Large-scale structure-informed multiple sequence alignment of proteins with SIMSApiper

Abstract

Summary

SIMSApiper is a Nextflow pipeline that creates reliable, structure-informed MSAs of thousands of protein sequences faster than standard structure-based alignment methods. Structural information can be provided by the user or collected by the pipeline from online resources. Parallelization with sequence identity-based subsets can be activated to significantly speed up the alignment process. Finally, the number of gaps in the final alignment can be reduced by leveraging the position of conserved secondary structure elements.

Availability and implementation

The pipeline is implemented using Nextflow, Python3, and Bash. It is publicly available on github.com/Bio2Byte/simsapiper.

1 Introduction

Proteins execute a myriad of functions based on combinations of 20 different types of amino acids arranged in distinct sequences, which can adopt a variety of 3D shapes. The investigation of protein function, and studies of how they evolved, rely largely on the analysis of Multiple Sequence Alignments (MSA) of proteins with a common evolutionary origin (O’Sullivan et al. 2004, Carpentier and Chomilier 2019). The evolutionary selection pressure on a protein is indeed driven by its ability to function correctly, with even drastic changes in a protein’s amino acid sequence tolerated as long as its overall behavior and shape maintains its function. Sequence-based alignment methods, while being very fast, struggle with the alignment of protein families with high structural and functional conservation but low sequence identity (SI) (<30%) (O’Sullivan et al. 2004, Carpentier and Chomilier 2019, Rajapaksa et al. 2023), such as RNA Recognition Motifs (RRMs) (Roca-Martínez et al. 2023), membrane proteins like G-Protein Coupled Receptors (Zhou et al. 2019) or designed proteins with no evolutionary connection (Figueroa et al. 2013, Huang et al. 2016). Moreover, sequence-based algorithms are optimized for single globular protein domains and lose accuracy when confronted with larger proteins with multiple domains (Santus et al. 2023). For these reasons, structure-informed MSA methods have provided the “gold standard” MSAs that serve to evaluate sequence-based alignment algorithms (Carpentier and Chomilier 2019).

The main challenge for structure-informed MSA methods has been a lack of available protein structures (Chatzou et al. 2016), which is now resolved through the emergence of accurate AI-based prediction of protein structure from sequence (Jumper et al. 2021, Lin et al. 2023): over 200 million protein structure models are now available through the AlphaFold Protein Structure Database (AF2 DB) (Varadi et al. 2022). Methods such as ColabFold (Mirdita et al. 2022) and ESMFold (ESMF) (Lin et al. 2023) can also provide protein structure models of reasonable accuracy within minutes. Even when such predicted protein structures are of reduced accuracy, they can significantly improve structure-informed MSAs (Baltzis et al. 2022, Rajapaksa et al. 2023).

For the structural alignment of larger datasets, several challenges are present. Access to structural information and/or the prediction of models can be time-intensive. Moreover, the computational resources required to find the optimal alignment grow exponentially with the number of proteins to align (Rubio-Largo et al. 2018, Lladós et al. 2021, Santus et al. 2023). For the efficient study of protein families, this process of structure-informed MSAs requires streamlining.

We introduce SIMSApiper, an automated pipeline to create structure-informed MSAs of protein sequences not limited by the size of the dataset nor by the SI between the protein sequences. SIMSApiper combines existing tricks and tools of current MSA workflows into a single framework implemented in Nextflow (Di Tommaso et al. 2017) to maintain reproducibility, scalability and convenience across users and platforms. The protein sequences to be aligned are divided into subsets, either provided by the user or automatically created based on SI cutoffs. This significantly reduces compute times and allows the method to scale to thousands of protein sequences. The required protein structures can be provided by the user and complemented with models from the AF2 DB or ESMF. Each subset is aligned using T-Coffee’s tool 3DCoffee (O’Sullivan et al. 2004) (referenced to as T-Coffee hereafter) and the TMalign and SAP algorithms (Taylor 2000, Zhang and Skolnick 2005). These sub-MSAs are then merged into the final MSA using MAFFT (Katoh et al. 2019), with a final secondary-structure based refinement step reducing the number of gaps. SIMSApiper’s accuracy was tested on three datasets, a subset of HOMSTRAD, and the TIM-barrel and GroEL protein families. The quality of SIMSApiper’s and T-Coffee’s alignments are comparable on challenging datasets. However, SIMSApiper is faster than T-Coffee on larger datasets (e.g. 20 times faster on 350 TIM-barrels proteins).

2 Materials and methods

2.1 Computation

SIMSApiper is implemented in Nextflow, with dependencies provided in containers compatible with Docker (Merkel et al. 2014) and Singularity(Kurtzer et al. 2017), ensuring reproducibility (Conte et al. 2023) and ease of use. Nextflow enables parallelization, produces extensive log files and provides check points from which execution can be resumed, all properties that contribute to the decrease in compute time while optimizing the use of compute resources.

SIMSApiper requires only a sequence file to run. The input sequences can first be cleaned (step 1, shown in Supplementary Fig. S1) to reduce the dataset’s redundancy using CD-Hit (Li and Godzik 2006). We recommend a cutoff of at least 90% SI, with 70% SI likely better to reduce sampling bias for a given protein family. In Fig. 1 step 2, SIMSApiper matches user-provided structures to the sequences and automatically collect structure models for the remaining ones. For sequences labeled with their corresponding Uniprot ID (UniProt Consortium 2023), models are automatically collected from the AF2 DB. Sequences shorter than 400 residues are submitted to the ESMF online resources to generate a model. Finally, sequences not matched to a structure after this step are called structureless. In step 3, similar sequences, with a matched structure, are grouped together into subsets. These can be based on user-provided sequence files or generated automatically using CD-Hit with a SI cutoff as low as 20%. Sequences that are too different can be collated with small clusters, thereby creating a minimal total amount of subsets. Alternatively, they can be labeled as orphan. These and the structureless sequences are not aligned in the next step and will be integrated later. Each sequence subset along with the corresponding structures are submitted to T-Coffee (step 4) for alignment with TMalign and SAP. Subsets can be aligned in parallel if the required hardware is available. After aligning each subset, the subset MSAs are submitted to MAFFT (step 5) with the “merge” setting, which preserves each subset MSA’s shape and therefore retains the structure information. The orphan and structureless sequences are then individually aligned to the main alignment to generate one final alignment. As step 5 tends to dramatically increase the number of gaps in variable and hard to align regions such as loops, SIMSApiper provides the option to reduce gaps by squeezing the alignment towards DSSP defined conserved secondary structure categories (steps 6,7) such as $\beta$-strands or $\alpha$-helices (Kabsch and Sander 1983). This step is based on the principle that the (spatial) position of amino acids relative to conserved secondary structure elements is the most relevant, not their SI (Roca-Martínez et al. 2023).

Figure 1.

Simplified workflow of data handling and integrated tools in SIMSApiper pipeline.

2.2 Performance evaluation

The overall quality of the MSA is quantified using the Column Score (CS). For each column of the query MSA (i), a score of 1 is assigned when all residues are aligned as in a reference MSA, a score of 0 is given if any misalignment is present. The number of matching columns between (regions of) the MSAs is then divided by the total number of evaluated columns (m) to produce a final score CS (Equation 1).

$$CS=\frac{{\sum }_{i=1}^{m}scor{e}_i}{m}$$ (1)

2.3 Validation datasets

SIMSApiper was evaluated on three datasets:

1. The TIM-barrel protein family [InterPro accession: IPR000652 (Paysan-Lafosse et al. 2023)]: 663 sequences with SI $⩾$20% and highly conserved structure of 4 $\beta \alpha \beta \alpha$ anti-symmetrical units (Wierenga 2001).

2. Similarly to T-Coffee’s validation strategy (O’Sullivan et al. 2004), we selected the most demanding MSAs (SI <25% and a size >4 proteins) from the hand-curated HOMSTRAD reference database (Mizuguchi et al. 1998), (release date: 07/05/2022). This resulted in 51 MSAs (HOM51) described in detail in Supplementary Fig. S4.

3. The GroEL protein family: Conserved obligate chaperonin across archaea and prokaryotes with a structural homolog in eukaryotes (Ansari and Mande 2018). 50k sequences were collected from Uniprot KB with estimated SI $⩽30\%$.

2.4 Hardware

A detailed description of the used hardware to compute the SIMSApiper validation MSAs is available in Supplementary Table S1.

3 Results

3.1 Fast and accurate alignment of the TIM-barrel protein family

SIMSApiper automatically cleaned the InterPro TIM-barrels protein family using a >90% SI cutoff, resulting in a dataset of 379 sequences to align. It then grouped the sequences into 6 subsets using a SI cutoff of 30%. Each subset was aligned by SIMSApiper with T-Coffee and required 10 CPUs, 15 GB memory and ran between 1 and 25 min depending on its size (Supplementary Fig. S2). Nextflow enabled efficient resource use in this step by running jobs in parallel, resulting in a total runtime of 33 min. Concurrently, this would result in an approximate runtime of 2h12. If no subsets are created, equivalent to running directly T-Coffee on the 379 sequences, the total runtime increases to 9h48 and requires 45 GB memory instead (Supplementary Fig. S2). This significant increase in computational resources highlights how SIMSApiper can more efficiently accommodate larger datasets than T-Coffee alone, by clustering the input data. From the TIM-barrel SIMSApiper generated MSA, we extracted the 10 sequences that are also present in a manually curated, gold standard, structure-informed TIM-barrel MSA (Maes et al. 1999) (MSA_SIMSApiper and MSA_ref, respectively).

A comparison of MSA_SIMSApiper and MSA_ref using the CS metric (Supplementary Table S2), illustrates that SIMSApiper effectively aligns conserved structured regions (CS = 99,3%), with a lower score for the diverse and unstructured loops (CS = 91.3%). The overall CS (CS = 95.8%) shows SIMSApiper’s capability to generate a quality MSA despite the use of clustering to accommodate larger datasets (step 3). The CS also highlights the importance of the squeezing step (step 6,7), which helps to reduce the number of gaps in the MSA and improving its quality (CS from 85.2% to 91.3% in the loops). Running SIMSApiper without the squeezing mode is equivalent to running T-Coffee. By squeezing, SIMSApiper thus performs slightly better than T-Coffee for this dataset. SIMSApiper was further evaluated by comparing the alignment of two designed TIM-barrels, sTIM 11 (Huang et al. 2016), and Octarellin VI (Figueroa et al. 2013), to the 10 natural TIM-barrels mentioned above. Despite the low SI (<20%) and different 3D structure (>4Å) between the designed and natural proteins (Supplementary Fig. S3a–d), SIMSApiper identified similar secondary structure elements ($\beta$-sheets and $\alpha$-helices) and effectively aligned those with each other (Supplementary Fig. S3e).

3.2 Alignment of the challenging datasets from HOMSTRAD

When comparing the HOM51 MSAs with the equivalent SIMSApiper MSAs, a median CS of 80% is obtained for the conserved structured regions (Supplementary Fig. S5). SIMSApiper can thus successfully align datasets with low SI. As a similar overall median CS is obtained with or without squeezing (53% and 55%, respectively), the squeezing (step 7) does not impact the overall alignment quality in these examples. This is because of the way the HOM51 MSAs are constructed, with many gaps in the loop regions. Squeezing the loops of the obtained SIMSApiper MSAs towards their conserved regions is therefore in this case counter-effective. Nevertheless, because running SIMSApiper with the no squeezing setting is equivalent to running native T-Coffee (no pre-processing), we can conclude that T-Coffee also struggles with aligning these loop regions. In our opinion, because of the sequence and structural diversity of the loops, the alignment of the loops in the HOM51 MSAs is not more accurate than the alignment obtained with T-Coffee, and thus by extension SIMSApiper. Analyzing the CS of the unconserved regions is therefore of limited relevance here.

3.3 Alignment of large datasets: the GroEL protein family

We applied a 70% sequence similarity cutoff on the large GroEL dataset to reduce compute times, leaving 1600 diverse proteins to be aligned. Four sequences were excluded due to a high number of unresolved amino acids. For 400 sequences, no matching models could be found in the AF2 DB. As the average sequence length exceeds 400 residues for GroEL, we supplied the structural information with ESMF predictions run on the VSC Tier-2 general-purpose clusters provided by VUB-HPC. Assembling the complete structural dataset took 7 min to retrieve models from AF2 DB and 2.5 h to predict the remaining models locally. Thirty-eight subsets were generated and aligned in parallel. The alignment process took 45 min, including post-processing. Supplementary Figure S6a highlights the final MSAs well aligned secondary structure elements and Supplementary Figure S6b shows that SIMSApiper has maintained the conservation of experimentally confirmed residues for ATP binding (Xu et al. 1997, Koike-Takeshita et al. 2014), which are crucial for GroEL to perform its function.

4 Discussion

The rise of AI-driven protein fold predictors has greatly impacted the field of structural bioinformatics by providing accurate structure models for well-folded protein regions, with their accessibility ensured through both sequence- and structure-based methods [BLASTP (Camacho et al. 2009), Foldseek (Barrio-Hernandez et al. 2023)] on AF2 DB.

Recent reviews (Rajapaksa et al. 2023, Santus et al. 2023) highlight the need for a MSA pipeline that can leverage this new data for the study of protein families with low SI but conserved secondary structure. Such approaches will allow to pinpoint similarities and differences, enabling a deeper understanding of their function. SIMSApiper provides this functionality by combining user-provided structures and automatically retrieved models from AF2 DB or ESMF and by enabling the use of the alignment tool T-Coffee, known for generating high-quality MSAs, on larger datasets. The exponential scaling of the resources related to this alignment step is successfully mitigated by SIMSApiper through creating sequence subsets. SIMSApiper generates high quality structure-informed MSAs, which is illustrated by the excellent match with manually curated MSAs and its ability to align designed proteins with low sequence and 3D structural identity but conserved secondary structure elements. The validation further highlights the importance of the pre-processing step subdividing the sequences, as well as the post-processing squeezing step to reduce the computing time and the number of gaps, respectively. SIMSApiper represents not only a convenient way to create MSAs based on experimental and modeled structures, but also, it can significantly reduce calculation times while maintaining the accuracy of the generated structure-informed MSA.

Author contributions

W.V. conceptualized the method. C.C. developed the initial methodology. S.-L.H. created the Nextflow pipeline and corresponding documentation. C.C. contributed scripts. C.C. conducted the validation with the HOMSTRAD and TIM-barrels dataset while S.-L.H. conducted the validation with the GroEL dataset. A.D. provided technical support. W.V. provided supervision. S.-L.H., C.C., and W.V. wrote and reviewed the manuscript.

Supplementary data

Supplementary data are available at Bioinformatics online.

Conflict of interest

None declared.

Funding

This work was supported by Research Foundation Flanders (FWO) SB PhD fellowship [1SE5923N] to C.C.; the FWO large infrastructure grant [I000323N] to A.D.; and the Fonds de la Recherche Scientifique (FNRS) Aspirant fellowship to S.L.H. The resources used in this work were provided in part by the VSC (Flemish Supercomputer Center), funded by the Research Foundation—Flanders (FWO) and the Flemish Government.

Data availability

The complete dataset and scripts to reproduce the results are available on GitHub (github.com/Bio2Byte/simsapiper).