
While we do not disagree that the structure prediction about SurA/OmpA complex may contain artifacts, there are several reasons why our predications may be insightful, as we explained in the manuscript. First, it is well-known in experimental studies (references 41, 42, 45) that the SurA/OmpA complex is very dynamic and unlikely to possess a stable structural complex as in a typical crystal structure. As such, the low confidence score by AF2Complex is expected, as it reflects uncertainty due to the existence of many possible conformations. Second, it makes physical sense to have loose wrapping of OmpA around SurA, as it reduces the energetic costs to dissociate OmpA from SurA when SurA approaches BAM for its delivery. Our point is a qualitative assessment, rather than claiming a specific complex model as in a typical structure prediction scenario. To be cautious as the reviewer suggested, we added a sentence in the Discussion, from line 309,
“Despite the low confidence due to weak interactions, the predicted structures delineate a picture for how SurA prevents OmpA from aggregating. Moreover, since it transports OmpA with a relatively small number of intermolecular contacts, the free energy required to dissociate OmpA from SurA is small. Notwithstanding these considerations, we caution that artifacts likely exist in these predicted structural models.”
“Figure 5. How is (does) this predicted structure compare with the known structure of the complex? In particular, how similar are the predicted and known structures of the individual subunits, and how similar are the predicted docking poses to the known ones?”
The BAM complex has been studied extensively, with over one hundred experimental structures of its individual subunits or the full complex. Therefore, a thorough structural comparison is a subject of a review beyond the scope of this study. In our computational models, the structures of the individual subunits or of the full BAM complex closely mimic their known experimental structures, which is expected because some of these structures were likely employed in the training of deep learning models and/or structure predictions. We added a comparison to the highest resolution crystal structure in the revised manuscript after line 225,
“Because BAM has been extensively studied structurally [7,47], we focus on describing its interaction with SurA, though the predicted BAM complex model closely mimics a known crystal structure of the complex determined at 2.9 Å resolution (PDB 5D0O, [48]). The alignment of the two complex structures yields a very high TM-score of 0.94.”
“Authors should make the results easily accessible to all. Maybe as Cytoscape and CyToStruct sessions for easy visualization.”
Cytoscape and the add-on CytoStruct are very useful tools to visualize large networks. In our case, however, we are presenting only a handful of complexes, not a massive protein-protein interaction network like those resulting from all-against-all screening at proteome-scale. A diagram such as Fig. 7 is sufficient for our visualization purposes. Moreover, we provide the atomic coordinates in the standard PDB format for readers who wish to examine the respective structures in detail. In the future, if we have opportunity to expand PPI screening to a large number of targets, Cytoscape and add-ons will be handy to display the resulting gigantic network.
“Finally, AlphaFold was trained and tested mostly with water-soluble protein. Thus, application to outer membrane proteins is a bit risky. Maybe authors can comment on this.”
While it is true that most experimental structures used for training AlphaFold models are of water-soluble proteins, there are also structures of many membrane proteins available for training, as over 10,000 structures of membrane-proteins were already deposited in the Protein Data Bank, though there are redundancy within these structures and some domains are outside the transmembrane regions. These structures are likely sufficient for machine-learning approaches such as AlphaFold 2 to learn the sequence and structural patterns unique to transmembrane proteins. This view is supported by our empirical experience, because transmembrane regions of membrane proteins are typically among those with high confidence scores, e.g., complex models for a transmembrane molecular system CcmI presented in our AF2Complex work (Ref. 23). And one of these computational models (of CcmA2B2CD) was just confirmed to have high quality by cyro-EM models (Li et.al., Nature Communications 13:6422, 2022) at TM-score 0.89. We note that this was a non-trivial prediction as this structure was not present in the PDB and was long sought by the experimentalists. The view also agrees with the conclusion of a recent published study on AF2 models of transmembrane proteins (Hegedűs, et. al. Cell. Mol. Life Sci. 79:73, 2022).
-
eLife assessment
In this valuable contribution, the authors apply an artificial intelligence method to predict the three-dimensional structure of complexes of outer membrane proteins of the Gram-negative bacterium E. coli. Some of the cases presented are compelling, as they explain previously published biochemical data and/or reproduce existing structural data.
-
Reviewer #1 (Public Review):
Earlier this year Skolnick and colleagues managed to tweak AlphaFold to predict protein complexes (reference 23 in the current manuscript). They also added a score that allows the detection of true protein-protein interactions among arbitrary protein pairs. Thus, their methodology allows reliable prediction of homo- and hetero-meric protein-protein interactions, and predicting the structure of the corresponding protein complexes. Leveraging this methodology, the current manuscript describes a very interesting application to a set of about 1,500 E. coli proteins of the outer membrane, the periplasm and the inner membrane of this Gram negative bacteria. They explore protein-protein interactions among this protein set, which they refer to as 'envelome'. Their results reproduce known protein complexes, such as …
Reviewer #1 (Public Review):
Earlier this year Skolnick and colleagues managed to tweak AlphaFold to predict protein complexes (reference 23 in the current manuscript). They also added a score that allows the detection of true protein-protein interactions among arbitrary protein pairs. Thus, their methodology allows reliable prediction of homo- and hetero-meric protein-protein interactions, and predicting the structure of the corresponding protein complexes. Leveraging this methodology, the current manuscript describes a very interesting application to a set of about 1,500 E. coli proteins of the outer membrane, the periplasm and the inner membrane of this Gram negative bacteria. They explore protein-protein interactions among this protein set, which they refer to as 'envelome'. Their results reproduce known protein complexes, such as the translocon, and suggest many yet unknown interactions that make biological sense.
A main strength here is the generation of ample hypotheses to be tested in experiment, i.e., all protein-protein interactions of high predicted accuracy. Another strength is that the methodology is readily applicable to other systems. However, a few outstanding issues need to be clarified.
1. Even though the methodology was already introduced, it should be described in some detail. Most importantly, AlphAfold's measures of accuracy have been part of the loss function during training/testing. What about the measure of protein-protein interaction accuracy? Was it also in the loss function?
2. Figure 1a (upper panel, PpiD) includes quite a few promising hits but only the first, third, and 12th were considered. How were these chosen? For example, why not consider the second? The lower panel (YfgM) also shows many promising hits but only the first was chosen. Why not more?
3. Likewise, only two of the top hits in Figure 4 are considered. What about the rest? For example, why taking into account the second best hit while skipping the first?
4. Authors argue that the unstructured part of OmpA, which wraps around SurA, is to be trusted, which may be the case. But a more likely explanation is that it is an artefact, in agreement with the very low confidence assigned by AlphaFold.
5. Figure 5. How is this predicted structure compare with the known structure of the complex? In particular, how similar are the predicted and known structures of the individual subunits, and how similar are the predicted docking poses to the known ones?
6. Authors should make the results easily accessible to all. Maybe as Cytoscape and CyToStruct sessions for easy visualization.
7. Finally, AlphaFold was trained and tested mostly with water-soluble protein. Thus, application to outer membrane proteins is a bit risky. Maybe authors can comment on this. -
Reviewer #2 (Public Review):
It is known that bacterial outer membrane proteins must interact with a variety of cellular factors to reach their final destination safely. There is considerable biochemical evidence in the literature (primarily from crosslinking studies) that these factors interact to promote the movement of client proteins and to prevent their aggregation or misfolding, but the details of the interactions are unknown. The authors showed that they could use a novel virtual screening method together with known crystal structures of individual factors to predict the three-dimensional structures of several pairs or groups of interacting factors (supercomplexes). The predicted supercomplex structures are both fascinating and compelling because they are consistent with the published results and they help to explain the …
Reviewer #2 (Public Review):
It is known that bacterial outer membrane proteins must interact with a variety of cellular factors to reach their final destination safely. There is considerable biochemical evidence in the literature (primarily from crosslinking studies) that these factors interact to promote the movement of client proteins and to prevent their aggregation or misfolding, but the details of the interactions are unknown. The authors showed that they could use a novel virtual screening method together with known crystal structures of individual factors to predict the three-dimensional structures of several pairs or groups of interacting factors (supercomplexes). The predicted supercomplex structures are both fascinating and compelling because they are consistent with the published results and they help to explain the mechanism by which the cellular factors promote outer membrane protein biogenesis. I think that this study will be of interest to a wide audience because it serves as a proof-of-concept that although Alpha Fold is incredibly useful for predicting the structures of protein monomers, more sophisticated applications can be used to successfully predict the structures of protein complexes which are often the workhorses of the cell. I have only two significant concerns. First, the authors focused on high confidence supercomplexes that have known biological significance. Their method also identified other high confidence supercomplexes, but they need to explain how they can distinguish predicted supercomplexes that have potential biological significance from those that are simply "false positives". Second, one of the proposed functional models does not seem to be consistent with the results of a previous study.
-
Reviewer #3 (Public Review):
In this paper, the authors apply AlphaFold2 to predict the structure of membrane protein complexes in E.Coli. They scan ~1500 membrane proteins starting with one protein to predict the interactions. They present the results for four proteins and analyse them carefully to propose novel models for complexes.
The main problem with the manuscript is that the authors first claim that the method is highly specific but then cherry-pick a subset of interactions that they believe are correct (most likely they are). But the authors do not discuss the other high-scoring predictions. Are these false positives (in which case the method has very limited value) or novel interactions (which would be really interesting but needs further examination)?
-
