Reconstructing the transport cycle in the sugar porter superfamily using coevolution-powered machine learning
Curation statements for this article:-
Curated by eLife
eLife assessment
This potentially important work proposes a novel approach, based on co-evolution analysis, machine-learning protocols and molecular dynamics simulations, to predict structures and energetics of the main states of the alternating access cycle of a family of membrane transporters, the sugar porters. The approach is compelling, especially the application of co-evolution and Alphafold to generate highly accurate models in different conformational states of a given protein, but the work is currently incomplete due to shortcomings in the calculation of the energy landscape. With this aspect strengthened, the manuscript will be of interest to the transporter and computational modeling communities.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Structural Biology and Molecular Biophysics (eLife)
Abstract
Sugar porters (SPs) represent the largest group of secondary-active transporters. Some members, such as the glucose transporters (GLUTs), are well known for their role in maintaining blood glucose homeostasis in mammals, with their expression upregulated in many types of cancers. Because only a few sugar porter structures have been determined, mechanistic models have been constructed by piecing together structural states of distantly related proteins. Current GLUT transport models are predominantly descriptive and oversimplified. Here, we have combined coevolution analysis and comparative modeling, to predict structures of the entire sugar porter superfamily in each state of the transport cycle. We have analyzed the state-specific contacts inferred from coevolving residue pairs and shown how this information can be used to rapidly generate free-energy landscapes consistent with experimental estimates, as illustrated here for the mammalian fructose transporter GLUT5. By comparing many different sugar porter models and scrutinizing their sequence, we have been able to define the molecular determinants of the transport cycle, which are conserved throughout the sugar porter superfamily. We have also been able to highlight differences leading to the emergence of proton-coupling, validating, and extending the previously proposed latch mechanism. Our computational approach is transferable to any transporter, and to other protein families in general.
Article activity feed
-
-
eLife assessment
This potentially important work proposes a novel approach, based on co-evolution analysis, machine-learning protocols and molecular dynamics simulations, to predict structures and energetics of the main states of the alternating access cycle of a family of membrane transporters, the sugar porters. The approach is compelling, especially the application of co-evolution and Alphafold to generate highly accurate models in different conformational states of a given protein, but the work is currently incomplete due to shortcomings in the calculation of the energy landscape. With this aspect strengthened, the manuscript will be of interest to the transporter and computational modeling communities.
-
Reviewer #1 (Public Review):
This manuscript harnesses recent advances in co-evolution based modeling and computational approaches to provide molecular details about the transport cycles and mechanisms of an entire family of transporters, the sugar porters. The authors evaluate the validity of their approach in a number of ways, including comparison to structurally characterized proteins/states excluded from the training set, comparison to the GLUT5 transport free energy landscape determine through conventional enhanced MD methods in a companion paper, and a global evaluation of RMSDs between models. Based on these structural models, the authors are able to generate a number of interesting insights into the networks of co-evolving contacts that form in different conformational states, and different why certain sugar porters are or are …
Reviewer #1 (Public Review):
This manuscript harnesses recent advances in co-evolution based modeling and computational approaches to provide molecular details about the transport cycles and mechanisms of an entire family of transporters, the sugar porters. The authors evaluate the validity of their approach in a number of ways, including comparison to structurally characterized proteins/states excluded from the training set, comparison to the GLUT5 transport free energy landscape determine through conventional enhanced MD methods in a companion paper, and a global evaluation of RMSDs between models. Based on these structural models, the authors are able to generate a number of interesting insights into the networks of co-evolving contacts that form in different conformational states, and different why certain sugar porters are or are not proton-coupled.
-
Reviewer #2 (Public Review):
Transporters cycle between several conformational states; however, developing a unifying cycle for a single transporter is often difficult, as different homologs are often used to experimentally determine the structures of different conformations. The manuscript of Mitrovic et al. is a clever and inspiring combination of computational methods to reconstruct the transport cycle and free-energy landscape of a single sugar transporter. Using co-evolution and machine learning, the authors extracted state-specific residue contacts, many of which were previously unobserved, and potentially describe subtle yet important structural features. Using these contacts, they bias AlphaFold2 structure determination and MD simulations to accurately predict any conformation. These structures combined with enhanced sampling …
Reviewer #2 (Public Review):
Transporters cycle between several conformational states; however, developing a unifying cycle for a single transporter is often difficult, as different homologs are often used to experimentally determine the structures of different conformations. The manuscript of Mitrovic et al. is a clever and inspiring combination of computational methods to reconstruct the transport cycle and free-energy landscape of a single sugar transporter. Using co-evolution and machine learning, the authors extracted state-specific residue contacts, many of which were previously unobserved, and potentially describe subtle yet important structural features. Using these contacts, they bias AlphaFold2 structure determination and MD simulations to accurately predict any conformation. These structures combined with enhanced sampling methods facilitate the inference of free-energy landscapes of the transport cycle. Notably, this work continues to push the limits of using and interpreting AlphaFold2 past static snapshots of highly dynamic proteins. This combination of techniques represents the forefront of structural biology, clearly demonstrating how static protein structures can be leveraged using bioinformatic and computational techniques to understand the biophysical mechanisms of proteins. Though the methodology is technically and theoretically exciting, it is as of yet unclear if this represents a substantial enough improvement over existing techniques for wider adoption. Nevertheless, this work represents an innovative combination of existing approaches to create a cohesive framework of the sugar transport cycle, and the authors provide detailed methods and supplementary information to recreate these approaches in other transporter families.
-
Reviewer #3 (Public Review):
This work proposes a novel computational methodology that, using available structures of homologous proteins in different structural states, evolutionary couplings and machine-learning protocols, allows to predict structural states of a membrane transporter during the transport cycle. The core of the methodology is to use convolutional neural networks to distinguish state specific evolutionary contacts and drive alphafold2 models into a specific state based on the predicted contacts (using rosettaMP and short MD relaxation). The authors then derived the free energy landscape of the alternating access transition of GLUT5 (in absence of substrate) from enhanced sampling simulations biased along variables based on the previously mentioned contacts. The variables are constructed using a machine learning approach …
Reviewer #3 (Public Review):
This work proposes a novel computational methodology that, using available structures of homologous proteins in different structural states, evolutionary couplings and machine-learning protocols, allows to predict structural states of a membrane transporter during the transport cycle. The core of the methodology is to use convolutional neural networks to distinguish state specific evolutionary contacts and drive alphafold2 models into a specific state based on the predicted contacts (using rosettaMP and short MD relaxation). The authors then derived the free energy landscape of the alternating access transition of GLUT5 (in absence of substrate) from enhanced sampling simulations biased along variables based on the previously mentioned contacts. The variables are constructed using a machine learning approach that allows distinguishing different structural states.
The advantage of this approach is that it uses a combination of advanced modeling and innovative computational techniques that might help the structural characterization of the alternating access cycle of membrane transporters. An important innovation is the use of machine learning methods that, based on previous structural information, allow to construct collective variables for free energy calculations in an objective, data-driven manner.
The results of the modellng part of the work are encouraging but could benefit from using more specific descriptors that better distinguish structural differences between states.
An important weakness of this work is that there are critical flaws in the simulation analysis. Another weakness is that the different free energy landscapes calculated do not appear strongly consistent to each other, which suggests the presence of significant errors in the calculations that are not discussed. An additional important point is that a quantitative assessment of the quality of the models used in the simulations is currently lacking and this could affect the reliability of the simulation results. In this regard, previous systematic studies (Proteins 2012; 80:2071-2079) have shown that small imperfections in the predicted models (such as in backbone and side chains conformations) could lead to simulations that drift away from the initial structure in the multi microseconds time domain.
-
-