CoMR: an integrative scoring pipeline for Comprehensive Mitochondrial proteome Reconstruction across eukaryotes

The authors present CoMR, an integrative scoring pipeline that combines targeting prediction, homology searches, and phylogenetic analysis to significantly improve the accuracy of mitochondrial proteome reconstruction across diverse eukaryotes compared to standalone predictors.

The Gist

Imagine your cell is a bustling city. Inside this city, there is a very special power plant called the mitochondrion. It generates the energy the city needs to run. To keep this power plant working, it needs a specific crew of workers (proteins) who know how to get inside and do their jobs.

The problem is: How do we know which workers belong in the power plant and which ones belong in the library or the bakery?

For a long time, scientists tried to guess by looking at the workers' "badges" (called Mitochondrial Targeting Signals). If a worker had a specific badge, they were assumed to be a power plant employee. But this method has a big flaw: the badges look different in different cities (species), and some workers have no badges at all but still work in the power plant. Relying only on badges meant scientists were missing a lot of important workers or misidentifying others.

Enter CoMR (Comprehensive Mitochondrial Reconstructor). Think of CoMR not as a single detective, but as a super-intelligent hiring committee that uses a "scorecard" to decide who gets the job.

How CoMR Works: The Hiring Committee

Instead of just checking the badge, CoMR asks four different questions to build a complete picture of every worker in the cell:

1. The Badge Check (Targeting Prediction): "Do you have the standard power plant badge?" (This is what old methods did).
2. The Family Reunion (Homology Search): "Do you look like other workers we know work in power plants?" CoMR compares the worker's DNA to a massive database of known power plant workers from many different species.
3. The "Look-Alike" Search (Large-Scale Search): "Does anyone in the entire world of biology look like you and work in a power plant?" It scans the entire internet of biological data.
4. The Evolutionary Tree (Phylogenetic Analysis): "Does your family tree show that your ancestors worked in power plants?" It builds a family tree to see if your lineage belongs in the energy sector.

The Magic Score:
CoMR takes the answers to these four questions and gives the worker a score.

• If they have a badge, +1 point.
• If they look like a known power plant worker, +1 point.
• If they match a global search, +1 point.
• If their family tree says "yes," +1 point.

If a worker gets a high score (e.g., 4 or 5 out of 6), the committee is very confident: "This person belongs in the power plant!" If they get a low score, they probably belong elsewhere.

Why is this a big deal?

The authors tested this new "Hiring Committee" in two very different scenarios:

1. The Model City (Yeast):
They tested it on yeast, a well-studied organism where we already know most of the workers.

• Old Method (Just checking badges): Got it right about 72% of the time.
• CoMR (The Committee): Got it right 92% of the time.
• Analogy: It's like upgrading from a security guard who only checks IDs to a team that checks IDs, fingerprints, background checks, and family history.

2. The Mystery City (A weird, ancient protist):
Then they tested it on Paratrimastix pyriformis, a strange, ancient single-celled organism that lives in low-oxygen environments. Its "power plant" is a weird, reduced version, and its workers have very strange or missing badges.

• The Challenge: In this city, the power plant workers are incredibly rare (only 32 out of 13,000 workers). It's like finding a needle in a haystack.
• Old Method: Failed miserably, mostly guessing wrong because the "badges" didn't exist.
• CoMR: Even with the badges missing, CoMR used the other clues (family history and look-alikes) to find the workers. It was 78 times better at finding the needle than random guessing and 10 times better than the old badge-checking method.

The Takeaway

Before CoMR, scientists were trying to solve a complex puzzle with only one piece of evidence (the badge). If the badge was missing or weird, they gave up.

CoMR is like giving the puzzle solver a flashlight, a magnifying glass, and a map all at once. By combining different types of evidence, it can reconstruct the full list of power plant workers even in the most difficult, weird, and ancient organisms. This helps us understand how life evolved and how these tiny power plants changed over billions of years.

In short: Don't judge a book by its cover (badge); read the whole story (CoMR) to know what it's really about.

Technical Summary

1. Problem Statement

Reconstructing mitochondrial proteomes from eukaryotic sequence data is a critical challenge in evolutionary biology and cell biology. Current methods rely heavily on Mitochondrial Targeting Signal (MTS) predictors (e.g., TargetP, MitoProt). However, these tools have significant limitations:

• Training Bias: They are primarily trained on model organisms (e.g., humans, yeast) and assume canonical N-terminal targeting sequences.
• Failure in Divergent Lineages: In phylogenetically divergent or anaerobic organisms (e.g., metamonads), MTSs are often atypical, reduced, or absent, leading to high false-negative rates.
• False Positives: Standalone predictors often generate high false-positive rates, particularly in non-model organisms.
• Lack of Integration: There is no unified framework that combines MTS prediction with homology, profile-based, and phylogenetic evidence to improve accuracy across diverse eukaryotes.

2. Methodology: The CoMR Pipeline

The authors developed CoMR (Comprehensive Mitochondrial Reconstructor), a modular, Snakemake-based workflow designed to integrate multiple independent lines of evidence into a unified scoring framework. The pipeline accepts predicted proteomes (FASTA) and executes three parallel evidence streams:

A. Evidence Streams

1. Targeting Signal Prediction:
   • Runs four independent predictors: TargetP 2.0, MitoProt, MitoFates, and DeepMito.
   • Outputs are parsed to standardize scores and categorical calls. Note: DeepMito is used for sub-mitochondrial localization but is excluded from the composite score to avoid bias.
2. Homology and Profile-Based Searches:
   • DIAMOND BLASTP: Searches against the Subtractive Mitochondrial Database (SMD) (curated mitochondrial vs. non-mitochondrial proteins from six diverse organisms) and the NCBI Non-Redundant (NR) database.
   • HMM Profile Search: Uses HMMER to scan against a curated CoMR mitochondrial profile HMM database derived from the SMD.
   • Taxonomy: NCBI taxonomy is reconstructed for NR hits to filter lineage-specific biases.
3. Phylogenetic Analysis:
   • Sequences with significant HMM hits are aligned (MAFFT), trimmed (trimAl), and used to infer phylogenetic trees (IQ-TREE 2).
   • Tree Parsing: Automated parsing (ete3) classifies sequences as mitochondrial or non-mitochondrial based on their placement relative to known reference sequences (Fitch-style traversal).

B. Evidence Integration and Scoring

• Unified Scoring: All evidence layers are merged into a per-sequence table.
• Scoring Logic: A rule-based system assigns points for positive evidence:
  • +1 for each positive MTS predictor (TargetP, MitoProt, MitoFates).
  • +1 for curated homology support (SMD matches).
  • +1 for phylogenetic placement in a mitochondrial clade.
  • +1 for large-scale homology support (NR hits with mitochondrion-related keywords).
  • Composite Score: Ranges from 0 to 6.
• Flexibility: The scoring weights are defined in external "scorecard" files, allowing users to adjust weights based on specific lineage characteristics or to perform ablation studies.

C. Implementation

• Distributed as a fully containerized workflow (Docker/Singularity) to ensure reproducibility.
• Uses Python 3.11, Snakemake, and standard bioinformatics tools (HMMER, DIAMOND, MAFFT, IQ-TREE).

3. Key Contributions

• Integrative Framework: CoMR is the first pipeline to systematically combine MTS prediction, curated homology, large-scale homology, and automated phylogenetics into a single, tunable scoring metric.
• Customizable Scoring: The ability to re-weight evidence layers allows the tool to adapt to different phylogenetic contexts where specific evidence types (e.g., homology vs. targeting signals) may be more or less reliable.
• Reproducibility: The pipeline is fully containerized with public access to code, databases, and benchmarking data via GitHub and Figshare.

4. Results and Performance Evaluation

The authors benchmarked CoMR on two distinct datasets: a model organism (Saccharomyces cerevisiae) and a divergent anaerobic protist (Paratrimastix pyriformis).

A. Model Organism (S. cerevisiae)

• Performance: CoMR achieved a ROC-AUC of 0.92, significantly outperforming standalone TargetP2 (ROC-AUC = 0.72).
• Ablation Analysis: Removing the NCBI NR homology search caused the largest drop in performance, indicating that for well-represented lineages (fungi), large-scale homology is the strongest signal.
• Robustness: Down-weighting or removing individual predictors (like MitoProt, which had higher false positives) resulted in only modest performance changes, demonstrating the pipeline's robustness to individual tool biases.

B. Divergent Organism (P. pyriformis)

• Context: This organism has a highly reduced mitochondrial-related organelle (MRO) with only 32 annotated proteins in a proteome of ~13,500 (extreme class imbalance).
• Performance: CoMR maintained strong performance with a ROC-AUC of 0.86.
• Precision-Recall: Under extreme imbalance, ROC-AUC can be misleading. CoMR achieved a Precision-Recall AUC (PR-AUC) of 0.183.
  • This represents a ~78-fold enrichment over random expectation.
  • It is nearly a 10-fold improvement over TargetP2 (PR-AUC = 0.018).
• Evidence Contribution: In this divergent lineage, both curated SMD homology and NR homology were critical, highlighting the need for curated databases when public databases lack representation for specific clades.

C. Computational Efficiency

• Runtime scales with proteome size and resource allocation. On a high-performance cluster (128 CPUs), CoMR processed S. cerevisiae in ~3 hours and P. pyriformis in ~5 hours.
• The NCBI NR search is the most computationally intensive step, but the pipeline remains functional even when disabled, relying on other evidence layers.

5. Significance

• Overcoming Limitations of Single-Method Approaches: The study proves that relying solely on MTS prediction is insufficient, especially for non-model organisms. Integrating phylogenetic and homology evidence significantly boosts accuracy.
• Enabling Research in Non-Model Lineages: CoMR provides a robust tool for studying the evolution of mitochondria and mitochondrial-related organelles (MROs) in anaerobic and phylogenetically distant eukaryotes, where traditional tools fail.
• Scalability and Adaptability: The modular, scorecard-based design allows researchers to tailor the pipeline to their specific organism of interest, making it a versatile resource for comparative genomics.
• Resource Availability: By providing open-source code, databases, and containerized environments, the authors facilitate reproducible mitochondrial proteome reconstruction across the scientific community.

In conclusion, CoMR represents a major advancement in computational biology for organelle reconstruction, shifting the paradigm from single-signal prediction to a comprehensive, evidence-integrated approach that is effective across the full spectrum of eukaryotic diversity.