MitoAtlas: a Domain-Resolved Spatial Map of the HumanMitochondrial Proteome

The paper introduces MitoAtlas, a comprehensive spatial systems biology resource that utilizes super-resolution proximity labeling, machine learning, and AlphaFold modeling to generate a high-resolution domain-level map of the human mitochondrial proteome, resolving membrane topologies, annotating previously uncharacterized proteins, and identifying novel protein-protein and protein-metabolite interactions.

The Gist

Imagine a mitochondrion (the powerhouse of your cells) not just as a single battery, but as a high-tech factory with four distinct floors: a basement (matrix), a middle hallway (intermembrane space), an inner wall (inner membrane), and an outer wall (outer membrane).

For decades, scientists had a list of the workers (proteins) in this factory, but they didn't know exactly where on the factory floor each worker stood, or which way they were facing. Was a worker standing in the basement, or were they stuck in the wall? This lack of detail made it hard to understand how the factory actually runs.

MitoAtlas is the new, ultra-detailed blueprint that solves this mystery. Here is how the researchers did it, explained simply:

1. The "Spray Paint" Strategy (Super-Resolution Proximity Labeling)

Imagine you want to know exactly which rooms in a dark building are connected. You can't just look at a map; you have to go inside.

• The Old Way: Scientists used to take the factory apart, separate the floors, and see what was in each pile. But this is messy; things get mixed up, and you lose the "who is standing next to whom" information.
• The New Way (MitoAtlas): The team used 42 different "bait" proteins, each acting like a smart spray paint gun. They placed these guns in specific spots (the basement, the walls, the hallway).
  • When they turned them on, the guns sprayed a tiny, sticky tag (biotin) onto any protein that was standing right next to them.
  • They used two types of guns: APEX2 (which sprays a fine mist that reaches into tiny cracks) and BioID (which sprays a bigger blob that only sticks to things it touches directly).
  • By using 42 different guns in different locations, they created a massive, overlapping web of sticky tags.

2. The "Detective Algorithm" (Machine Learning)

Now they had a huge pile of sticky tags on thousands of proteins. The question was: Which floor is this tag on?

• They fed this data into a computer detective (Machine Learning).
• The detective looked at the "fingerprint" of tags. For example, if a protein was tagged heavily by the "basement guns" but ignored by the "outer wall guns," the computer knew: "Ah, this protein lives in the basement."
• This allowed them to pinpoint the location of 861 proteins with incredible precision, down to the specific amino acid (the individual brick) that was tagged.

3. The "3D Puzzle" (AlphaFold)

Sometimes, the tags were confusing. Was a protein sticking out of the wall, or was it fully inside?

• The team combined their experimental data with AlphaFold, an AI that predicts what proteins look like in 3D.
• It's like taking a blurry photo of a person in a crowd and overlaying a perfect 3D model of that person to see exactly how they are standing.
• This helped them figure out the topology: Is the protein's head in the basement and its feet in the hallway? Or is it a barrel shape sitting in the wall?

What Did They Discover?

This new map revealed things we didn't know before:

• The "Orphans": They found 160 proteins that were missing from previous maps entirely. They were like workers who were on the payroll but nobody knew where they worked. Now we know exactly where they stand.
• The "Wrong Address" Fixes: Some proteins were thought to be in the basement, but the map showed they were actually in the hallway. It's like realizing the CEO was actually working in the breakroom all along.
• New Interactions: They found proteins that were holding hands (interacting) that we didn't know were friends. For example, they found a specific complex between LETM1 and Citrate Synthase, which helps explain how the factory manages energy and calcium.
• Disease Clues: By mapping where disease-causing mutations happen, they found that if a mutation is on the "basement side" of a wall protein, it causes one type of disease, but if it's on the "hallway side," it causes a different one. This is like realizing that a broken door handle on the inside of a room causes a different problem than a broken handle on the outside.

The Big Picture

MitoAtlas is like upgrading from a 2D floor plan to a 3D, augmented-reality tour of the mitochondrion.

Instead of just saying, "This protein is in the mitochondrion," we can now say, "This protein is a 7-story skyscraper embedded in the inner wall, with its lobby facing the basement and its roof facing the hallway."

This level of detail is crucial because, in biology, location is everything. Knowing exactly where a protein stands tells us exactly what it does, how it talks to its neighbors, and what happens when it breaks. This map is now available online for any scientist to use, acting as a GPS for the next generation of mitochondrial research.

Technical Summary

1. Problem Statement

The mitochondrion is a complex organelle defined by a double-membrane system partitioning it into four distinct physiological compartments: the outer mitochondrial membrane (OMM), intermembrane space (IMS), inner mitochondrial membrane (IMM), and matrix. While the mitochondrial proteome has been cataloged (e.g., in MitoCarta3.0), current resources lack domain-level spatial resolution.

• Limitations of Existing Data: MitoCarta3.0 assigns ~90 proteins to ambiguous categories like "membrane" or "unknown" and fails to resolve the topological orientation (e.g., which side of the membrane a domain faces) for over 400 proteins.
• Methodological Gaps: Traditional biochemical methods (e.g., protease protection assays) often suffer from low throughput, non-physiological conditions, and artifacts. Single-bait proximity labeling approaches can yield false positives or mislocalization due to enzyme fusion artifacts (e.g., trapping proteins on the wrong membrane).
• Need: A high-resolution, systems-level approach is required to map protein localization, membrane topology, and functional domain architecture simultaneously to understand mitochondrial metabolism and disease mechanisms.

2. Methodology

The authors developed MitoAtlas, a spatial systems biology resource built on an expanded Super-Resolution Proximity Labeling (SR-PL) workflow.

• Multiplexed Bait Library:
  • Generated 42 stable HEK293T-REx cell lines expressing fusion proteins of APEX2 (peroxidase) and BioID/BioID2 (biotin ligase) targeted to specific sub-mitochondrial compartments.
  • Strategic Deployment: APEX2 (contact + diffusive labeling, ~10 nm radius) was biased toward the dense Matrix and IMS (17 baits). BioID (strictly contact-dependent labeling) was primarily deployed to the OMM (11 baits) to capture confined neighborhoods and avoid the diffusive artifacts seen with APEX2 on OMM proteins.
• SR-PL Workflow:
  • Cells were labeled with desthiobiotin-phenol (for APEX2) or biotin (for BioID).
  • Labeled peptides were enriched via streptavidin pulldown and identified via LC-MS/MS.
  • Data Output: 13,440 unique proximity-labeled sites across 4,058 proteins (6,046 tyrosine sites from APEX2; 7,394 lysine sites from BioID).
• Computational Pipeline (Two-Step Classification):
  • Step 1 (Site Classification): Logistic Regression (LR) models were trained on MS intensity profiles across all 42 baits to classify individual labeled sites into Matrix, IMS, OMM, or Non-mitochondrial compartments.
  • Step 2 (Protein Classification): Site-level probabilities were aggregated to assign protein-level localizations.
    • Non-TM Proteins: Assigned "Gold" (≥2 concordant sites) or "Silver" (1 site) confidence tiers.
    • Transmembrane (TM) Proteins: Integrated site-level classifications with TM domain predictions from four sources (UniProt, DeepTMHMM, TMbed, AF2-TMDET) and AlphaFold2 structures to resolve membrane topology (e.g., N-terminus in Matrix vs. IMS).
• Integration with AI: The study leveraged AlphaFold2/3 and AlphaFold-Multimer to validate topologies, predict protein-protein interactions (PPIs), and model protein-metabolite docking.

3. Key Contributions & Results

A. High-Resolution Spatial Map

• Coverage: MitoAtlas resolves the sub-mitochondrial architecture of 861 proteins, covering 79% of OXPHOS subunits, 91% of the mitochondrial ribosome, and 86% of the TOM complex.
• Resolution: Unlike MitoCarta3.0, which provides broad compartment labels, MitoAtlas provides domain-level resolution, distinguishing between matrix-facing, IMS-facing, and transmembrane domains.

B. Membrane Topology Determination

• Transmembrane Proteins: Successfully resolved the topology of 196 transmembrane proteins (136 IMM-TM, 60 OMM-TM).
• Corrections & Discoveries:
  • DNAJC11: Reclassified from a peripheral OMM protein to an OMM-TM protein with a novel 17-stranded $\beta$-barrel architecture, suggesting a role as an accessory import pore.
  • LETM1: Resolved the long-standing debate by reclassifying LETM1 as a peripheral matrix protein (not an IMM-TM antiporter), consistent with its interaction with Citrate Synthase (CS).
  • SLC30A9 & SFXN2: Proposed a 6-TM domain model (correcting UniProt's 5-TM annotation) based on labeling patterns and AlphaFold structures.
  • SMIMs: Resolved the topology of seven small integral membrane proteins (SMIMs), reclassifying SMIM8 and SMIM12 as IMM-TM proteins.

C. Expansion of the Mitochondrial Proteome ("Mito-Orphans")

• Identified 160 mitochondrial proteins absent from MitoCarta3.0.
• Validated specific "Mito-Orphans" (e.g., GTPBP8, C17orf80) as matrix residents and MINPP1 as an IMS resident.
• MINPP1 Discovery: Demonstrated that MINPP1, previously thought to be ER-localized, resides in the IMS at ER-mitochondria contact sites, where it hydrolyzes InsP6 to IP3 to regulate mitochondrial calcium flux.

D. Interaction Networks

• Protein-Protein Interactions (PPIs): Identified 47 high-confidence interactions by combining bait-specific labeling signatures with AlphaFold-Multimer.
  • Validated the LETM1–Citrate Synthase (CS) complex. Mutations at the interface (R380, R393) abolished binding, linking structural data to pathogenic variants in mitochondrial disease.
• Protein-Metabolite Interactions: Predicted 155 protein-metabolite interactions. Validated MTX1 as a heme-binding OMM protein, suggesting a role in heme transport.

E. Functional Insights from Topology

• Disease Variants: Mapping ClinVar pathogenic variants onto MitoAtlas topologies revealed topology-dependent disease segregation. For example, in SLC25A4 (ANT1), IMS-facing variants cause progressive external ophthalmoplegia (adPEO), while matrix-facing variants cause mtDNA depletion syndrome.
• Post-Translational Modifications (PTMs): Revealed compartment-specific PTM landscapes (e.g., succinylation enriched on matrix-facing loops; ubiquitination restricted to cytosolic OMM faces).
• Evolutionary Constraint: Showed that transmembrane domains of IMM proteins are under stronger purifying selection than soluble loops.

4. Significance

• Resource Creation: MitoAtlas (available at mitoatlas.org) is the most detailed sub-mitochondrial proteome map to date, offering an interactive platform for exploring protein localization, topology, and interactions.
• Methodological Advance: Demonstrates that integrating multi-bait proximity labeling with machine learning and AI structural modeling can overcome the limitations of traditional fractionation and single-bait approaches.
• Biological Impact:
  • Resolves long-standing topological ambiguities (e.g., LETM1, DNAJC11) that hindered functional understanding.
  • Provides a framework for interpreting disease-causing mutations based on their specific topological context (membrane face vs. loop).
  • Establishes a new standard for mapping organelle architecture, enabling the discovery of novel protein functions, interaction networks, and metabolic regulation mechanisms.

In summary, MitoAtlas transforms the understanding of mitochondrial biology from a compartment-level view to a domain-level structural and functional map, bridging the gap between proteomics, structural biology, and systems medicine.