Modeling Flexible Protein Structure With AlphaFold2 and Crosslinking Mass Spectrometry

Abstract

•Multiple conformations of proteins can be predicted with AlphaFold2.•Crosslinks and monolinks can be used to select the relevant protein conformation.•Tools for data analysis are available at https://gitlab.com/topf-lab/xlms-tools. We propose a pipeline that combines AlphaFold2 (AF2) and crosslinking mass spectrometry (XL-MS) to model the structure of proteins with multiple conformations. The pipeline consists of two main steps: ensemble generation using AF2 and conformer selection using XL-MS data. For conformer selection, we developed two scores—the monolink probability score (MP) and the crosslink probability score (XLP)—both of which are based on residue depth from the protein surface. We benchmarked MP and XLP on a large dataset of decoy protein structures and showed that our scores outperform previously developed scores. We then tested our methodology on three proteins having an open and closed conformation in the Protein Data Bank: Complement component 3 (C3), luciferase, and glutamine-binding periplasmic protein, first generating ensembles using AF2, which were then screened for the open and closed conformations using experimental XL-MS data. In five out of six cases, the most accurate model within the AF2 ensembles—or a conformation within 1 Å of this model—was identified using crosslinks, as assessed through the XLP score. In the remaining case, only the monolinks (assessed through the MP score) successfully identified the open conformation of glutamine-binding periplasmic protein, and these results were further improved by including the “occupancy” of the monolinks. This serves as a compelling proof-of-concept for the effectiveness of monolinks. In contrast, the AF2 assessment score was only able to identify the most accurate conformation in two out of six cases. Our results highlight the complementarity of AF2 with experimental methods like XL-MS, with the MP and XLP scores providing reliable metrics to assess the quality of the predicted models. The MP and XLP scoring functions mentioned above are available at https://gitlab.com/topf-lab/xlms-tools. We propose a pipeline that combines AlphaFold2 (AF2) and crosslinking mass spectrometry (XL-MS) to model the structure of proteins with multiple conformations. The pipeline consists of two main steps: ensemble generation using AF2 and conformer selection using XL-MS data. For conformer selection, we developed two scores—the monolink probability score (MP) and the crosslink probability score (XLP)—both of which are based on residue depth from the protein surface. We benchmarked MP and XLP on a large dataset of decoy protein structures and showed that our scores outperform previously developed scores. We then tested our methodology on three proteins having an open and closed conformation in the Protein Data Bank: Complement component 3 (C3), luciferase, and glutamine-binding periplasmic protein, first generating ensembles using AF2, which were then screened for the open and closed conformations using experimental XL-MS data. In five out of six cases, the most accurate model within the AF2 ensembles—or a conformation within 1 Å of this model—was identified using crosslinks, as assessed through the XLP score. In the remaining case, only the monolinks (assessed through the MP score) successfully identified the open conformation of glutamine-binding periplasmic protein, and these results were further improved by including the “occupancy” of the monolinks. This serves as a compelling proof-of-concept for the effectiveness of monolinks. In contrast, the AF2 assessment score was only able to identify the most accurate conformation in two out of six cases. Our results highlight the complementarity of AF2 with experimental methods like XL-MS, with the MP and XLP scores providing reliable metrics to assess the quality of the predicted models. The MP and XLP scoring functions mentioned above are available at https://gitlab.com/topf-lab/xlms-tools. AlphaFold2 (AF2) has revolutionized structural biology with unprecedented accuracy in protein structure prediction, even on sequences for which related structures are unavailable (1Jumper J. Evans R. Pritzel A. Green T. Figurnov M. Ronneberger O. et al.Highly accurate protein structure prediction with AlphaFold.Nature. 2021; 596: 583-589Crossref PubMed Scopus (13358) Google Scholar). AF2 has been used to exhaustively predict the structures of more than 200 million UniProt sequences (2Varadi M. Anyango S. Deshpande M. Nair S. Natassia C. Yordanova G. et al.AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models.Nucleic Acids Res. 2022; 50: D439-D444Crossref PubMed Scopus (2700) Google Scholar), which is both an invaluable resource to the structural biology community, as well as a test of AF2’s current structure prediction capacity. In particular, it has been shown that AF2 predicts ordered protein domain structures well but performs less well on proteins with predicted flexibility or disorder (3Perrakis A. Sixma T.K. AI revolutions in biology: the joys and perils of AlphaFold.EMBO Rep. 2021; 22: e54046Crossref PubMed Scopus (82) Google Scholar, 4Chakravarty D. Porter L.L. AlphaFold2 fails to predict protein fold switching.Protein Sci. 2022; 31: e4353Crossref PubMed Scopus (43) Google Scholar). A possible cause is that AF2 and other protein structure prediction approaches have been trained with the assumption that one protein sequence corresponds to one structure, something we know to be untrue for a wide variety of proteins, such as molecular switches which transition between two different conformations as part of their function (4Chakravarty D. Porter L.L. AlphaFold2 fails to predict protein fold switching.Protein Sci. 2022; 31: e4353Crossref PubMed Scopus (43) Google Scholar, 5Ha J.-H. Loh S.N. Protein conformational switches: from nature to design.Chemistry. 2012; 18: 7984-7999Crossref PubMed Scopus (98) Google Scholar), as well as proteins that are either fully or partially disordered. The issue is compounded by the fact that the Protein Data Bank (PDB) (6Burley S.K. Bhikadiya C. Bi C. Bittrich S. Chao H. Chen L. et Protein Data Bank of structures one million structure of proteins from Acids Res. PubMed Scopus Google protein structure used to structure prediction proteins that are are to and are T. disorder in the Protein Data PubMed Scopus Google Scholar). In proteins that have multiple possible from domain to other have of their conformations in the AF2 for structure and to of a In the of molecular switches that have that AF2’s predicted is of between but experimental methods that can protein in a For in crosslinking mass spectrometry proteins are at with such as the most used and which both with to on the protein J. The of a spectrometry and of proteins and PubMed Scopus Google Scholar). The at only one to a or both of the to different of the protein, to a crosslink of the by mass spectrometry of the protein are on the surface. can the between the two on the protein are the has a Å for and J. mass methods and in molecular and PubMed Scopus Google Scholar). The by a has been as the of the and the of the two the of protein and protein by mass PubMed Scopus Google Scholar), with a of 3 Å to for flexibility and model in a of Å S. A. from crosslinking mass a molecular to Sci. PubMed Scopus Google Scholar). have been used to protein structure by either the that can be by the is et protein fold by using experimental from and mass Sci. S. A. PubMed Scopus Google Scholar, A. L. C. S. T. et of the by and mass J. PubMed Scopus Google Scholar, M. of by mass and Sci. S. A. 2012; PubMed Scopus Google Scholar, C. C. A. analysis of by spectrometry conformational Rep. PubMed Scopus Google Scholar), or by more scoring functions that are based on the The more to is a of and was shown to than in protein structures from a of A. L. R. and in PubMed Scopus Google Scholar, A. A. G. R. L. molecular with PubMed Scopus Google Scholar, T. J. S. Protein structure prediction by crosslinking of the of the crosslinking PubMed Scopus Google Scholar, J. J. M. The of and in proteins with from crosslinking mass PubMed Scopus Google Scholar, C. protein in the of PubMed Scopus Google Scholar). The of methods like XL-MS that can be in is that For this a experimental pipeline proteins is to both structural from for with in such as to flexibility and XL-MS or to that be in the In the of of proteins have been used in with in in to model the structure of with the of protein structure prediction methods such as AF2, this pipeline be on proteins that have sequence to structures in the or proteins for which only of the conformations are In this we XL-MS data can be used to structure from AF2, on proteins has previously been shown that AF2 a variety of conformations using a than depth D. D. J. conformational of and with 2022; PubMed Scopus Google Scholar). We this to an ensemble of conformations for a of proteins that have an and a conformation in the We screened these AF2 using experimental XL-MS with two functions we have developed to score crosslink and monolink that of the of between and the depth of the residue from the of the We that our scores are able to select for that are the open and closed conformations from the AF2 that AF2 was able to predict these conformations. we a in which monolinks are more than in the relevant conformer from an we propose a pipeline to model proteins by first generating an ensemble with AF2, then the conformer in a experimental using and monolinks from We two scoring functions to structures from a of decoy protein based on of and monolinks from XL-MS and and monolinks that were to structures were first from the J. A. A. M. for analysis and of data in the of PubMed Scopus Google Scholar). A of structures were to be with either or The of both and were in the structures using the which is a for generating protein as well as to this using D. by PubMed Google Scholar, D. H. Protein depth and the for accuracy of protein fold PubMed Scopus Google Scholar). The were used to The residue by were to from which is a more accurate but that the protein structure in a in a of for the of residue from the R. a to depth and predict in Acids Res. PubMed Scopus Google Scholar). the we that the was able to to a of residue mentioned in this were with the The depth of is shown in with the depth of the structures In the between of were with the of shown in The depth were with a the was with a function using R. M. T. D. et for in The monolink probability (MP) score for a monolink in a structure is as with the first to the between probability and A at a depth in the structure and the a for monolinks that are a depth of more than Å in the structure, with the Å from the residue The MP score of a model is the MP score of monolinks in the The of monolink can be we are only the or of but can In particular, we as is the and is the The crosslink probability score for a crosslink and in a structure is then with the first to the of the by a crosslink by the monolink probability MP at both and The is a for to or the is a for either residue a depth of as in the MP score. A of Å was as this is the for and previously A. L. C. S. T. et of the by and mass J. PubMed Scopus Google Scholar, J. J. M. The of and in proteins with from crosslinking mass PubMed Scopus Google Scholar). The for the were to for and for residue depth to the MP XLP is in the The of crosslink can be we are only the or of crosslinks, but can the of the of the monolink M. S. M. from and monolinks in the of protein PubMed Scopus Google and crosslink J. J. M. The of and in proteins with from crosslinking mass PubMed Scopus Google scores we previously both MP and XLP were to from to with a score of 1 that the structure the probability that the monolink or crosslink can be such a that the of the scores of a model is of the of other predicted and on the assumption that the are that the ensemble the in other methods of score such as the of a The MP and XLP scores were on the structure decoy dataset a of 200 proteins from the with structural with from the structure from to Å H. generation of and protein structure PubMed Scopus Google Scholar). of and monolinks were at different using the structures in the to our M. S. M. from and monolinks in the of protein PubMed Scopus Google Scholar), we monolinks as in the structure with a depth of Å or as by R. a to depth and predict in Acids Res. PubMed Scopus Google Scholar). were as of in the structure with a depth of at most and with a of at most from the structure with J. J. M. The of and in proteins with from crosslinking mass PubMed Scopus Google Scholar). We a of the of from possible crosslinking from to in of possible crosslinking were to that in XL-MS we monolinks in this by possible monolinks at a of of to the XLP to crosslinks, were in from to at crosslink the MP and XLP scores were for structure, using and and the MP and XLP scores were with a and score for model in the is as and are the is the between and is the from Å from the crosslink J. J. M. The of and in proteins with from crosslinking mass PubMed Scopus Google Scholar). A scoring function to structures that the structure and can structures from less accurate assess we the between the scores and the from the In we to identify structures by the the The the between the of and at different scoring In our case, were as scoring above the were scoring above the J. function for assessment of protein structure PubMed Scopus Google Scholar). were between XLP and we the and at using the test in R. M. T. D. et for in Scholar). We a test dataset of three proteins with experimental XL-MS data from the protein in the dataset at one and one conformation in the which were based on their of R. L. A. M. Structure of by spectrometry and PubMed Scopus Google Scholar, C. et protein structures from PubMed Scopus Google Scholar, T. G. L. A. et for conformational of protein by mass PubMed Scopus Google Scholar). of the and glutamine-binding periplasmic protein of a is a protein of two one of which a large conformational of the dataset are in For or monolinks were for the open and monolinks using our from the we that has one that is only in the open test this is to open was and for XL-MS as proteins with structures and experimental XL-MS between conformations open closed XL-MS component A. A. Structure of conformational that PubMed Scopus Google A. et of component the function and of PubMed Scopus Google and R. L. A. M. Structure of by spectrometry and PubMed Scopus Google structure of on a of PubMed Google structure of in a conformation a domain 2012; PubMed Scopus Google and T. G. L. A. et for conformational of protein by mass PubMed Scopus Google periplasmic protein J. The structure of glutamine-binding protein from PubMed Scopus Google J. The structure of glutamine-binding protein with at A with other PubMed Scopus Google C. et protein structures from PubMed Scopus Google monolinks this Protein Data of XL-MS, crosslinking mass in a Protein Data of XL-MS, crosslinking mass a for the of was from and were in at to an by of for at were in with and was from in a using an was a and with A was in a for of was to a of M. was in and through of were and in and proteins were by to and by was at for to using an to in For was to was to was to were then at for with The with and with were were with and at for using with in for by a with was then at a mass of and at The was then by of to a of were in at a of with a and a 3 was to a was at and was to A and were with an from to was A mass was used for this The mass was using was by the of a to a a and was by the and the and was to was to mass was at a of with a mass of to and a of were with a analysis was to a of was to were with a using the mass and selection was For mass analysis based was with a of was to mass was to was to and was to For two experimental were with experimental the in were in three of both This and within the experimental Data analysis was using with the used to the from were were a the sequence and 200 protein sequences from UniProt was used with monolink and monolink that for is to mass was to and mass was to score for was to to or than For was using For used for this was the within which the was the mass for was from and and were from the The monolinks and are in which the that was predicted to be in the open conformation but in the closed conformation of in and protein has been on dataset ensembles were predicted for and using a of (1Jumper J. Evans R. Pritzel A. Green T. Figurnov M. Ronneberger O. et al.Highly accurate protein structure prediction with AlphaFold.Nature. 2021; 596: 583-589Crossref PubMed Scopus (13358) Google Scholar), which the in a in predicted than the D. D. J. conformational of and with 2022; PubMed Scopus Google Scholar). of and were used for AF2 prediction for a of 200 predicted For which consists of two R. M. Pritzel A. A. Green T. et prediction with 2022; PubMed Scopus Google was to model the In this case, were predicted in models. depth was for as the was for the was to a to the first related structure, to structural function D. S.K. structural of the Protein Data PubMed Scopus Google Scholar). The AF2 were based on their predicted an of in The of scores for protein in the test dataset can be in with the score used to from further XLP and MP scores were for AF2 model and The AF2 were by XLP and MP and the for protein and conformation were using et structure for and Sci. 2021; PubMed Scopus Google Scholar). a to the of by AF2, we structural ensembles for and by the structures from the open to closed This was using the function in which conformations by The of were to using with the T. R. S. et molecular through from to Scopus Google Scholar, et generation for the of proteins, and Scopus Google Scholar). the XLP score was to these ensembles using from both the open and closed conformations. The were using for further We first the of our model assessment MP and using monolink and crosslink by the accuracy of the scores and selection of score on the the of MP and XLP to the and crosslink score scoring functions improved accuracy and structure selection with This is as more or monolinks the of possible conformations. XLP in both accuracy and selection in the crosslink depth from which only the in to both XLP and the of depth The monolink score MP showed the the assessed scores. it selection accuracy to XLP depth at of to an XLP score crosslink XLP a of of the are XLP at to We AF2 to predict ensembles of for luciferase, and using as in the the AF2 from open and closed conformations. were for for the closed conformation of and the open conformation of depth for and in an in model to AF2 only the closed conformation was predicted was that the for and a between the open and closed conformations This that these ensembles a of including and as well as between these for luciferase, the of model from the to the closed This the of This is further by the in which scores this for and that AF2 is less these models. We XLP and MP scores for AF2 model using experimental XL-MS data from either the open or closed conformation of luciferase, and In cases, the model with the XLP or MP score to a structure, a model was predicted by AF2, to a conformation within 1 Å of the most accurate model in the ensemble and The selection as by the from for the closed conformation of to for the other AF2 predicted a model In contrast, using the the different for the open conformation of and the closed conformation of for the open conformation of and for the closed conformation of XLP and MP scores more accurate of to as by for the closed conformation of that the XLP and MP scores the of the conformer by a of XL-MS data from an ensemble of structural models. In contrast, the of at only for one XLP and MP scores select for the open conformation from the AF2 ensembles of and The are shown as with three structures from the AF2 the XLP or MP the most accurate from and the model with the score. The XLP or MP scores and from the conformation are shown with the of the model in of accuracy shown structure in In case, the and or monolinks are are by monolinks are by residue depth AF2, glutamine-binding periplasmic Protein Data predicted crosslink For in our test we the monolinks and as the MP and XLP In the of open we data and the of monolink the as improved the selection and accuracy from and for the to and for the MP score This that including have a on the scores but this have to be benchmarked using a experimental XL-MS is that our scoring functions select for the model it was by this issue in the of and luciferase, we the of AF2 by an ensemble of the open and closed conformation and structures between these ensembles in of XLP in and selection for open and and open and and selection for closed conformation For cases, the scoring were and within the in of to the AF2 a in structural biology in but it the of experimental in the of We that AF2 can well in an pipeline for proteins by providing an ensemble of structures that can be using and from In scoring functions based on XL-MS we one that was more than a of et protein fold by using experimental from and mass Sci. S. A. PubMed Scopus Google Scholar, A. L. C. S. T. et of the by and mass J. PubMed Scopus Google Scholar, M. of by mass and Sci. S. A. 2012; PubMed Scopus Google Scholar, C. C. A. analysis of by spectrometry conformational Rep. PubMed Scopus Google Scholar), to structures that are possible within the large by and A. L. C. S. T. et of the by and mass J. PubMed Scopus Google Scholar, J. J. M. The of and in proteins with from crosslinking mass PubMed Scopus Google Scholar). the of scoring based on were to than A. A. G. R. L. molecular with PubMed Scopus Google Scholar, T. J. S. Protein structure prediction by crosslinking of the of the crosslinking PubMed Scopus Google Scholar, J. J. M. The of and in proteins with from crosslinking mass PubMed Scopus Google Scholar, C. protein in the of PubMed Scopus Google Scholar). is more as it the possible of the the protein only a This results in a in using of to our using crosslinks, the of the score can be and even by our scoring function the for more and the of using XLP to assess large conformational results are with our that that including can the of an scoring function to the of the scoring function J. J. M. The of and in proteins with from crosslinking mass PubMed Scopus Google Scholar, M. protein using from crosslinking mass PubMed Scopus Google Scholar). of in our scoring function we used residue in improved of XLP at crosslink of or in the crosslink are experimental crosslink are to be in this A. A. G. R. L. molecular with PubMed Scopus Google Scholar, J. J. M. The of and in proteins with from crosslinking mass PubMed Scopus Google Scholar, M. protein using from crosslinking mass PubMed Scopus Google Scholar, A. D. R. A. T. et of a protein by and mass 2012; PubMed Scopus Google Scholar, C. et structural analysis of protein crosslinking PubMed Scopus Google Scholar), in the in in crosslinking O. AlphaFold2 model experimental test using mass spectrometry of PubMed Scopus Google Scholar, A. C. R. S. et in by structural PubMed Scopus Google Scholar). We showed that the XLP score even to of the are This to is a is conformational in the experimental such that from more than one conformation are on the it is most to that as as the crosslink the XLP score can be used to select the conformation the to the The XLP score was shown to between different conformations of the proteins in our test dataset using crosslinks, with the scoring model in either the most accurate one or within 1 Å of the most accurate We and benchmarked a scoring function based on the MP which is a of the that an at a depth from the protein be with the crosslink we have that monolinks less than crosslinks, can be used to select for structures from a of at of at M. S. M. from and monolinks in the of protein PubMed Scopus Google Scholar). Our current has these the MP score was to select for the open conformation of as were to this This is to the of such that possible crosslinking can be by in open or closed was in this is the residue depth that previously of the protein are to it is in the open even one residue was to the open conformation of from the of the conformational ensemble and the of this in more monolinks be residue depth data can be from such as C. S. et for and mass spectrometry PubMed Scopus Google Scholar), but one can this in XL-MS in the of which are used for data We this of the of monolinks to monolink data more and to in available such as the J. C. D. S. S. et in a for mass Acids Res. 2022; 50: PubMed Scopus Google Scholar). We tested the of data to our scoring function by a in to monolink in open in a in both and selection to the monolinks were it be more accurate to model and crosslinking in this it to be this to a experimental XL-MS it is a that We have both the and of that crosslink is to the accuracy and has been previously shown that the of in the protein their A. C. R. S. et in by structural PubMed Scopus Google Scholar, A. T. A. O. M. et protein structures by mass and PubMed Scopus Google Scholar). For both of a crosslink a of an protein, then it between different conformations of the This can be by crosslink with the of crosslinking with different to the of in an XL-MS to the of XL-MS based scoring functions is to providing of T. J. S. Protein structure prediction by crosslinking of the of the crosslinking PubMed Scopus Google Scholar, L. C. A. mass spectrometry for protein conformations and for 2022; PubMed Scopus Google Scholar). This is to the in conformation in a protein like in our In particular, have been to of the such that have been shown to their and be S. structures of large protein using PubMed Scopus Google Scholar). the of to of by further to the these a be to the to as well as more data and to to the of a of or the crosslink out of in a Å in to providing in be that our scoring function is based on of this have to be for different crosslinking or further the to experimental XL-MS data in we have the of the XLP and MP scores on a large dataset at crosslink and monolink and the of the XLP score in the of of We showed that the scores are able to select or using experimental crosslink and monolink In our current has two main the XLP and MP scores are a of model quality within an it is to a score that has both and in the XLP score to in the model of closed conformation to be the to to this model in for other proteins like closed of the XLP scores are above This that other such as to protein a in which can be as a XLP or MP score

Topics & Concepts