A Protein Language Model Reveals Organellar Calcium ATPases at Neuronal Synapses

The study introduces Synapse Gigamapper (SyGi), a protein language model that predicts synaptic localization and reveals the unexpected presence of ER-bound SERCA pumps at neuronal synapses, where they colocalize with spine apparatuses and synaptic vesicles to regulate calcium.

The Gist

Imagine the human brain as a bustling, hyper-advanced city. In this city, the most important interactions happen at tiny "meeting points" called synapses, where neurons (brain cells) pass messages to one another. For decades, scientists have been trying to map out exactly which proteins (the workers and machines) are stationed at these meeting points.

However, the city is so complex that traditional mapping tools have missed a huge number of workers. Many proteins were found in the data but were labeled as "unknowns" or "contaminants" because they appeared in such small numbers that scientists thought they were mistakes.

This paper introduces a new, super-smart detective named SyGi (Synapse Gigamapper) that changes how we look at the brain. Here is the story of what they found, explained simply:

1. The New Detective: SyGi

Think of the protein sequence (the string of amino acids that makes up a protein) as a unique barcode or a secret code.

For a long time, scientists had to physically go to the synapse, take a sample, and look for these barcodes. It was slow, expensive, and often missed the "ghost" proteins that were there but hard to see.

The authors built SyGi, an Artificial Intelligence (AI) detective trained on a massive library of known protein codes. Instead of looking at the physical synapse, SyGi looks at the text of the protein's code. It asks: "Based on the letters in this code, does this protein belong at a synapse?"

It's like having a librarian who can tell you exactly which book belongs on the "Science Fiction" shelf just by reading the first few sentences of the story, without ever seeing the book cover.

2. The Big Discovery: The "Hidden Plumber"

The AI detective scanned thousands of proteins and found a specific group of workers that everyone had been ignoring: SERCA pumps.

• The Old Belief: Scientists thought SERCA pumps were like plumbing pipes that only lived in the "basement" of the cell (the Endoplasmic Reticulum). They were thought to be too big and clumsy to ever fit into the tiny, crowded synapses.
• The New Reality: SyGi predicted that SERCA pumps were actually hanging out at the synapses. When the scientists went to check with high-powered microscopes, they found SyGi was right!

But here is the twist: The SERCA pumps weren't in the basement. They had moved into two very specific, tiny "apartments" right at the meeting point:

1. The Spine Apparatus: A tiny, specialized storage unit inside the receiving end of the synapse (the postsynaptic side).
2. The Synaptic Vesicles: Tiny bubbles that carry messages out of the sending end of the synapse (the presynaptic side).

3. Why This Matters: The "Fire Extinguisher" Analogy

To understand why this is a big deal, imagine a neuron firing a message is like a fireworks display.

• When the fireworks go off, they release a lot of calcium (the sparks).
• If the sparks (calcium) stay around too long, they cause a mess and ruin the next show.
• The cell needs fire extinguishers (calcium pumps) to clean up the sparks immediately.

Before this paper: We thought the fire extinguishers (SERCA) were kept in a warehouse far away (the ER) and had to run over to the fireworks when needed.
After this paper: We discovered that the fire extinguishers are actually parked right next to the fireworks, inside the tiny bubbles and storage units.

This means the brain can clean up the "sparks" (calcium) much faster and more precisely. It allows for sharper, faster, and more complex thinking and memory formation.

4. The Takeaway

This paper isn't just about finding one protein; it's about a new way of thinking.

• The Problem: We were throwing away "low-abundance" proteins (the quiet, hard-to-see workers) because our old tools couldn't find them.
• The Solution: The AI (SyGi) acts as a magnifying glass for the "text" of life. It found patterns in the code that human eyes missed.
• The Result: We now know that the brain has a much more sophisticated, localized cleanup crew for calcium than we ever imagined.

In short, the authors used a "protein language model" to translate the secret code of the brain, revealing that the "plumbers" (SERCA) have been working right at the front door of the synapse all along, keeping the brain's communication lines clear and efficient.

Technical Summary

1. Problem Statement

The synaptic proteome of neurons is critical for information transfer and storage, yet a significant fraction of proteins identified via mass spectrometry in synaptic compartments (synaptosomes) remain unannotated or unassigned in expert-curated databases like SynGO.

• The Gap: Current databases cover only a subset of identified proteins. Many low-abundance proteins, which may exist in specific synapse subsets and drive heterogeneity, are often discarded as contaminants or false positives due to technical limitations in proteomics.
• The Challenge: There is a need for a high-throughput, hypothesis-generating method to predict the subcellular localization of these "hidden" proteins specifically at excitatory and inhibitory synapses, distinguishing them from general cellular compartments.

2. Methodology: Synapse Gigamapper (SyGi)

The authors developed SyGi, a transformer-based protein language model designed to predict protein localization to specific subcellular compartments based solely on amino acid sequences.

• Architecture:
  • Encoder: Utilizes ESM-2 (Evolutionary Scale Modeling, 8M parameters) to generate contextualized sequence embeddings. A parallel version using ESM-C (600M parameters) was tested to evaluate the impact of model capacity.
  • Classifier: A multi-layer perceptron (MLP) trained for multi-label classification across six compartments: Excitatory Synapse, Inhibitory Synapse, Endoplasmic Reticulum (ER), Mitochondria, Nucleus, and Cytosol.
• Training Data:
  • Integrated proteomic datasets from cortical excitatory/inhibitory synaptosomes and global organelle profiling.
  • Dataset Size: 6,627 unique human UniProt IDs with experimentally validated localization annotations.
  • Class Imbalance Handling: Synaptic proteins represent only ~14% of the dataset. The model employs ROC-AUC as the primary metric (robust to imbalance) and tested weighted loss functions for the larger ESM-C model to improve recall for minority classes.
• Feature Analysis:
  • SyGi calculates attribution scores for individual amino acids to identify Short Linear Motifs (SLMs) indicative of synaptic localization.
  • Motif discovery was performed using the MEME Suite (XSTREME) on the predicted synaptic proteins.

3. Key Contributions

1. Development of SyGi: A specialized protein language model capable of predicting localization to excitatory vs. inhibitory synapses with high accuracy (average ROC-AUC > 0.89).
2. Motif Discovery: Identification of 152 amino-acid motifs (80 excitatory-specific, 72 inhibitory-specific) that serve as localization signals, many of which were previously unknown or not annotated in the ELM database.
3. Hypothesis Generation: Application of SyGi to screen key cellular pathways (calcium homeostasis, kinases, proteasomes, etc.), shortlisting >100 high-confidence candidate proteins not previously annotated as synaptic constituents.
4. Experimental Validation: The discovery and experimental verification of SERCA (Sarco/Endoplasmic Reticulum Ca2+ ATPase) localization at synapses, challenging the canonical view that SERCA is exclusively ER-bound.

4. Key Results

A. Model Performance and Motif Analysis

• Accuracy: SyGi achieved ROC-AUC values >0.8 across all six compartments. It successfully filtered expert-curated SynGO genes, enriching for core synaptic components while depleting housekeeping proteins.
• Motifs: The model identified 152 motifs ranging from 4–16 amino acids.
  • Shared Motifs: 7 motifs were shared between excitatory and inhibitory synapses.
  • Specificity: Motifs included known signals (e.g., proline-rich motifs for PSD proteins, transmembrane domains) and novel acidic docking motifs (e.g., WFRGLNRI for calcium pumps).
  • Biophysical Properties: UMAP analysis revealed five distinct biophysical clusters (polar, charged, aromatic, hydrophobic), with no strict clustering by synapse type, suggesting a shared "vocabulary" of localization signals.

B. Discovery of SERCA at Synapses

SyGi predicted high-confidence localization of calcium homeostasis proteins, specifically SERCA, at excitatory synapses. This led to a multi-modal experimental investigation:

• Observation 1: Punctate Distribution: Unlike the diffuse distribution of Plasma Membrane Ca2+ ATPase (PMCA), SERCA formed distinct clusters at excitatory synapses (vGluT1-positive) in dendrites.
• Observation 2: Postsynaptic Host (Spine Apparatus):
  • SERCA clusters were found in dendritic spines without the presence of the general ER marker (Calnexin).
  • Instead, SERCA colocalized with Synaptopodin, a specific marker for the Spine Apparatus (a specialized ER extension in spines).
  • DNA-PAINT super-resolution microscopy confirmed SERCA hotspots overlap with the spine apparatus contours.
• Observation 3: Presynaptic Host (Synaptic Vesicles):
  • Surprisingly, SyGi predicted and experiments confirmed SERCA presence in the presynaptic compartment.
  • Expansion Microscopy & DNA-PAINT: SERCA clusters were found colocalized with vGluT1 (presynaptic marker).
  • Purified Vesicles: Immunofluorescence on purified synaptic vesicles (Vamp2-positive) showed that 86% of vesicles contained SERCA or PMCA, with 47% containing both.
  • Topology: SERCA was detected on non-permeabilized vesicles, indicating its cytoplasmic domains face the cytosol, consistent with a functional role in vesicular calcium uptake.

C. Quantitative Findings

• Copy Number: Despite lower overall dendritic density compared to PMCA, the synaptic clustering of SERCA resulted in a higher average copy number per synapse (~32.8 SERCA vs. ~19.5 PMCA).
• Presynaptic Bias: Quantification in brain slices revealed significantly higher SERCA levels in the presynaptic compartment compared to the postsynaptic compartment.

5. Significance and Implications

• Redefining Calcium Clearance: The study reveals that calcium clearance at synapses is not solely a function of the plasma membrane or bulk ER. Instead, it involves compartment-specific organelles:
  • Postsynaptic: The Spine Apparatus acts as a localized calcium sink.
  • Presynaptic: Synaptic vesicles themselves may act as calcium sinks via SERCA/PMCA, potentially providing negative feedback for vesicle fusion and regulating release probability.
• Methodological Impact: SyGi demonstrates that protein language models can effectively "salvage" low-abundance proteins from proteomic noise, transforming them into testable biological hypotheses. It highlights that "contaminants" in proteomics may actually be genuine, subset-specific synaptic components.
• Biological Insight: The presence of SERCA on synaptic vesicles suggests the ER contributes more than just lipids to vesicle assembly; it may provide machinery for active calcium regulation directly at the release site, a mechanism previously unappreciated in synaptic physiology.

In summary, this paper bridges the gap between AI-driven protein sequence analysis and experimental neurobiology, uncovering a novel, compartmentalized mechanism for calcium regulation at the synapse mediated by the unexpected localization of SERCA.