1. Computational and Systems Biology
  2. Structural Biology and Molecular Biophysics

Deep learning-driven insights into super protein complexes for outer membrane protein biogenesis in bacteria

  1. Mu Gao  Is a corresponding author
  2. Davi Nakajima An
  3. Jeffrey Skolnick  Is a corresponding author
  1. Center for the Study of Systems Biology, School of Biological Sciences, Georgia Institute of Technology, United States
  2. School of Computer Science, Georgia Institute of Technology, United States
https://doi.org/10.7554/eLife.82885
Share this article
Cite this article
  1. Mu Gao
  2. Davi Nakajima An
  3. Jeffrey Skolnick
(2022)
Deep learning-driven insights into super protein complexes for outer membrane protein biogenesis in bacteria
eLife 11:e82885.
https://doi.org/10.7554/eLife.82885
  2. Download BibTeX
  3. Download .RIS

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.

Editor's evaluation

The authors show that an artificial-intelligence method can be used to predict the three-dimensional structure of protein-protein complexes formed between cellular factors that promote the assembly of bacterial outer membrane proteins. The structures are compelling because they explain previously published biochemical data and provide novel insights into the function of these factors.

https://doi.org/10.7554/eLife.82885.sa0

eLife digest

All living cells are contained within a fatty cell membrane that allows water and only certain other molecules to pass through with ease. Bacteria only consist of a single cell, making their membrane the only interface with the surrounding environment. Gram-negative bacteria – which include Escherichia coli, a bacterium found in the gut of all humans – have an extra layer of protection, the ‘outer membrane’. Proteins in this membrane are called ‘outer membrane proteins’ or OMPs and allow nutrients to enter the cell. But OMPs, which are made inside the cell, need to be transported to the outer membrane and folded correctly before they can perform their role. This multistep process, which involves interactions between many different proteins, is not fully understood.

The journey of an OMP from the center of the cell where it is made to the outer membrane is complicated. First, the OMP needs to pass through the cell’s inner membrane. To do this, it must interact with ‘channel proteins’ in the inner membrane that feed the OMP into the space between the two membranes, known as the bacterial envelope. This step requires the OMP to be unfolded. Once in the bacterial envelope the OMP interacts with proteins that help it fold correctly and integrate into the outer membrane.

The interactions between proteins in the bacterial envelope are short-lived, making them difficult to study using lab-based experiments. An alternative approach is predicting a protein’s structure from its amino acid sequence which is a difficult computational problem to solve. However, in 2020 developers behind the AlphaFold2, a deep learning program, were able to use a set of equations organized in a ‘neural network’ that can ‘learn’ from a library of known protein structures to predict unknown structures with high accuracy. Gao et al. used AF2Complex, a tool based AlphaFold2, tailored to predicting interactions between proteins, to investigate what interactions OMPs could be involved with on their way to the outer membrane.

With the help of a supercomputer at the Oakridge National Laboratory, Gao et al. screened nearly 1,500 E. coli proteins within the bacterial envelope to see how they might interact with OMPs. The screen identified previously unknown interactions between proteins that suggest that the formation of the bacterial outer membrane and the integration of proteins into it involve protein complexes and molecular mechanisms that have not yet been characterized. Additionally, the screen also identified interactions that had been previously described, confirming that the deep learning approach can correctly capture real interactions.

Overall, Gao et al.’s work inspires new hypotheses about the mechanisms through which OMPs are transported to the outer membrane, although further work will be needed to confirm the roles of protein interactions predicted by the computational model experimentally. Furthermore, the ability to design experiments based on computational predictions is exciting. If confirmed, the new protein interactions could help scientists better understand OMP transport, which is essential for bacterial biology. In the future, this could lead to the discovery of new targets for antibiotic drugs.

Introduction

A structural component unique to gram-negative bacteria is the outer membrane (OM), composed of lipopolysaccharides and phospholipids in an asymmetric bilayer with embedded lipoproteins and transmembrane β-barrel proteins (Silhavy et al., 2010). The latter group, OM proteins (OMPs), play vital functional roles, e.g., exchanging small molecules with the environment through their transmembrane β-barrel porins. OMPs are synthesized by cytosolic ribosomes, translocated across the inner membrane (IM), and finally delivered to the OM via the OMP biogenesis pathway (Rollauer et al., 2015).

The translocation and folding of OMPs involve many proteins, including essential and auxiliary ones that form multiple complexes in cooperation (De Geyter et al., 2016; Troman and Collinson, 2021). Two core complexes are the SecYEG translocon (Rapoport et al., 2017; Oswald et al., 2021) and the β-barrel assembly machine (BAM; Tomasek and Kahne, 2021; Noinaj et al., 2017). SecYEG, a hetero-trimer composed of the secretion channel SecY and two additional subunits SecE and SecG, is anchored to the IM and is responsible for moving most cell envelope proteins across the IM (Rapoport et al., 2017). Periplasmic and OM proteins enter the channel in SecY via the SecA-dependent translocation pathway (Oswald et al., 2021). They possess signal peptides recognized by SecA, which inserts a protein substrate into SecY and powers it through the channel using energy from ATP hydrolysis (Collinson, 2019). The signal peptide, located at the N-terminus of a substrate, contains a hydrophobic segment that folds into a transmembrane α-helix once it exits the lateral gate of SecY. To release the substrate from the IM, a type I signal peptidase (SPase I) cleaves the signal peptide (Paetzel, 2014). At this point, a periplasmic protein has reached its destination, but an OMP, escorted by the chaperones SurA or Skp, continues its journey toward BAM harbored in the OM (Sklar et al., 2007). Five subunits constitute BAM, of which BamA plays the major role in folding a β-barrel and releasing the matured product (Tomasek and Kahne, 2021; Noinaj et al., 2017).

There are many open questions concerning the OMP biogenesis pathway. In Escherichia coli, the SecYEG translocon recruits additional proteins, such as SecA (Li et al., 2016), SecDF (Tsukazaki, 2018), YidC (Kumazaki et al., 2014), and PpiD (Antonoaea et al., 2008; Sachelaru et al., 2014), to form a variety of supercomplexes in different scenarios. We do not know the identities of all members of these supercomplexes, let alone their atomic structures. SurA plays a major role in chaperoning a nascent OMP (Sklar et al., 2007; Behrens-Kneip, 2010). How does it handle and deliver a substrate to BAM? Likewise, BAM recruits additional helpers, e.g., BepA (Narita et al., 2013), but how do they work together?

To answer these questions experimentally is challenging (Babu et al., 2018; Maddalo et al., 2011; Alvira et al., 2020; Carlson et al., 2019). Recently, deep learning approaches have made tremendous progress in predicting the structures of protein complexes (Gao et al., 2022b; Humphreys et al., 2021; Evans et al., 2021; Bryant et al., 2022). Here, we use one such method to address the above questions. The centerpiece of our approach is AF2Complex (Gao et al., 2022b), built on AlphaFold2 (AF2) (Evans et al., 2021; Jumper et al., 2021). Using AF2Complex, we combine virtual screening for protein-protein interactions (PPIs) and supercomplex modeling and apply this strategy to several important proteins in the OMP biogenesis pathway.

Results

Virtual screening for PPIs in the E. coli envelopome

A workflow employing AF2Complex was implemented to search for interacting proteins within the E. coli cell envelope. We refer to these proteins collectively as the ‘envelopome,’ which consists of ~1450 proteins, or ~35% of the complete E. coli proteome. Given a query envelope protein sequence, screening is conducted against every envelope protein, including the query itself to test if it forms a homo-oligomer. For each pair of input proteins, AF2Complex predicts their structures simultaneously as a putative dimer and evaluates the interface score (iScore) of the resulting structural models to decide if they interact. The value of iScore ranges between 0 and 1. Interacting protein pairs typically return a positive iScore, whereas non-interacting proteins usually return 0 or low iScores. Typically, 20 final structures are produced in multiple independent runs using 10 different AF2 deep learning models. The highest iScore among all predicted structures of a target is used for ranking the entire envelopome.

According to previous benchmarks on a set of ~7000 putatively non-interacting protein pairs from E. coli, minimum iScore thresholds of 0.40, 0.50, and 0.70 yield false positive rates of 1.2, 0.4, and <0.01%, respectively (Gao et al., 2022b). With respect to the capability of identifying and modeling true PPIs, on a set of 440 heterodimeric complexes whose experimental structures have been recently determined and were not used for AF2 model training, AF2Complex recalls 81, 74, and 34% of these benchmark targets at the same three iScore thresholds, respectively, and yields medium- or high-quality complex structures for 84, 87, and 93% in the top ranked models of these positively identified targets (Gao et al., 2022b). Hence, we consider predictions of medium, high, and very high confidence, progressively, at iScore cutoffs of 0.4, 0.5, and 0.7. Given a query protein, the top iScores of all scanned envelope proteins are then sorted to identify potential interacting partners with the query protein. 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 a 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 PPIs predicted by this pipeline do not necessarily have biological relevance. Nevertheless, since cognate PPIs required by their functions are more likely to be detected than randomly selected proteins, biologically interesting PPIs 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.

As a first test, we applied our PPI pipeline to the chaperones PpiD and YfgM because strong experimental evidence indicates that these two proteins belong to a super SecYEG translocon (Antonoaea et al., 2008; Sachelaru et al., 2014; Maddalo et al., 2011; Dartigalongue and Raina, 1998; Fürst et al., 2018; Jauss et al., 2019; Matern et al., 2010; Götzke et al., 2014), yet their structures remain elusive. Figure 1a shows the screening results of PpiD and YfgM against the envelopome. Reassuringly, these two proteins stand out at the very top of the predicted PPI rankings among the E. coli envelopome with iScores of ~0.73 in each case (Figure 1a, Supplementary file 1). Furthermore, PpiD hits SecY at the 12th rank (iScore = 0.43), corroborating previous experimental studies that SecY is an interacting partner of PpiD (Antonoaea et al., 2008; Sachelaru et al., 2014; Fürst et al., 2018).

Figure 1 with 1 supplement see all
Download asset Open asset
E. coli super-translocon SecYEG/PpiD/YfgM.

(a) Computational screening for protein-protein interaction partners of PpiD and YfgM within the E. coli envelopome, respectively. A histogram displays the distribution of the top interface scores (iScores) of all envelope proteins screened with each query. Black arrows mark the top hits that were further studied, along with their names and overall ranks. (b) The top AF2Complex model of a supercomplex made of PpiD (blue), YfgM (green), SecY (silver), SecE (cyan), and SecG (tan) in three different views. Proteins are shown in a cartoon representation. Viewpoint transition, from either left to right or top to bottom, is indicated by a rotation axis (dashed line) and the rotation angle in degrees (circled arrow). (c, d, and e) Predicted PPI sites. The corresponding locations in b are indicated by black boxes. For clarity, the viewpoints and representations are adjusted. In the surface representations c and e, the color code is hydrophobic (white), polar (green), positive (blue), and negative (red), except for Phe122PpiD (yellow) in e. The same color code for the surface representation is employed below unless noted otherwise. PPI residues are shown in a ball-and-stick representation for PpiD in c, SecY in d, and SecG in e; the color scheme of atoms is carbon (cyan), oxygen (red), nitrogen (blue), and sulfur (yellow). The same scheme of atoms is adopted throughout this work.

SecYEG associates with the chaperone PpiD and assistant YfgM

Subsequently, we modeled SecYEG together with PpiD and YfgM and acquired another very high confidence structure (iScore = 0.73, mean predicted local distance difference test [pLDDT] = 79, see Methods; Figure 1b–e). Between this supercomplex and the dimer structures obtained from screening, the PPI interfaces between PpiD and YfgM are very similar according to iAlign (Gao and Skolnick, 2010), which yields a root-mean-square deviation (RMSD) of 0.6 Å for the Cα atoms of the interface residues. Thus, we focus on the supercomplex structure below.

Individually, PpiD and YfgM in our complex model display the same architecture as their respective monomer models calculated by applying AF2 (Varadi et al., 2022). PpiD by itself has four domains in an open-arm shape, fitting its proposed chaperone role (Antonoaea et al., 2008; Matern et al., 2010), plus an N-terminal transmembrane α-helix (Figure 1b). The membrane anchoring helix is linked to a domain consisting of two discontinuous segments (residues 40–189, 581–623) from the terminal regions. Within this domain, two α-helices (residues 167–189, 581–609) are held by the tetratricopeptide repeat (TPR) domain of YfgM, which inserts itself into the IM via a single hydrophobic α-helix. Structurally, the PpiD/YfgM interface resembles a pair of ‘chopsticks’ (two PpiD helices) grasped by a ‘palm’ (the TPR domain). Their PPI interactions are quite extensive, including ~100 inter-protein residue-residue contacts involving both hydrophobic and polar partners (Figure 1b and c), which may explain the strong binding affinity observed in experiments (Maddalo et al., 2011).

In addition to the channel that connects the cytosol to the periplasm, SecY features a lateral gate embedded in the IM. It opens to release transmembrane segments, including the signal peptide of a substrate such as OmpA (Li et al., 2016). In vivo photo cross-linking studies found numerous sites at the lateral gate probed by PpiD (Sachelaru et al., 2014; Jauss et al., 2019). Such cross-linking suggests direct physical contacts between the two proteins because the covalent linkers formed upon UV light radiation are very short, lying within 4 Å. Because PpiD has only a single transmembrane α-helix, a logical speculation is that the α-helix interacts with the lateral gate of SecY. Indeed, our structure reveals that PpiD guards the lateral gate of SecY with its α-helix (Figure 1b and d), which occupies the same position as a signal peptide leaving the SecY gate (Figure 1—figure supplement 1). In superposition with a SecY crystal structure co-crystallized with an OmpA signal peptide, the Cα RMSD of SecY is 2.4 Å between the X-ray and computed structures. Thus, SecY in our model is open to accommodate the α-helix of PpiD, in comparison to an electron microscopy (EM) structure of E. coli SecY in which the gate is closed (Figure 1—figure supplement 1). Our model further elucidates why SecY residues located in transmembrane helices (TM2/3/7/8) can be cross-linked to PpiD, because they are located at the PpiD/SecY interface. For example, strong cross-linking signals have been found for five SecY residues (Sachelaru et al., 2014; Jauss et al., 2019), Ile91SecY, Leu94SecY, Leu127SecY, Phe130SecY, and Phe286SecY, and all make direct physical contacts (<4.5 Å) with PpiD in our structural model (Figure 1d). Moreover, we identified direct contacts between a hydrophobic pocket of PpiD and SecG, explaining the experimentally observed crosslinking of Phe122PpiD to SecG (Fürst et al., 2018; Figure 1e). The interaction assigns SecG a functional role in coordinating with PpiD. Overall, the predicted structure rationalizes the results of several experimental studies.

The chaperone PpiD interacts with the disulfide isomerase DsbA

Among the top hits for PPI partners of PpiD, a fascinating discovery is DsbA, ranked third with a high confidence iScore of 0.55. DsbA donates its disulfide bond to a substrate in need; thus, it is critical to the folding of a nascent protein as it leaves the SecY channel (Kadokura and Beckwith, 2009; Goemans et al., 2014). How does DsbA coordinate with the SecY translocon? The predicted interaction immediately provides an answer, i.e., by associating with the chaperone PpiD that is a part of a super translocon complex (SecYEG/PpiD/YfgM/DsbA). In a predicted structure of this supercomplex (iScore = 0.71, mean pLDDT = 78, Figure 2a), PpiD uses its ‘hand’ to grip DsbA, as observed in the top model of the pair because their interactions do not interfere with other members of the translocon. The hand of PpiD is a parvulin-like domain, but it is devoid of the peptidyl-prolyl isomerase (PPIase) activity canonical to parvulin, due to the mutation of a critical histidine (Weininger et al., 2010). As our structure reveals, the catalytically inactive pocket of the PPIase domain is hydrophobic and buries the aromatic sidechain of Phe148DsbA, which becomes a hot-spot residue surrounded by six hydrophobic residues of PpiD upon complexation (Figure 2b). In addition, three pairs of salt-bridges are formed around the hydrophobic contacts, contributing to specific recognition. Notably, the catalytic cysteines of DsbA, Cys49DsbA and Cys52DsbA, are on the opposite side to the protein-protein interface (Figure 2a); thus, they are open to engage a substrate as it emerges from the SecY channel. Based on this model, we hypothesize that DsbA resides on PpiD transiently to improve its chance of encountering a substrate.

Figure 2
Download asset Open asset
Structural model of the SecYEG/PpiD/YfgM/DsbA supercomplex.

(a) Two views of the predicted structure. DsbA is shown in red, while the other proteins are colored the same as in Figure 1. Two cysteines, Cys49 and Cys52, essential to the enzymatic function of DsbA, are shown as spheres. (b) Protein-protein interaction sites between PpiD and DsbA. For clarity, tertiary structures are transparent. Key interacting residues are shown in the licorice representation for PpiD and in the ball-and-stick representation for DsbA.

SPase I LepB accesses an OMP substrate received by PpiD/YfgM

Polypeptide translocation triggers the dissociation of PpiD from SecY (Sachelaru et al., 2014). The modeled complex suggests that PpiD dissociation is realized by pushing a substrate out of the SecY gate, e.g., the signal peptide, which then repels the PpiD helix bound to the gate. For an OMP, a key next step is the removal of the signal peptide by the peptidase LepB in E. coli (Paetzel, 2014). How does it operate? The dissociation of PpiD from SecYEG vacates a space for LepB to approach a substrate. We generated a structural model of LepB in complex with a precursor OmpA (proOmpA) chain (residue 1–87) grasped by PpiD (iScore = 0.64, mean pLDDT = 80; Figure 3). The OmpA cleavage site, Ala21OmpA, directly contacts the LepB active site triad, Ser89LepB, Ser91LepB, and Lys146LepB at ~4 Å away from the alanine. In our model, the proOmpA chain contributes a β-strand to form a parallel β-sheet with LepB, as has been speculated (Paetzel, 2014). Overall, the structure provides a model for how LepB accesses a substrate received by PpiD/YfgM to cleave a signal peptide.

Figure 3
Download asset Open asset
Predicted structure of the PpiD/YfgM/LepB/OmpA supercomplex.

(a) Two views in the cartoon representation are shown. Colors: PpiD (blue), YfgM (green), LepB (magenta), and OmpA (residue 1–87, yellow). For clarity, representations of PpiD and LepB are transparent. (b) Close-up view of the OmpA signal peptide in the active site of LepB. Essential catalytic residues, Ser89, Ser91, Lys146, and Ser279 of LepB are shown in a licorice representation, and the cleavage site Ala21 and Ala19 of OmpA is shown as spheres.

The chaperone SurA opens to load an OMP substrate

After passage through a SecYEG translocon, a nascent OMP polypeptide is relayed to SurA, which convoys the substrate toward its next stop, BAM. A crystal structure of SurA exhibits three domains: a core domain (split into N- and C-terminal subdomains) and two PPIase domains, P1 and P2 (Bitto and McKay, 2002). P2 retains PPIase activity but P1 does not (Behrens et al., 2001). In this crystal structure, P1 further packs with the core domain forming a cradle-like structure (Bitto and McKay, 2002), referred to as the closed conformation of SurA. By contrast, another crystal structure of the SurA homodimer lacking P2s displays an open conformation, where two P1s are uncoupled from the core domains to hold a dodecapeptide at the dimeric interface between P1s (Xu et al., 2007). To explore conformations of SurA, we predicted structures for a single, full-length SurA and obtain two major conformations (Figure 4a). One conformation closely resembles the cradle-like closed conformation with a backbone Cα RMSD of 2.4 Å from the crystal structure. The other mimics the open conformation but incorporates P2 absent in the X-ray structure. Overall, the open conformation resembles a ‘three-prong hook,’ where both P1 and P2 swing away from their resting position in the closed conformation.

Figure 4 with 1 supplement see all
Download asset Open asset
Structural models of SurA in the absence and presence of an OmpA substrate.

(a) Open and closed conformations of monomeric SurA, consisting of the core domain (N-terminal region in gray and C-terminal in tan), P1 (purple), P2 (red). (b and c) Two structures of SurA in the presence of an OmpA substrate from two separate modeling runs. In both, SurA is open as in a. The β-barrel domain of OmpA is completely unfolded and generally does not maintain the same residue-residue contacts with SurA, except that an OmpA aromatic residue consistently makes π-π interactions with Tyr128SurA located in the crevice of the SurA core domain. Two β-signal residues, Y189 and F191 of OmpA, are also shown as spheres. The folded periplasmic domain of OmpA is bound to the P2 domain of SurA. (d) Envelopome protein-protein interaction screening of SurA identifies itself and BamA among the top hits. (e) Two views of the top predicted structure of a SurA dimer (green and cyan). (f) Superimposition of two open conformations from the monomeric and dimeric SurA. Subscripts indicate the stoichiometry. The color schemes correspond to those used in a and e. Only a single SurA from the dimeric model is shown in the superposition.

SurA functions mainly as a monomer but may act as a dimer for a large client (Li et al., 2018; Calabrese et al., 2020). How does SurA chaperone a substrate such as OmpA? OmpA consists of a β-barrel (N-terminal) (Pautsch and Schulz, 1998) and a periplasmic (C-terminal) domain (Ishida et al., 2014). To model the OmpA polypeptide, we reduced the size of multiple sequence alignments provided to AF2Complex and removed all its structural templates. Consequently, we generated a monomeric OmpA structure with a non-native, collapsed N-terminal domain and a native-like periplasmic domain (Figure 4—figure supplement 1). In the presence of SurA, the periplasmic domain maintains the same fold, but remarkably, the non-native β-barrel region completely unravels and wraps around SurA (Figure 4b–c). This is consistent with small angle neutron scattering data that suggests an expanded and unfolded OmpA substrate (Marx et al., 2020). The SurA/OmpA dimer is not stable as different wrapping configurations were observed in two separate structure models, leading to very low iScores of ~0.05, reflecting uncertainty due to the possibility of many conformations. Nevertheless, the SurA/OmpA models appear physical and provide a hypothetical basis for how the chaperone SurA could prevent a polypeptide chain from aggregating and present an unfolded polypeptide to BAM for its final assembly. Intriguingly, in both predicted complex structures, the OmpA polypeptide passes through a SurA crevice at its core domain, where Tyr126SurA consistently attracts a substrate aromatic residue with π-π interactions. This may be one mechanism by which SurA recognizes OMPs, which are enriched with aromatic residues (Xu et al., 2007; Hennecke et al., 2005). Moreover, the disordered β-barrel region has few interactions with SurA could explain how the system avoids high-energy expenditure for β-barrel release. Of note, two key ‘β-signal’ residues (Tommassen, 2010), Y189 and F191 of OmpA, do not make direct physical contact with SurA in both models.

To identify other interacting partners, we searched the E. coli envelopome for PPIs with SurA. Interestingly, SurA recognizes itself at a high confidence iScore of 0.53 (Figure 4d). The top model of a SurA homodimer exhibits twofold rotational symmetry, in which P1 domains are swapped, leading to an open conformation quite different from the above one (Figure 4e–f). This conformation, like a four-prong hook, provides SurA with another configuration to handle clients, potentially large ones, consistent with experimental conclusions (Li et al., 2018; Calabrese et al., 2020).

SurA and BamA specifically recognize each other

Excitingly, envelopome screening also confidently detected that SurA interacts with BamA, the anchoring subunit of BAM at the final stop of OMPs. To understand how SurA interacts with BAM, we subsequently used AF2Complex to probe potential interactions of SurA with all five BAM constituents (BamABCDE) and acquired a high confidence complex structure (iScore = 0.75, mean pLDDT = 84; Figure 5a–b). Because BAM has been extensively studied structurally (Tomasek and Kahne, 2021; Wu et al., 2020), we focus on describing its interaction with SurA, though the predicted BAM structure closely mimics a known crystal structure of the complex determined at 2.9 Å resolution (PDB 5D0O; Gu et al., 2016). The alignment of the two complex structures yielding a very high TM-score of 0.94. Overall, SurA mainly interacts with BamA, with similar interactions observed in both the top supercomplex and the BamA/SurA dimer structures, in the N-terminal domains of both proteins (Figure 5b–c). Gln23SurA to Val28SurA, largely missing in the crystal structures of SurA (Bitto and McKay, 2002; Xu et al., 2007), now form a β-sheet with the β2 strand of POTRA1, the N-terminal domain of BamA. In addition, hydrophobic contacts and two salt bridges are present in the same region. The structure explains the result that Asp26SurA is cross-linked to BamA (Wang et al., 2016) because Asp26SurA is located at the center of the SurA/BamA interface. Furthermore, SurA adopts a closed conformation in our model, but the P2 domain rotates ~45° from its crystal position (Bitto and McKay, 2002) to engage additional contacts with BamA and BamE (Figure 5d–e). For example, Asn336SurA forms a hydrogen bond with the backbone oxygen atom of Gly275BamA; in the same cross-linking study (Wang et al., 2016), Asn336SurA is implied to interact with BamA. These results invite a hypothesis that SurA and BamA initiate payload transfer via specific docking at their N-terminal domains.

Figure 5
Download asset Open asset
Structural model of SurA docked to β-barrel assembly machine (BAM).

(a and b) Two views of the top ranked supercomplex model in the cartoon representation. The BAM constituents are BamA (green), BamB (pink), BamC (yellow), BamD (blue), and BamE (black). The N-terminal POTRA1 domain of BamA provides the main interaction sites for SurA (violet). (c) Close-up view of the interaction sites at POTRA1 and the core domain of SurA. Interacting residues are shown in the licorice (SurA) and ball-and-stick (BamA) representations. (d) Crystal structure of SurA (magenta, PDB 1M5Y) superimposed onto the computed structure of the supercomplex. The magenta arrow indicates the change of location in P2 between two structures. BamCD are omitted for clarity. (e) Protein-protein interaction between P2 of SurA and POTRA4 of BamA and BamE.

Metalloprotease BepA flips a lid in complex with BAM

SurA led us to BamA, for which we conducted another envelopome PPI screening. The top six hits include all expected ones, the other four BAM factors and SurA, but unexpectedly, BepA ranked third with high confidence (Figure 6a–b). BepA was not anticipated because we were not aware of its relevance. However, a literature search quickly revealed that BepA is highly relevant because it cleans up stalled OMP folding at BAM (Narita et al., 2013; Soltes et al., 2017), and its interactions with BAM have been documented (Narita et al., 2013; Daimon et al., 2017), though no structure of the complex is available. Consequently, we modeled BepA and the full BAM complex altogether, obtaining a high confidence model (iScore = 0.68, mean pLDDT = 85, Figure 6c). Overall, the BepA/BamA interfaces are quite similar (interface Cα RMSD 1.2 Å) between the dimer and the supercomplex structure. Extensive contacts are present between all five periplasmic domains (POTRA1−5) of BamA and the two domains (protease and TPR) of BepA. They can be clustered into two main groups scattered between the POTRA1 and TPR and between POTRA3-4 and the protease domains (Figure 6d). Strong photo cross-linking signals to BamA have been observed previously for several residues of TPR including Phe404BepA (Narita et al., 2013; Daimon et al., 2017). According to our model, Phe404BepA belongs to the first interaction cluster with five BamA residues, particularly Pro47BamA. Another residue cross-linked to BamA is Gln428BepA, which establishes a specific hydrogen bond with Arg49BamA in our model.

Figure 6 with 1 supplement see all
Download asset Open asset
Structural models of β-barrel assembly machine (BAM) and BepA.

(a) Computational protein-protein interaction screening identifies BepA as a top hit to BamA. (b) Top structural model of the heterodimeric complex of BamA (green) and BepA (purple) in cartoon representation. The lid of BepA is colored red. The active sites with the protease domain of BepA are shown in a surface representation (orange). The five POTRA domains of BamA are labeled P1−P5. (c) Predicted structure of the BAM/BepA supercomplex. The lid of BepA extends to an open conformation. The image was created from the same viewpoint as b. (d) Specific residue-residue contacts between BamA and the tetratricopeptide repeat (TPR) and protease domains of BepA. (e) Close-up views of the lid and the BamA β-barrel in the surface (top) and cartoon representations (bottom). A hydrophobic contact between Ala180BepA and Leu780BamA is shown as spheres, and the lateral gate of BamA is between the β1 and β16 strands (dark blue).

Crystal structures of BepA alone have been solved, but a segment (residue 154–192) of the metalloprotease domain is absent in these structures, presumably due to its flexibility (Daimon et al., 2017; Bryant et al., 2020; Shahrizal et al., 2019; Figure 6—figure supplement 1). The segment, termed an ‘active-site lid,’, was speculated to be important to the protease function of BepA because its movement could either expose (open state) or cover (closed state) the catalytic site, the HEXXH motif (Bryant et al., 2020). Intriguingly, we only see the closed conformation in 40 models of the BamA/BepA heterodimer (Figure 6b), but both the open and closed states in 8 supercomplex models. In the open state, two α-helices comprising the lid are in a straight-up configuration (Figure 6c), whose upper half (residue 157–186) rotates ~105° to cover the enzymatic site in the closed configuration (Figure 6b). Moreover, the flexible lid, with a large hydrophobic surface similar to the β-barrel of BamA, is in a position as if it is inserted into the inner leaflet of the OM and is in the proximity of the lateral gate of BamA (Figure 6e). As such, BepA could control its protease activity via probing a substrate β-barrel stemming out of BamA. Hence, it is hypothesized that a normal budding β-barrel blocks the lid opening, whereas a stalled OMP within BamA permits lid opening and subsequent substrate cleavage. It is also possible that the lid of BepA somehow detects an abnormal β-barrel popping out of BamA, as suggested in an experimental study that shows BepA degrades a LptD mutant stalled at a late stage of β-barrel folding (Soltes et al., 2017).

Discussion

Here, we have demonstrated a deep learning strategy that combines virtual PPI screening over the E. coli envelopome and supercomplex structure modeling. By applying it to several key proteins in the OMP biogenesis pathway, we have identified their functional partners within the top 1% ranking of ~1450 proteins screened for PPIs per query. Thanks to high confidence structures underlying the top predictions, one can understand many experimental phenomena, particularly in vivo site-directed photo cross-linking data. For example, cross-linked products found from the SecYEG or BAM supercomplexes may be explained by direct physical interactions revealed in our predicted structures. Moreover, previously speculated conformations are captured for SurA and BepA. Most importantly, these revealing atomic structures suggest mechanistic hypotheses for various steps of the OMP biogenesis pathway as summarized in Figure 7, where we present their diagrams along with some predicted supercomplex structures.

Figure 7
Download asset Open asset
Proposed mechanisms involved in the outer membrane protein (OMP) biogenesis pathway in E. coli.

Complex structures resulting from this study accompany relevant cartoon diagrams. Powered by SecA, a precursor OmpA polypeptide (orange line) first passes through the SecYEG translocon in complex with PpiD and YfgM. PpiD, held in place by YfgM, senses the translocating substrate via its N-terminal α-helix bound to the lateral gate of SecY and temporarily dissociates from the translocon upon receiving the substrate OmpA. Protein disulfide isomerase DsbA is recruited by PpiD and promotes formation of a disulfide bond between two cysteine residues (yellow spheres) of OmpA. Meanwhile, peptidase LepB fills the vacancy left by SecYEG and cleaves the transmembrane signal peptide from OmpA, which is then handed over to chaperone SurA. At this point, the periplasmic domain of OmpA is folded, but the unfolded β-barrel region wraps around SurA, which carries OmpA to BAM. Lastly, SurA docks to POTRA1, the N-terminal domain of BamA, where the β-barrel domain of OmpA is folded and released from the lateral gate of BamA. If this folding and assembly process are stalled for some reason, metalloprotease BepA senses the failure with its flexible lid and cleans up by cleaving a stuck substrate. For clarity, the peptidoglycan layer in the periplasm is not shown, and the schematic drawings are not to scale.

One unexpected discovery is the DsbA/PpiD interaction. It was known that DsbA crucially transfers its disulfide bond to a nascent polypeptide translocated by SecYEG (Kadokura and Beckwith, 2009; Goemans et al., 2014), but how does the translocon interact with DsbA? The predicted supercomplex of SecYEG/PpiD/YfgM/DsbA provides a compelling answer for three reasons: first, the transient residence of DsbA on PpiD at the translocon dramatically improves its efficiency versus a random search in a crowded periplasm. Second, it predicts a function for the PPIase domain of PpiD that mysteriously lacks the PPIase function. Third, since PpiD is anchored to the IM, it can keep DsbA close to DsbB, an IM protein that recycles DsbA. Notably, two other E. coli chaperones, SurA and FkpA, possess two PPIase domains each, and AF2Complex predicts that DsbA interacts with FkpA (iScore = 0.47 versus 0.55 for PpiD) but not with SurA (iScore = 0.03).

Interestingly, the AF2 neural networks can model a partially unfolded polypeptide accompanied by chaperones, even though AF2 was trained on folded proteins. We attained structural models that appear to be at least partially physical. One example is the model of PpiD/YfgM/LepB/proOmpA, in which proOmpA is posed for cleavage by peptidase LepB. The cleavage alanine of OmpA is ~4 Å away from the catalytic triad of LepB. Considering that only apo or inhibitor bound structures of LepB are available, this computed model implies that AF2 has learned physical representations. More intriguing examples are the models of SurA/OmpA, where the periplasmic domain of OmpA is folded, but the β-barrel domain is completely unfolded and loosely wrapped around SurA. The predicted structures echo the NMR structures of an unfolded polypeptide bound to SecB, a cytosolic chaperone involved in the early stage of the SecA-dependent translocation pathway (Huang et al., 2016). 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.

These results reinforce the notion that deep learning is a promising way to explore the conformational ensemble of proteins and to uncover the molecular mechanisms of biosystems (Gao et al., 2022b). The combination of advanced AF2 deep learning models, an effective PPI ranking metric, and a workflow optimized for large-scale screening generates illuminating structures of protein complexes. In the presented examples, we focused on those with obvious biological relevance according to the literature. There are of course other confident PPI predictions that were not described here. While some are promiscuous interactions or physically possible but biologically irrelevant, or simply false, there are likely functional interactions yet to be explored. More generally, these results are an example of a deep learning-based strategy that helps to elucidate mechanistic aspects of complex biochemical pathways.

Methods

E. coli cell envelopome

The proteome for E. coli strain K12 MG1655, consisting of 4400 protein sequences, was retrieved from UniProt (Wu et al., 2006) in March 2022 (Proteome ID UP000000625). Then, the subcellular location of each protein was parsed to collect all known and predicted cell envelope proteins. To be conservative, we included all proteins whose primary subcellular location is not the cytosol, yielding 1466 proteins defined as the envelopome. The set contains all proteins located in the IM, periplasm, OM, and extracellular surface, and some cytosolic proteins located primarily at the periphery of the periplasmic IM, e.g., SecA. For AF2Complex modeling, we removed 11 sequences that contain at least one non-standard amino acid.

AF2Complex

An updated version of AF2Complex was built upon AlphaFold version 2.2.0 (Evans et al., 2021). AF2Complex supports three different set of deep learning neural network models provided with AF2, i.e., the original models for monomer prediction (AF version 2.0.1 Jumper et al., 2021) and two set of models for multimer prediction (version 2.1.0 and version 2.2.0; Evans et al., 2021). Numerous changes were made; we list three major ones here. First, the iScore metric was introduced to rank the confidence of a predicted complex model (Gao et al., 2022b). The iScore metric was derived from the predicted alignment errors that give 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 (Gao et al., 2022b; Jumper et al., 2021). To better estimate confidence, the contribution of each interface residue to the iScore is calculated using local frames not located within the same protein chain, i.e., residue i and j belonging to different chains. Second, the data pipeline for generating input features was split from the neural network inference. This allows rapid assembly of input features for individual proteins of a putative complex from pre-generated input features of the full proteome. Third, five options for multiple sequence alignment (MSA) pairing are provided: no pairing, all paired, cyclic, linear, and arbitrary pairing. The last three options are experimental, and we only employed the first two MSA pairing modes in this work. All structural models, either of a monomer or a multimer, were predicted by AF2Complex using either the ‘monomer_ptm’ deep learning models without any MSA pairing or the ‘multimer_v2’ models with all MSA pairing; the latter only used for structure prediction of complexes. The input features of all monomers were derived using standard sequence libraries (The Uniprot Consortium, 2019; Mitchell et al., 2020; Tunyasuvunakool et al., 2021) and a version of the Protein Data Bank (Berman et al., 2000) released in November 2021. The confidence of an output structure is evaluated using three different metrics: the iScore for protein interface (Gao et al., 2022b), the predicted TM-score for global structure of a monomer (Zhang and Skolnick, 2004), and the pLDDT score for local domain structure (Jumper et al., 2021; Mariani et al., 2013). According to AF2, a mean pLDDT score higher than 70 indicates high confidence in a predicted structure when evaluated on individual domain(s).

Envelopome PPI screening

We adapted a workflow for proteome-scale monomeric structure prediction with AF2 implemented on the Summit supercomputer at Oak Ridge National Laboratory (Gao et al., 2022a). The workflow was applied to four query proteins: PpiD, YfgM, SurA, and BamA. Given a query sequence and an envelopome protein sequence, AF2Complex assembles the pre-generated monomeric input features and then feeds the composite features to the AF2 neural network models for inference. 10 different neural network models are used as mentioned above, and each deep learning model is invoked twice with two different random seeds and up to eight recycles. This procedure typically gives 20 final structural models per target for ranking, and the top ranked model by iScore is retained. Although the ‘multimer_v2’ set of models greatly reduces unphysical clashes in predicted structures compared to structures predicted by the original set of multimer deep learning models, there is still a small chance of generating severe clashes, which could yield an artifactually high iScore. We filter out these structures by applying a minimum interface clash indicator of 0.4 (Gao et al., 2022b). We did not apply model relaxation to these predicted structures because most are not actual protein complexes; as such, it is not worth expending the extra computing time. To avoid memory overflow, we limited the input MSA depth to 5000 for each monomer and allowed a maximum of four structural templates per monomer. A maximum of 1600 amino acids was imposed on the total size of each complex target. This limitation slightly reduced the total number proteins in the envelopome for screening, typically to ~1450 for the four proteins studied. Because about 30% of envelopome proteins have a signal peptide that is absent in their mature chains, only mature chains were used for screening. This is readily realized with AF2Complex because it provides an option to crop arbitrary segments of a monomer input feature during feature assembly. The residue ranges of the mature chains were obtained from the UniProt knowledgebase.

Modeling the OmpA polypeptide

It was necessary to model OmpA as a substrate of PpiD or SurA. OmpA was chosen mainly because it is a model OMP for studying the OMP biogenesis pathway (Reusch, 2012). Many experimental data exist for validation. To minimize potential ‘memory’ effects due to large MSAs or structural templates, we reduced the number of sequences in the MSAs and removed all structural templates in the input features of OmpA. We tested MSA depths of 1, 10, 20, 50, and 100 to predict a structural model of proOmpA in complex with SecYEG. The goal is to generate a structural model that mimics the crystal structure of SecYEG translocating a proOmpA polypeptide (Li et al., 2016). Our tests found that an MSA depth below 50 can yield a model structure of an unfolded OmpA peptide through the channel of SecY, when used with either ‘model_1_ptm’ or ‘model_3_ptm’ of all AF2 neural network models. In this study, we used an MSA depth of 20 and the 2 AF2 models to predict structural models involving OmpA.

Computational resources

E. coli envelopome PPI screening was performed on the Summit supercomputer, typically using 923 nodes for several hours of wall clock time. Each Summit node hosts six Nvidia 16 GB V100 GPUs. The structure predictions of various supercomplexes were conducted locally using about 10 workstations each with four Nvidia RTX6000 GPUs, each with 24 GB of GPU memory.

Analysis

The program VMD (Humphrey et al., 1996) was used to inspect predicted structural models and create all molecular images. APoc was used to align monomeric protein structures (Gao and Skolnick, 2013), and iAlign was used to perform protein-protein interface comparison (Gao and Skolnick, 2010).

Data availability

The input features for the full E. coli proteome used for envelopome screening by AF2Complex, and the computational models presented in this study are available at Zenodo (https://doi.org/10.5281/zenodo.6846915). The source code of AF2Complex is freely available at https://github.com/FreshAirTonight/af2complex.

References

    1. Rollauer SE
    2. Sooreshjani MA
    3. Noinaj N
    4. Buchanan SK
    (2015) Outer membrane protein biogenesis in gram-negative bacteria
    Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 370:1679.
    https://doi.org/10.1098/rstb.2015.0023

Decision letter

  1. José D Faraldo-Gómez
    Senior and Reviewing Editor; National Institutes of Health, United States
  2. Nir Ben-Tal
    Reviewer; Tel Aviv University, Israel
  3. Arne Elofsson
    Reviewer; Stockholm University, Sweden

Our editorial process produces two outputs: (i) public reviews designed to be posted alongside the preprint for the benefit of readers; (ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "Deep learning-driven insights into super protein complexes for outer membrane protein biogenesis in bacteria" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, overseen by José Faraldo-Gómez as the Senior Editor. Two of the reviewers have agreed to reveal their identity, namely Nir Ben-Tal (Reviewer #1); Arne Elofsson (Reviewer #3).

As you will see below, the reviewers conclude the manuscript is not suitable for publication for eLife in its current form, but they make specific recommendations to resolve their concerns. I recognize some of these revisions are substantial – nevertheless, I encourage you to take the time required to address these concerns convincingly and then submit a revised version.

Reviewer #2 (Recommendations for the authors):

In this manuscript, Gao et al., combine a very elegant virtual screening method (AF2Complex) based on Α Fold with known crystal structures to predict the structure of super-protein complexes in the E. coli cell envelope that play important roles in the biogenesis of outer membrane proteins (OMPs). The authors do an excellent job of validating their methodology. Their predicted structures both explain a considerable amount of biochemical data that has appeared in the literature (e.g., the crosslinking of PpiD to SecY and SurA to BamA) and provide compelling explanations for unresolved conceptual issues (e.g., how DsbA is recruited to the Sec machinery to form disulfide bonds as proteins are translocated into the periplasm). With one exception (see comment 4 below), the predicted structures also provide potentially important insights into the function of the assembly factors. In essence, this work describes an impressive example of the power of AI that should be of interest not only to investigators who study OMP biogenesis but also to the broader research community. I should also note that the manuscript is clearly and concisely written for a general audience.

Specific comments:

1) Lines 83-86 and Figures 1a, 4a and 6a: The authors selected predicted structures of known biological relevance for further study, but they do not comment on high confidence predictions of supercomplex structures of unknown biological significance. How can they distinguish predicted structures that have potential biological significance from those that simply represent noise/false positives? This is an important issue that requires further explanation, especially because the authors provide false-positive rates on lines 76/77.

2) In the paper by Alvira et al., (ref. 21) the authors argue that E. coli produce a "holotranslocon" in which SecDF and BAM interact across the periplasm. I was surprised that AF2Complex did not detect this interaction with a high confidence score. Does AF2Complex have a significant rate of "false negatives"? The authors should discuss this issue.

3) Lines 181-183 and Figure 4c: As suggested by many different studies, the C-terminal β signal of OmpA must be accessible to bind to the first β strand of BamA at a relatively early stage of assembly. Would the last β strand be accessible if the β barrel is wrapped around SurA? The authors should comment on this issue.

4) Lines 246-249: While the structural model of the BAM-BepA supercomplex shown in Figure 6 is intriguing, the idea that the protease activity of BepA is controlled by the probing of substrate β barrels as they bud from the BamA β barrel does not appear to be consistent with the literature. The authors propose that β barrels that are stalled within BamA permit the opening of the BepA lid and activation of the protease. This suggests that BepA degrades β barrels that are stalled at an early stage of assembly, but not at later stages. The work of Soltes, et al., (ref. 48), however, shows that BepA degrades an LptD mutant that has formed a nearly complete barrel (and that likely protrudes from the BamA barrel) while another protease (YcaL) degrades a second LptD mutant that engages BAM at an earlier stage of assembly. I realize that the stages of OMP assembly are currently rather vague, but the authors should discuss the work of Soltes et al., and state that it imposes a possible caveat on their model.

Reviewer #3 (Recommendations for the authors):

1) What is the advantage of running both monomer and multimer versions of AF2? Does it provide an advantage? In our benchmarks, the multimer is slightly better and roughly doubles the computational cost. Also, did you run 2 multimer versions (version 2.1.0 should not be used it makes bad predictions often)?

2) Why is not the top-hit for surA (oppA) discussed? Or any other of the high-scoring pairs?

3) The authors claim that an iScore of 0.4 is related to a 1.2% FPR. This would mean that yfgM would have roughly 100 interactions. Is this correct? If so, this would need to be discussed (and possibly proven experimentally). Here only discussion with PPID at rank 12 is discussed. Should it not be more likely that the other 11 higher-ranked models interact? Why are these ignored?

4) Would it not be computationally more efficient to use MMseqs2 to make the MSAs?

5) Why is it only tested on 4 proteins? I think it should be computationally possible to run it on many more (in particular as you already have the MSAs). At a minimum, all complexes shown in Figure 7 should be run through the pipeline to ensure these can be modelled.

6) How would the pipeline handle stoichiometry?

7) WHat happens if a set of the high-scoring pairwise interactions are fed into alphafold-multimer (which can handle up ~5000 residues). Can the large complexes be modelled?

https://doi.org/10.7554/eLife.82885.sa1

Author response

Reviewer #2 (Recommendations for the authors):

1) Lines 83-86 and Figures 1a, 4a and 6a: The authors selected predicted structures of known biological relevance for further study, but they do not comment on high confidence predictions of supercomplex structures of unknown biological significance. How can they distinguish predicted structures that have potential biological significance from those that simply represent noise/false positives? This is an important issue that requires further explanation, especially because the authors provide false-positive rates on lines 76/77.

We would like to clarify that a true positive of our method is about predicting physical interactions between the two proteins subjected to modeling alone. It does not consider many other factors such as stoichiometry, binding competition with other proteins, post-translational modifications, and functional relevance. As a result, we must review other sources of evidence and select ones that likely have biological relevance according to existing literature. Ideally, such a selection requires the involvement of experts of these proteins or pathways. For this reason, we provide the lists of all top predictions in case that the hits we ignored could turn out to be useful for future studies by the respective experts.

2) In the paper by Alvira et al., (ref. 21) the authors argue that E. coli produce a "holotranslocon" in which SecDF and BAM interact across the periplasm. I was surprised that AF2Complex did not detect this interaction with a high confidence score. Does AF2Complex have a significant rate of "false negatives"? The authors should discuss this issue.

The paper referred to by the reviewer is one of several intriguing works that led us to this study. While we were able to obtain high confidence models for some complexes found in previous experimental studies, in the case of BamA/SecDF, unfortunately we did not observe a confident computational model, despite multiple run efforts. In retrospect, it is unclear to us whether or not direct interactions between BamA and SecDF alone would be feasible in a normal periplasmic space, which is about 20 nm wide from the inner membrane to outer membrane. Having said that, it is entirely possible that our computational method missed the target as a false negative, as the reviewer speculated. In this revision, we have added statistics of the recall (1 – false negative rate), from line 77,

“With respect to the capability of identifying and modeling true protein-protein interactions, on a set of 440 heterodimeric complexes whose experimental structures have been recently determined and were not used for AF2 model training, AF2Complex recalls 81%, 74% and 34% of these benchmark targets at the same three iScore thresholds, respectively, and yields medium- or high-quality complex structures for 84%, 87%, and 93% in the top ranked model of these positively identified targets [23].”

3) Lines 181-183 and Figure 4c: As suggested by many different studies, the C-terminal β signal of OmpA must be accessible to bind to the first β strand of BamA at a relatively early stage of assembly. Would the last β strand be accessible if the β barrel is wrapped around SurA? The authors should comment on this issue.

We thank the reviewer for this suggestion. We have verified that the so-called β-signal residues are accessible in our OmpA/SurA complex models.

We marked the two referred to residues in the revised Figure 4 and added the following sentence in line 211,

“Of note, two key ‘β-signal’ residues [47], Y189 and F191 of OmpA, do not make direct physical contact with SurA in both models.”

4) Lines 246-249: While the structural model of the BAM-BepA supercomplex shown in Figure 6 is intriguing, the idea that the protease activity of BepA is controlled by the probing of substrate β barrels as they bud from the BamA β barrel does not appear to be consistent with the literature. The authors propose that β barrels that are stalled within BamA permit the opening of the BepA lid and activation of the protease. This suggests that BepA degrades β barrels that are stalled at an early stage of assembly, but not at later stages. The work of Soltes, et al., (ref. 48), however, shows that BepA degrades an LptD mutant that has formed a nearly complete barrel (and that likely protrudes from the BamA barrel) while another protease (YcaL) degrades a second LptD mutant that engages BAM at an earlier stage of assembly. I realize that the stages of OMP assembly are currently rather vague, but the authors should discuss the work of Soltes et al., and state that it imposes a possible caveat on their model.

We thank the reviewer for pointing out these experimental observations that we had inadvertently overlooked. Our hypothesis is a quite simplified reasoning based on computational models and some experimental studies that we were aware of at the time we wrote the original version of this paper. We have now revised the hypothesis and added an alternative possibility on line 271,

“It is also possible that the lid of BepA somehow detects an abnormal β-barrel popping out of BamA, as suggested in an experimental study that shows BepA degrades a LptD mutant stalled at a late stage of β-barrel folding [51].”

Reviewer #3 (Recommendations for the authors):

1) What is the advantage of running both monomer and multimer versions of AF2? Does it provide an advantage? In our benchmarks, the multimer is slightly better and roughly doubles the computational cost. Also, did you run 2 multimer versions (version 2.1.0 should not be used it makes bad predictions often)?

We agree with the reviewer that the multimer_v2 deep learning models by DeepMind (AF version 2.1.0) provide improved accuracy in predicted models, and we employed this version of multimer models instead of the multimer_v1 version. However, in our benchmarks, we found that the original monomer deep learning models (AF version 2.0.1) make good predictions not seen with the multimer models in some cases. For example, as we mentioned in the Methods (line 382), only certain monomer deep learning models could generate models of an unfolded OmpA peptide through the channel of SecY mimicking the crystal structure (Ref. 12). For this reason, we included the monomer models in our procedure as well. This practice slightly improves the coverage of our approach.

2) Why is not the top-hit for surA (oppA) discussed? Or any other of the high-scoring pairs?

3) The authors claim that an iScore of 0.4 is related to a 1.2% FPR. This would mean that yfgM would have roughly 100 interactions. Is this correct? If so, this would need to be discussed (and possibly proven experimentally). Here only discussion with PPID at rank 12 is discussed. Should it not be more likely that the other 11 higher-ranked models interact? Why are these ignored?

We thank the reviewer for these good questions, which we have addressed in R1.2 as they were raised similarly by the other two reviewers. We again stress that our selections from predicted interacting protein pairs were based on biological relevance that we were aware of. With respect to the false positive rate, it was based on benchmarking a set of putatively non-interacting protein pairs in E. coli (Ref 23). It is possible that some of these false positives with high scores could physically interact if the proteins were actually present at the same subcellular location, but in the cell they cannot as they are located in different cellular compartments, one main criterion employed for selecting these control negatives. Obviously, such interacting pairs are unlikely to be biologically relevant. Moreover, even though we screened about 1,500 envelope proteins, the numbers of confident hits are generally low, e.g., 9 for BamA, 12 for PpiD, and it is clear that at least some of these hits are not false positives. The query protein with highest number of confident hits is YfgM, which has 98 hits above an iScore of 0.4. They are all attracted to the plam-like surface of YfgM (Figure 1c), the same surface where PpiD is predicted to make extensive contacts with YfgM. Thus, PpiD likely forms a very stable complex with YfgM that other hits may not be able to compete with. Therefore, the biological relevance of other YfgM hits is unclear. We skipped all hits that we could not convince ourselves by some molecular hypothesis that they are relevant to the OMP biogenesis pathway.

4) Would it not be computationally more efficient to use MMseqs2 to make the MSAs?

We cannot comment on the efficiency of MMseqs2 because we have not benchmarked its performance. However, as mentioned in the Methods, we have implemented a workflow in AF2Complex such that it can re-use the input features including the MSAs of monomers for multimer predictions. This means that one needs to generate MSAs only once for each protein in a proteome and reassemble them quickly for predicting arbitrary combinations of multimers. This eliminates the need for feature derivation for each multimer target as is required in DeepMind’s workflow, which is far less efficient, especially for large-scale screening. We have shared the input features of the full E. coli proteome at Zenodo (see Data availability), and we hope that they could be helpful for modeling other E. coli protein complexes.

5) Why is it only tested on 4 proteins? I think it should be computationally possible to run it on many more (in particular as you already have the MSAs). At a minimum, all complexes shown in Figure 7 should be run through the pipeline to ensure these can be modelled.

All complexes presented in Figure 7 actually resulted from the modeling efforts of this study, though parts of these complexes were experimentally determined in previous studies. We wish that we could conduct all-against-all PPI screening of E. coli., but we could not afford the huge computational costs to do such a study. Another limiting factor is the brain power required for analyzing the results, as we have neither the resources nor the expertise to analyze hundreds or thousands of PPI predictions by ourselves. The main purpose of the study is to demonstrate the usefulness of a new approach to the study of biological pathways through what we believe biologically relevant and important examples.

6) How would the pipeline handle stoichiometry?

The pipeline requires a user to specify a stoichiometry of a complex target. In this study, we rely on the literature to determine the stoichiometry. If such information is not available, one could run some trials at different stoichiometries and select the stoichiometry that gives the most confident model. While this is possible in some cases, it is in general a challenging issue.

7) WHat happens if a set of the high-scoring pairwise interactions are fed into alphafold-multimer (which can handle up ~5000 residues). Can the large complexes be modelled?

As we show for the BAM/BepA complex, it is possible to make confident models for large complexes that are made of multiple high-scoring pairwise interactions. However, it is also our experience that large scale modeling is challenging when the total size of the complex is over 3000 residues. This is very much a frontier research topic that requires further exploration.

https://doi.org/10.7554/eLife.82885.sa2

Article and author information

Author details

  1. Mu Gao

    Center for the Study of Systems Biology, School of Biological Sciences, Georgia Institute of Technology, Atlanta, United States
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Supervision, Funding acquisition, Investigation, Visualization, Methodology, Writing - original draft, Writing - review and editing
    For correspondence
    mu.gao@gatech.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0378-3704

  2. Davi Nakajima An

    School of Computer Science, Georgia Institute of Technology, Atlanta, United States
    Contribution
    Software, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared

  3. Jeffrey Skolnick

    Center for the Study of Systems Biology, School of Biological Sciences, Georgia Institute of Technology, Atlanta, United States
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition, Project administration, Writing - review and editing
    For correspondence
    skolnick@gatech.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-1877-4958

Funding

U.S. Department of Energy (DE-SC0021303)

  • Jeffrey Skolnick

National Institute of General Medical Sciences (R35GM118039)

  • Jeffrey Skolnick

Advanced Scientific Computing Research

  • Mu Gao
  • Jeffrey Skolnick

Advanced Computing Environment (PACE)

  • Jeffrey Skolnick

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Jerry M Parks for stimulating discussions and critical reading of the manuscript. This work was supported in part by the DOE Office of Science, Office of Biological and Environmental Research (DOE DE-SC0021303) and the Division of General Medical Sciences of the National Institutes Health (NIH R35GM118039). The research used resources supported in part by the Advanced Scientific Computing Research (ASCR) Leadership Computing Challenge (ALCC) program, and by the Partnership for an Advanced Computing Environment (PACE) at the Georgia Institute of Technology.

Senior and Reviewing Editor

  1. José D Faraldo-Gómez, National Institutes of Health, United States

Reviewers

  1. Nir Ben-Tal, Tel Aviv University, Israel
  2. Arne Elofsson, Stockholm University, Sweden

Publication history

  1. Received: August 22, 2022
  2. Preprint posted: August 25, 2022 (view preprint)
  3. Accepted: November 28, 2022
  4. Version of Record published: December 28, 2022 (version 1)

Copyright

© 2022, Gao et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,970
    Page views
  • 257
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Mu Gao
  2. Davi Nakajima An
  3. Jeffrey Skolnick
(2022)
Deep learning-driven insights into super protein complexes for outer membrane protein biogenesis in bacteria
eLife 11:e82885.
https://doi.org/10.7554/eLife.82885

Categories and tags

Research organism

Further reading

    1. Computational and Systems Biology
    2. Neuroscience

    High-resolution volumetric imaging constrains compartmental models to explore synaptic integration and temporal processing by cochlear nucleus globular bushy cells

    George A Spirou, Matthew G Kersting ... Paul B Manis
    Research Article

    Globular bushy cells (GBCs) of the cochlear nucleus play central roles in the temporal processing of sound. Despite investigation over many decades, fundamental questions remain about their dendrite structure, afferent innervation, and integration of synaptic inputs. Here, we use volume electron microscopy (EM) of the mouse cochlear nucleus to construct synaptic maps that precisely specify convergence ratios and synaptic weights for auditory- nerve innervation and accurate surface areas of all postsynaptic compartments. Detailed biophysically-based compartmental models can help develop hypotheses regarding how GBCs integrate inputs to yield their recorded responses to sound. We established a pipeline to export a precise reconstruction of auditory nerve axons and their endbulb terminals together with high-resolution dendrite, soma, and axon reconstructions into biophysically-detailed compartmental models that could be activated by a standard cochlear transduction model. With these constraints, the models predict auditory nerve input profiles whereby all endbulbs onto a GBC are subthreshold (coincidence detection mode), or one or two inputs are suprathreshold (mixed mode). The models also predict the relative importance of dendrite geometry, soma size, and axon initial segment length in setting action potential threshold and generating heterogeneity in sound-evoked responses, and thereby propose mechanisms by which GBCs may homeostatically adjust their excitability. Volume EM also reveals new dendritic structures and dendrites that lack innervation. This framework defines a pathway from subcellular morphology to synaptic connectivity, and facilitates investigation into the roles of specific cellular features in sound encoding. We also clarify the need for new experimental measurements to provide missing cellular parameters, and predict responses to sound for further in vivo studies, thereby serving as a template for investigation of other neuron classes.

    1. Developmental Biology
    2. Computational and Systems Biology

    Precise temporal control of neuroblast migration through combined regulation and feedback of a Wnt receptor

    Erik S Schild, Shivam Gupta ... Hendrik C Korswagen
    Research Article Updated

    Many developmental processes depend on precise temporal control of gene expression. We have previously established a theoretical framework for regulatory strategies that can govern such high temporal precision, but experimental validation of these predictions was still lacking. Here, we use the time-dependent expression of a Wnt receptor that controls neuroblast migration in Caenorhabditis elegans as a tractable system to study a robust, cell-intrinsic timing mechanism in vivo. Single-molecule mRNA quantification showed that the expression of the receptor increases non-linearly, a dynamic that is predicted to enhance timing precision over an unregulated, linear increase in timekeeper abundance. We show that this upregulation depends on transcriptional activation, providing in vivo evidence for a model in which the timing of receptor expression is regulated through an accumulating activator that triggers expression when a specific threshold is reached. This timing mechanism acts across a cell division that occurs in the neuroblast lineage and is influenced by the asymmetry of the division. Finally, we show that positive feedback of receptor expression through the canonical Wnt pathway enhances temporal precision. We conclude that robust cell-intrinsic timing can be achieved by combining regulation and feedback of the timekeeper gene.

    1. Computational and Systems Biology
    2. Neuroscience

    Model discovery to link neural activity to behavioral tasks

    Jamie D Costabile, Kaarthik A Balakrishnan ... Martin Haesemeyer
    Research Article

    Brains are not engineered solutions to a well-defined problem but arose through selective pressure acting on random variation. It is therefore unclear how well a model chosen by an experimenter can relate neural activity to experimental conditions. Here we developed 'Model identification of neural encoding (MINE)'. MINE is an accessible framework using convolutional neural networks (CNN) to discover and characterize a model that relates aspects of tasks to neural activity. Although flexible, CNNs are difficult to interpret. We use Taylor decomposition approaches to understand the discovered model and how it maps task features to activity. We apply MINE to a published cortical dataset as well as experiments designed to probe thermoregulatory circuits in zebrafish. MINE allowed us to characterize neurons according to their receptive field and computational complexity, features which anatomically segregate in the brain. We also identified a new class of neurons that integrate thermosensory and behavioral information which eluded us previously when using traditional clustering and regression-based approaches.