Database construction

The human mitochondrial proteome was downloaded from the MitoCarta 3.0 (1136 proteins) [44]. Additional annotations for the MitoCarta protein list were appended from the Integrated Mitochondrial Protein Index (Q4pre-2021) [23]. The yeast (Sacchromyces cerevisiae) mitochondrial proteome was derived from a high confidence dataset (901 proteins) [45]. Protein sequences were queried by UniProt ID and were submitted to: (1) iMLP–an internal MTS-like predictor using long short-term memory (LSTM) recurrent neural network architecture [30]; (2) TargetP2.0 –a presequence and cleavage site predictor using deep learning and bidirectional LSTM [31]. (3) MitoFates–a presequence and cleavage site predictor using support vector machine (SVM) classifiers [32]; (4) TPpred3 –a targeting and cleavage site predictor using Grammatical Restrained Hidden Conditional Random Fields [33]; (5) DeepMito–a sub-mitochondrial localization predictor using deep learning and convoluted neural networks [34]. For cleavage sites derived from N-terminomics, mass spectrometry data were aggregated from TopFIND 4.1 by Uniprot ID of both human and yeast proteins [39]. For variants and functional annotations of human proteins, dbNSFP v4.3a was parsed by Uniprot ID against GRCh38/hg38 coordinates [46]. The resulting list was filtered using an in-house Python script into separate datasets for gnomAD v3.1 and ClinVar. Variants unique to the ExAC database were ignored. AlphaFold models for the Homo sapiens proteome (UP000005640) and Sacchromyces cerevisiae (UP000002311) were downloaded and matched by Uniprot ID [47]. An in-house Python script based on BioPandas [48] was used to parse the PDB files and re-color B-factors according to iMLP scores (referred to interchangeably as iMTS scores) via iMLP. 3D visualization of protein structures was achieved using an adapted version of NGLViewer integrated into our R/Shiny application [49].

Database utility

MTSviewer serves as a user-friendly platform for investigating MTS’s from both genetic and structural perspectives. The database requires minimal bioinformatics knowledge and features both human and yeast mitochondrial proteomes. With MTSviewer, users are able to: (1) compare mitochondrial prediction outputs from a variety of algorithms; (2) visualize MTS likelihood on a folded protein structure; (3) compare experimentally identified and predicted proteolytic events; (4) map non-synonymous variants (gnomAD, ClinVar, or user uploaded) within these MTS’s and cleavage sites. Using this platform, we have also curated a list of disease-linked variants within human MTS’s as a resource for their functional characterization.

User interface

The MTSviewer user interface is intuitive and begins by selecting or searching a gene of interest. Users specify the desired database for variant visualization (currently gnomAD v3.1 or ClinVar), and variants are overlaid onto an XY plot with the iMTS probability from protein N- to C-terminus. Hovering over a variant reveals cursory details which are fully expanded in the variant table. For the structure viewer, two coloring schemes are toggleable: the iMTS score, or the AlphaFold per-residue predicted local distance difference test (pLDDT) confidence score. Users can investigate specific residues or variants by clicking on the iMTS plot or 3D structure, and the structure viewer will automatically highlight the interactions (i.e. polar contacts) and residues in proximity (5 Å) to the residue of interest (Fig 2). Users can also upload custom variant lists for their proteins of interest in CSV format, which will be added to the iMTS propensity curve, populated into the variant list data tables, and become visualizable on the 3D protein structure. This feature allows users to compare where their variants lie in terms of MTS propensity, cleavage sites and other pathogenic variants on primary sequence and structural levels.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. MTSviewer output for PINK1.

A sample output from MTSviewer investigating human PINK1, a mitochondrial kinase with a uniquely long MTS and multiple predicted cleavage sites. Protein sequence, prediction algorithms, and N-terminomics data tables have been omitted for clarity but are available in full on the interactive MTSviewer web server. Ser73Leu has been highlighted as a variant of interest, as Ser73 is found within a region of high MTS propensity near the predicted MitoFates cleavage site.

https://doi.org/10.1371/journal.pone.0284541.g002

The iMTS plot and structure viewer also contain toggleable visualizations to highlight cleavage site predictions from the various MTS predictors and/or experimentally determined N-terminomics sites. Aggregated comparisons of targeting predictors are pooled in table format, and data frames are exportable to facilitate downstream analyses. Taken together, these features enable users to rapidly generate protein-level hypotheses to test in vitro, or to rationalize previous in vitro findings with import- or protease-specific context.

PINK1 as a case study

To highlight the utility of MTSviewer we have chosen PINK1 as a case study (Fig 2), given its cryptic N-MTS and the innate coupling of its import and processing to gate its accumulation on the TOM complex. Briefly, PINK1 is known to be cleaved by the rhomboid protease PARL in the IMM at Ala103, which is validated by the N-terminomics outputs seen in MTSviewer. The precise MPP cleavage site within the PINK1 N-MTS remains unknown, though an MPP-cleaved PINK1 fragment accumulates upon PARL knockdown [50]. Based on the MTSviewer output for PINK1, there are many possibilities for the N-MTS MPP cleavage site, which will be critical to validate using in vitro assays, along with the effects of N-MTS variants (e.g. Gly30Arg, Pro52Leu, Arg57Cys, and Ser73Leu). While some of these PINK1 N-MTS variants are still cleaved by PARL in healthy mitochondria and accumulate following mitochondrial damage [51], their import rates and effects on MPP processing remain unstudied. Experiments which swap the PINK1 N-MTS with those from other mitochondrial proteins have shown that PINK1 can still be imported into mitochondria with chimeric N-MTS’s, though PINK1 accumulation is prevented [52]. While many of these N-MTS PINK1 chimeras can still be imported, their specific rates of import have also yet to be measured. This suggests that f elements of mitochondrial proteins (and variants within these regions) will be critical to study beyond the context of binary import success or blockage.

Another useful feature of MTSviewer is the ability to gauge the length of a protein’s N-MTS by looking at the iMTS propensity plots. For reference, it has been estimated that MTS’s are usually 15–50 amino acids long [4], yet the PINK1 N-terminus exhibits high MTS propensity across its first 90 amino acids. As all of the MTSviewer iMTS data is available to download, users will be able to analyze global trends in MTS length and propensity across protein families to investigate the downstream consequences of longer or atypical N-MTS’s within mitochondrial proteins. Beyond the N-MTS, the iMTS plot reveals a putative iMTS within the C-terminus of PINK1 (a.a. 460–500), which could regulate its import or processing rates at the mitochondrial surface. This is consistent with the knockdown of TOM70 reducing PINK1 import in vitro [53]. Furthermore, the PINK1 mRNA is co-transported with mitochondria [54], so it will also be important to investigate the role of TOM70 binding to PINK1 and this putative iMTS during translation and import. The MTSviewer output for PINK1 also highlights the need to consider the oligomerization status of proteins when investigating their monomeric AlphaFold structures. PINK1 is known to dimerize on the OMM following depolarization which could occlude its iMTS in the folded dimeric state [55, 56], even if partially unfolded PINK1 monomers bind to TOM70 upon import. Overall, MTSviewer will guide subsequent studies of atypically targeted proteins like PINK1 in the context of their MTS propensity, cleavage sites, and genetic variants.

Global profiling of MTS’s in human proteins

One caveat of the structure-based approach to MTSviewer is that N-terminal MTS’s within AlphaFold predictions are typically unstructured and have low confidence (i.e. pLDDT) scores. This is congruent with the majority of N-MTS’s being proteolytically removed prior to folding. Conversely, iMTS’s beyond the N-terminus could exist in both high or low pLDDT regions, depending on the AlphaFold prediction. To visualize the global relationship between structured or folded regions of a protein and their corresponding iMTS propensity, we generated a density plot for pLDDT and iMLP values on a per-residue basis for the entire human mitochondrial proteome (Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Global analysis of per-residue iMLP and pLDDT scores across the human mitochondrial proteome.

AlphaFold pLDDT and iMLP scores were extracted on a per-residue basis (426832 residues total) for all 1136 proteins in the human mitochondrial proteome and were visualized as a kernel density estimate (KDE) plot using the seaborn library in Python. Two density plots were generated and merged to visualize low density datapoints—one for all pLDDT and iMLP scores (blue), and one for residues with high MTS propensity (iMLP score > 3) (red). Based on the KDE plot, data were grouped into three subsets (Cluster 1: pLDDT 80–100, iMLP > 3, Cluster 2: pLDDT 25–50, iMLP > 3, and Cluster 3: pLDDT 80–100, iMLP < 1.5) and the % distance of each residue within their corresponding protein length were calculated and written as the average % distance from N-terminus ± standard deviation for all residues in that cluster.

https://doi.org/10.1371/journal.pone.0284541.g003

Of all residues, 69% were of high pLDDT values (pLDDT > 80) and low MTS propensity (iMLP score < 1.5) and were distributed within 25–75% of the protein sequence length (denoted cluster 3). This suggests that most residues in mitochondrial proteins are found in folded regions away from the extreme N- or C-termini and do not display significant MTS character. Next, we applied the same analysis to a subset of the data with iMLP scores > 3 in order to gauge what proportion of MTS’s exist in folded or unstructured regions, and where these MTS’s are found in the context of their protein sequence. From this subset of high MTS character residues, two distinct clusters emerged: one with high pLDDT scores (pLDDT > 80, denoted cluster 1), and one with low pLDDT scores (25 < pLDDT < 50, denoted cluster 2). Cluster 1 residues were found internally within protein sequences (average distance from N-terminus: 44.1%), while cluster 2 residues were found at protein N-termini (average distance from N-terminus: 7.4%). Taken together, this visualization provides quantitative evidence that (1) the majority of N-MTS’s in an AlphaFold protein structure will be unstructured with low pLDDT scores; and (2) the majority of iMTS’s within a protein will be found within a folded region of high pLDDT confidence.

Structural analysis of iMTS-containing proteins

These conclusions above raise the question: why is it useful to visualize MTS’s on a folded protein structure? To answer this, we first sought to characterize the iMTS’s of all mitochondrial proteins (i.e. how many iMTS’s does each protein contain and where are they located) to identify case studies. As such, we profiled the iMTS propensity curves of all mitochondrial proteins and counted their local maxima (defined as an amino acid with an iMLP score > 2 within a continuous 10 a.a. region of iMLP score > 1) using an in-house Python script (S1 File). To validate this approach, we first turned to the report that certain mitochondrial ribosomal proteins in yeast rely on iMTS’s for their targeting and internalization into mitochondria [12]. Within our dataset, 18/54 (33%) of yeast mitochondrial ribosomal proteins (MRPs–chosen in our dataset by gene name containing MRPL or MRPS) contained two or more MTS’s. In comparison, 34/78 (44%) of human MRPs contained two or more MTS’s, which suggests that this atypical targeting is likely conserved in humans. In yeast, Backes et al. previously highlighted Atp1 (human gene ATP5F1A), the alpha subunit of the mitochondrial ATP synthase, as a model substrate for their “stepping stone” import via TOM70 [9]. While Atp1 contains an N-MTS, its mitochondrial import was shown to be dependent on the presence of 3 iMTS’s–upon mutation of these iMTS’s, both TOM70 binding and mitochondrial import were abrogated for Atp1. Our maxima counting dataset correctly identified the same 3 iMTS’s of yeast Atp1, confirming the validity of our approach. In the AlphaFold structure of Atp1 on MTSviewer, these iMTS’s are tightly folded and have high pLDDT scores, which supports the model that TOM70 must bind to and maintain these iMTS-containing precursors in a partially unfolded, import-competent state to prevent their premature misfolding and/or cytosolic aggregation. In humans, ATP5F1A contains two iMTS’s, though specific import rates have not been studied. Other human proteins which contain multiple iMTS’s include POLRMT (6 iMTS’s), the mitochondrial RNA polymerase, and POLG (6 iMTS’s), the mitochondrial DNA polymerase. Given that TOM70 has recently been shown to regulate the transcriptional activity of mitochondrial proteins to drive biogenesis [57], the import dependency of POLRMT on TOM70 will remain an important question to address. Other hits with multiple iMTS’s include MAVS (3 iMTS’s), the mitochondrial antiviral signaling protein, which has been shown to interact with TOM70 upon RNA viral infection to stimulate the innate immune response [58]. While the AlphaFold structure of MAVS is highly disordered, MTSviewer can be used as a platform to design in vitro experiments that will clarify the role of these iMTS’s in the formation of an antiviral signaling complex and/or localization of MAVS.

Next, we used our local maxima counting dataset to identify other proteins which contain C-terminal MTS’s (C-MTS) by searching for iMLP maxima > 4 within the final 10% of their sequence (i.e. the extreme C-terminus). Within these hits was KMO (kynurenine 3-monooxygenase), a monooxygenase located primarily in microglia which regulates kynurenine metabolism and is implicated in neurological disorders including schizophrenia and Huntington’s disease [59, 60]. Experimental studies confirmed the C-terminus dependent localization of KMO to the OMM in pig liver—specifically, mitochondrial localization of KMO was lost when truncated at a.a. 448, while its enzymatic activity was still retained [61]. From our maxima counting dataset, the C-MTS of human KMO is located between a.a. 420–460, with a maximum iMTS score of 5.2 at Tyr439. Using our structural visualization in MTSviewer, the KMO C-MTS is not predicted to be a part of the main KMO fold and as such could be solvent-accessible to bind TOM70 or anchor to the OMM even when the rest of the protein is folded, which corroborates the deletion experiments in pig KMO (Fig 4). KMO possesses a variant (Arg452Cys) which is located within this C-MTS and is associated with bipolar disorder, though the functional consequences of this mutation remain uncharacterized [62]. The dependence of KMO OMM localization on specific import receptors (i.e. TOM70) remains to be elucidated, but we hope that MTSviewer will facilitate these future explorations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Structural visualization of human KMO and PUSL1 iMTS propensity via MTSviewer.

AlphaFold models of KMO (Uniprot: O15229) and PUSL1 (Uniprot: Q8N0Z8) colored by iMLP score within MTSviewer. MTS’s are annotated as blue rectangles on the sequence schematics. The C-MTS of KMO is predicted as helical, solvent-accessible, and is distinct from the folded domain of KMO. Arg452 is highlighted as a variant of interest (Arg452Cys). PUSL1 contains a putative N-MTS and two iMTS’s (denoted iMTS-1 and iMTS-2), which are both found within the folded domain and would likely be inaccessible for TOM70 binding in this mature state.

https://doi.org/10.1371/journal.pone.0284541.g004

To find another example of a protein with high MTS propensity within a confidently folded region, we calculated the product of iMTS and pLDDT scores across the human mitochondrial proteome. The top hit from this calculation was PUSL1 (tRNA pseudouridine synthase-like 1), with its iMTS spanning a.a. 90–120 (average pLDDT = 97.7; average iMTS score: 5.0; local iMTS max at Arg103). PUSL1 is a recently characterized mitoribosome-interacting protein (MIP) in the mitochondrial matrix which associates with the IMM and regulates mitochondrial translation, most likely via its interaction with the 39S assembly intermediate [63]. Based on the structural visualization, the PUSL1 iMTS’s would be inaccessible for TOM70 binding in its folded state, which suggests its potential reliance on the TOM70 “stepping stone” model for its import (Fig 4). Very little is known about the import regulation and structural function (i.e. which residues are required for its activity) of PUSL1, and as such PUSL1 remains an intriguing protein to study in the context of its targeting sequences and enzymatic function.

It is important to highlight that while regions of high iMTS propensity most likely bind to TOM70, high iMLP scores can also be suggestive of domains that bind to other chaperones or proteins with tetratricopeptide repeat (TPR) domains, as previously highlighted [30]. High iMLP scores also do not guarantee mitochondrial targeting–as shown in Bykov et al. 2022, the targeting of some yeast MRP’s (MRPL15) required the presence of an N-MTS, while others (e.g. MRP20, MRP35) were sufficiently targeted by iMTS’s in the absence of any N-MTS [12]. Also, in humans, creatine kinase B was shown to localize to mitochondria via its iMTS, though this localization was unchanged upon TOM70 silencing [11]. As such, in vitro experiments will remain critical to validate hypotheses generated with MTSviewer and to clarify the determinants for successful precursor targeting and TOM70 dependency in humans. With MTSviewer, users will be able to rapidly assess the distance between MTS’s and folded regions, to validate the "stepping stone" model for proteins of interest in vitro, and to assess whether variants in these regions would affect TOM70-mediated import and/or function of the mature, folded protein. Overall, our platform will also facilitate the discovery and characterization of novel TOM70 substrates in humans.