Ultrafast end-to-end protein structure prediction enables high-throughput exploration of uncharacterised proteins

DMPfold2 produces structural models of comparable or greater accuracy than DMPfold1.

“When supplemented with an approximate precision matrix, the learned representation can be used to produce structural models of comparable or greater accuracy as compared to our original DMPfold method, while requiring less than a second to produce a typical model.”

• cites [9] sourceNot checked yet

  Not checked in the free run.

DMPfold2 generated over 1.3 million models of uncharacterised proteins in the BFD.

“we generated over 1.3 million models of uncharacterised proteins in the BFD”

Accurate structure prediction is possible in very short timeframes (a few hundred milliseconds).

“we show that accurate structure prediction is also possible in very short timeframes (a few hundred milliseconds).”

During training with a rigid body coordinate loss, the final network is able to learn to automatically transform the output coordinates for any structures mirrored by the multidimensional scaling process, without manual intervention.

“During training with a rigid body coordinate loss, as opposed to a distance-based loss, this final network is able to learn to automatically transform the output coordinates for any structures mirrored by the multidimensional scaling process, without manual intervention.”

DMPfold2 requires less than a second to produce a typical model.

“while requiring less than a second to produce a typical model”

Updated models were provided for over 5,000 Pfam families studied in the original DMPfold paper.

“We also provide updated models for over 5,000 Pfam families studied in the original DMPfold paper.”

Models were generated for 5,193 of the 5,214 Pfam families using DMPfold2 and the same alignments as DMPfold1.

“Models were generated for 5,193 of these families, using DMPfold2 and the same alignments as those used for DMPfold1.”

A number of confident BFD models do not match structures in the PDB.

“we do find a number of confident models that do not match structures in the PDB”

Three DMPfold2 models match CATH superfamilies that are the only superfamily in the respective CATH topology type.

“Three examples are shown in Figure 4b of models that match CATH superfamilies that are the only superfamily in the respective CATH topology type.”

1,333,242 models were generated from the BFD alignments by running one iteration of DMPfold2.

“In total, we generated 1,333,242 models from the BFD alignments by running one iteration of DMPfold2.”

The best performance is obtained when using the learned MSA representation alongside an on-the-fly computed precision matrix.

“Although the best performance is obtained when using the learned MSA representation alongside an on-the-fly computed precision matrix, the work in this study opens up the possibility of using just the MSA representation itself as the sole input to the ResNet for predicting protein structure”

From the remaining 1,284,617 models, 33,548 (2.61%) had a network-predicted confidence score of 0.4 or greater.

“Next, from the remaining 1,284,617 models, 33,548 (2.61%) were found to have a network-predicted confidence score of 0.4 or greater.”

The vast majority of the 1.3 million BFD models are predicted to have a confidence score of less than 0.4.

“Although the vast majority of the models are predicted to have a confidence score of less than 0.4”

Embedding whole MSAs into a linear representation using standard language modelling approaches can enable the efficient generation of 3-D structures directly from sequence data.

“Overall, we have demonstrated that the idea of embedding whole MSAs into a linear representation using standard language modelling approaches can enable the efficient generation of 3-D structures directly from sequence data.”

It is possible to directly process an MSA into a learned representation that is effective for predicting structural features in proteins.

“Here, we show that it is possible to directly process a MSA into a learned representation that is effective for predicting structural features in proteins.”

A number of targets were identified as having a potentially novel fold, based on comparison to the PDB with TM-align.

“A number of targets were identified as having a potentially novel fold, based on comparison to the PDB with TM-align.”

Around 86% of the modelled BFD subsequences were fewer than 300 residues in length.

“Given that around 86% of these subsequences were fewer than 300 residues in length”

Approximately 60% of the generated Pfam models were at a confidence score of 0.4 or greater.

“Approximately 60% of the models generated were at a confidence score of 0.4 or greater”

A larger fraction of the iteratively refined models are found to have a hit to the PDB with a TM-score of 0.5 or greater.

“Concomitantly, a larger fraction of the models are found to have a hit to the PDB with a TM-score of 0.5 or greater.”

Of the 33,548 modelled BFD sequences used as HHsearch queries against PDB70, 31% have at least one hit at an E-value threshold of 1.0.

“We used the BFD sub-MSAs for the 33,548 modelled sequences as queries to search the PDB70 database using HHsearch.31% have at least one hit in the PDB70 database at an E-value threshold of 1.0.”

DMPfold2 produces more accurate structure models for CASP13 FM and FM/TBM domains than DMPfold1.

“Comparison against models generated using DMPfold1 (Figure 2 and Table S1) shows that DMPfold2 produces more accurate structure models for these domains.”

In DMPfold2, the use of two gated recurrent networks to embed the input MSA is effective.

“In DMPfold2, the use of two gated recurrent networks to embed the input MSA is clearly effective, but was something of a design compromise in that RNNs are relatively fast to train, but have known limitations in terms of the limits of modelling long-range dependencies in sequences.”

83% of the BFD queries have a DMPfold2 model that matches a structure in the PDB.

“At an E-value threshold of 0.001, only 12% of the queries retrieve significant hits despite 83% having a model that matches a structure in the PDB.”

DMPfold2 outperforms DMPfold1 in all difficulty classes in terms of mean TM-score for the 255 Pfam families.

“From our analysis, it can be seen that DMPfold2 outperforms DMPfold1 in all difficulty classes, in terms of the mean TM-score achieved for each category (Figure 2c).”

Structural models built using DMPfold2 are more accurate on average than those from DMPfold1.

“Structural models built using DMPfold2 are more accurate on average than those from DMPfold1.”

• Also stated in: § 4. Discussion · plain

  “Overall, DMPfold2 end-to-end achieves comparable or greater model accuracy than DMPfold1.”

23,495 BFD models were assigned to 2,394 CATH superfamilies on the basis of structural similarity.

“Overall, we were able to assign 23,495 models to 2,394 CATH superfamilies on the basis of structural similarity (Table S2)”

Five CASP13 domains could be folded correctly (TM-score > 0.5) by the end-to-end procedure but not by DMPfold1.

“Five CASP13 domains could be folded correctly (TM-score > 0.5) by the end-to-end procedure but not by DMPfold1.”

Applying DMPfold2 to sequence families from the BFD, even with a relatively coarse-grained approach, predicted structures with plausible topologies for tens of thousands of uncharacterised protein domains, including some with potentially novel folds.

“We showcased this approach by applying DMPfold2 to sequence families from the BFD, and even using a relatively coarse-grained approach, we were able to predict structures with plausible topologies for tens of thousands of uncharacterised protein domains, including some with potentially novel folds.”

At an E-value threshold of 0.001, only 12% of the BFD queries retrieve significant hits in PDB70.

“At an E-value threshold of 0.001, only 12% of the queries retrieve significant hits despite 83% having a model that matches a structure in the PDB.”

The DMPfold2 modelling process can be used to identify fold-level similarity to existing PDB entries where such similarity would otherwise have been missed by a purely sequence-based comparison.

“We also showed that the modelling process can be used to identify fold-level similarity to existing PDB entries, where such similarity would otherwise have been missed by a purely sequence-based comparison.”

DMPfold2 is able to fold 30 CASP13 domains to a TM-score of 0.5 or greater, as compared to 26 for DMPfold1.

“DMPfold2 is able to fold 30 domains to a TM-score of 0.5 or greater, as compared to 26 for DMPfold1.”

DMPfold2 produces models in much less time than RoseTTAFold and AlphaFold2.

“In comparison to these methods, DMPfold2 produces models in much less time (Figure S5), making it suitable for use at a very large scale.”

The selected models revealed putative novel protein folds.

“revealing putative novel protein folds”

Covarying residues in proteins are frequently found to be close together in the tertiary structure.

“Analysis of amino acid residue covariation in deep multiple sequence alignments (MSAs) has revealed that covarying residues in proteins are frequently found to be close together in the tertiary structure.”

• cites [1] sourceNot checked yet

  Not checked in the free run.

• cites [2] sourceNot checked yet

  Not checked in the free run.

• cites [3] sourceNot checked yet

  Not checked in the free run.

Iteration of the prediction process can make significant improvements in model quality.

“As we first showed with DMPfold1 (9), iteration of the prediction process, where the final model is fed back in as an input to the network, can make significant improvements in model quality.”

• cites [9] sourceNot checked yet

  Not checked in the free run.

A system of recurrent neural network (RNN) layers can be used to process MSAs of arbitrary size.

“Recently, we demonstrated that a system of recurrent neural network (RNN) layers can be used to process MSAs of arbitrary size (15).”

• cites [15] S. M. Kandathil , J. G. Greener , A. M. Lau , D. T. Jones , Deep learning-based prediction of protein structure using learned representations of multiple sequence alignments. bioRxiv, 2020.11.27.401232 (2020).Not checked yet

  Not checked in the free run.

DMPfold2 is fast enough to run at unprecedented scale.

“DMPfold2 is thus fast enough to run at unprecedented scale”

• AssumptionRests on an assumption: Existing efforts to model large sets of protein sequences have been limited by the computational cost of running more than a few thousand predictions in a reasonable time frame.

For MSAs with longer sequences or with many more homologous sequences, even gated RNNs will start to become ineffective.

“For MSAs with longer sequences or with many more homologous sequences, even gated RNNs will start to become ineffective.”

The large fraction of models with a similar structure in the PDB despite most sequences with templates being eliminated shows the value in de novo protein structure prediction for mapping sequences to the existing fold space and inferring homology.

“The large fraction of models with a similar structure in the PDB despite most sequences with templates being eliminated shows the value in de novo protein structure prediction for mapping sequences to the existing fold space and inferring homology.”

Most of the modelled BFD subsequences correspond to single domains.

“most will correspond to single domains”

These confident BFD models likely represent novel folds.

“and likely represent novel folds”

The lower performance of DMPfold2 on easy and trivial targets can be attributed to the fact that DMPfold2 does not make explicit use of template information, whereas C-I-TASSER does.

“The lower performance of our method on easy and trivial targets can be attributed to the fact that DMPfold2 does not make explicit use of template information, whereas C-I-TASSER does.”

Tertiary structure models are valuable for retrieving structural templates.

“This illustrates the value of tertiary structure models in retrieving structural templates: assuming that the DMPfold2 models are correct at the fold level, the TM-align search against the PDB reveals that these models are structurally similar to existing PDB entries, whereas an HMM-HMM comparison would miss these templates.”

• AssumptionThe authors state this as an assumption: The DMPfold2 models are correct at the fold level.

In practice, gated RNNs model dependencies beyond a window of a few hundred time steps poorly.

“In theory, gated RNNs are able to avoid the problems of vanishing gradients when modelling long sequences, but in practice, dependencies beyond a window of a few hundred time steps are poorly modelled.”

It is possible to speculate about the function of modelled sequences by matching the model to structures of known function.

“In some cases it is possible to speculate about the function of modelled sequences by matching the model to structures of known function (47).”

• cites [47] sourceNot checked yet

  Not checked in the free run.

This approach enables structural models to be built on a far larger scale than was previously possible.

“This approach enables structural models to be built on a far larger scale than was previously possible.”

For the modelled sequences matched to a CATH superfamily that is the only superfamily in its topology type, it is likely that the sequences share a function with other members of the matched superfamily.

“In these cases it is likely that the sequences share a function with other members of the matched superfamily.”

The ease and speed with which a rough candidate model can be generated for a query sequence has the potential to accelerate and improve the sensitivity of structural template searches.

“Of course, this principle has been well understood for some time, and is the basis of protein threading (45,46), but the ease and speed with which a rough candidate model can be generated for a query sequence has the potential to accelerate and improve the sensitivity of structural template searches.”

For some fraction of the BFD targets, domain boundaries may not have been correctly parsed, as the subsequence filtering procedures were based only on sequence profile HMM similarity.

“Additionally, for some fraction of these targets, domain boundaries may not have been correctly parsed, as our subsequence filtering procedures were based only on sequence profile HMM similarity.”

Model quality can usually be improved by generating a fresh MSA for the subsequence of interest, rather than merely extracting a region of the original BFD MSA for a longer sequence family.

“These results show the benefit of iterative structure refinement, but also demonstrate that model quality can usually be improved by generating a fresh MSA for the subsequence of interest, rather than merely extracting a region of the original BFD MSA for a longer sequence family.”

Most targets in the FM category at CASP do match a known fold, even though the template is not found with sequence-based searching.

“Indeed most targets in the FM category at CASP do match a known fold, even though the template is not found with sequence-based searching (44).”

• cites [44] 2019, CASP13 target classification into tertiary structure prediction categories, Proteins sourceNot checked yet

  Not checked in the free run.

The assignment of 23,495 models to CATH superfamilies suggests that the predictions are correct at the fold level.

“suggesting that the predictions are correct at the fold level.”

Confidence values are output per residue, which allows confidently modelled segments of the model to be distinguished from less confidently modelled segments.

“Confidence values are output per residue, which allows confidently modelled segments of the model to be distinguished from less confidently modelled segments.”