🧬 biology

CAPSUL: A Comprehensive Human Protein Benchmark for Subcellular Localization

The paper introduces CAPSUL, a comprehensive human protein benchmark that integrates diverse 3D structural representations with expert-curated subcellular localization annotations to address the lack of structural data in existing datasets and demonstrate the superior performance and interpretability of structure-based models for this biological task.

Original authors: Yicheng Hu, Xinyu Lin, Shulin Li, Wenjie Wang, Fengbin Zhu, Fuli Feng

Published 2026-03-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yicheng Hu, Xinyu Lin, Shulin Li, Wenjie Wang, Fengbin Zhu, Fuli Feng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your body is a bustling, high-tech city. Inside every cell of this city, there are thousands of tiny workers called proteins. Each protein has a specific job: some build roads, some deliver mail, and some act as security guards.

But here's the catch: a protein can only do its job if it's in the right neighborhood. A security guard protein needs to be at the city gates (the cell membrane), while a mail carrier needs to be in the post office (the Golgi apparatus). If a protein is lost in the wrong district, the whole city's system can break down, leading to diseases.

For a long time, scientists have tried to predict where these proteins belong just by looking at their names (their amino acid sequences). It's like trying to guess where a person lives just by reading their name on a mailbox. Sometimes it works, but often it's a guess.

This paper introduces a new, revolutionary tool called CAPSUL to solve this problem. Here's how it works, explained simply:

1. The Problem: We Only Had the "Name," Not the "Face"

Previously, the best maps scientists had (like a dataset called DeepLoc) only listed the "names" of the proteins. They didn't have the 3D blueprints.

The Analogy: Imagine trying to figure out if a key fits a lock. If you only know the key's serial number (the sequence), you might guess. But if you have a 3D scan of the key's teeth (the structure), you know exactly if it fits.
The Gap: No one had a massive library of these 3D scans linked to specific neighborhoods. Without the 3D scans, the smartest AI models couldn't "see" the shape of the protein, which is often the most important clue for where it lives.

2. The Solution: CAPSUL (The Ultimate Protein GPS)

The authors built CAPSUL, a massive new database that acts like a high-definition GPS for proteins.

The 3D Map: They used a super-smart AI (AlphaFold) to generate 3D blueprints for over 20,000 human proteins. Now, instead of just a list of letters, we have the actual shape of every worker.
The Detailed Neighborhoods: Old maps grouped neighborhoods into broad categories like "Downtown" or "The Suburbs." CAPSUL is much more precise. It breaks the city down into 20 specific districts (like "The Nucleus," "The Golgi," "The Power Plant"). It even distinguishes between the "Nuclear Membrane" and the "Nucleoli" (like telling the difference between the city hall building and the mayor's office inside it).
The Verified Addresses: They didn't just guess the addresses. They cross-referenced data from two massive, trusted libraries (UniProt and the Human Protein Atlas) and only kept the addresses that were confirmed by real experiments.

3. Testing the New Map

The researchers tested this new map using various AI models:

The Old Way (Sequence-only): These models looked at the protein's "name" (sequence). They were okay, but they missed a lot of details.
The New Way (Structure-based): These models looked at the 3D "blueprint." They won. Just like having a 3D scan of a key helps you open a lock, having the 3D shape of a protein helped the AI predict its location much more accurately.

4. The "Aha!" Moment: Finding the Hidden Clues

The coolest part of the paper is how they used the AI to find new biological secrets.

The Detective Work: They asked the AI, "Which part of the protein's shape made you decide it belongs in the Golgi district?"
The Discovery: The AI pointed to a specific shape called an $\alpha$ -helix (imagine a coiled spring). It turns out, proteins with this specific "spring" shape are almost always found in the Golgi.
Why it matters: This wasn't just a computer guess. It matched what human biologists had suspected for decades through expensive lab experiments. This proves that the AI isn't just memorizing data; it's actually "understanding" the physics of how proteins work.

5. Why This Matters for You

Drug Discovery: If we know exactly where a protein lives, we can design drugs that target it more precisely, like sending a delivery drone to the exact house instead of just the neighborhood.
Understanding Disease: Many diseases happen because proteins get lost or go to the wrong neighborhood. CAPSUL helps us spot these errors faster.
The Future: This paper sets a new standard. It tells the scientific community, "Stop guessing with just names; let's use the full 3D shape to understand life."

In a nutshell: The authors built the first massive, high-definition 3D map of the human cell's neighborhoods. By using this map, they proved that looking at the shape of a protein is the key to knowing where it lives, and they even used the map to rediscover some of nature's hidden rules.

1. Problem Statement

Subcellular localization (SL) is a critical task in cell biology for drug target identification and function annotation. While it is biologically established that protein 3D structure dictates localization (e.g., nuclear localization signals are conformation-dependent), existing AI benchmarks suffer from two major limitations:

Lack of Structural Data: Widely used datasets like DeepLoc contain only amino acid sequences, making them unsuitable for evaluating structure-based models.
Coarse-Grained Annotations: Existing datasets group organelles into broad categories (e.g., "Nucleus"), failing to distinguish between specific compartments (e.g., "Nucleoli" vs. "Nuclear Membrane"), which hinders the discovery of specific biological mechanisms.

2. Methodology: The CAPSUL Benchmark

The authors introduce CAPSUL (Comprehensive humAn Protein benchmark for SUbcellular Localization), a new dataset and benchmark designed to bridge the gap between protein structure and localization prediction.

A. Dataset Construction

Source Data:
- Structures: Retrieved 20,504 human protein structures from AlphaFold2.
- Annotations: Cross-referenced with UniProt and the Human Protein Atlas (HPA).
Preprocessing & Filtering:
- Filtered for active proteins and removed fragmented AlphaFold predictions to ensure coordinate consistency.
- Final dataset size: 20,181 proteins.
Feature Extraction:
- Sequence: Amino acid sequences.
- Structure: Extracted $C_\alpha$ (alpha carbon) coordinates and generated 3Di tokens using FoldSeek to represent 3D structural patterns compactly.
Fine-Grained Categorization:
- Mapped annotations from UniProt and HPA to a unified schema of 20 distinct subcellular compartments (e.g., separating Nucleoli, Nuclear Membrane, and Nucleoplasm).
- Validated by domain experts and aligned with standard cell biology textbooks.
Evidence Levels: Annotated labels with evidence codes (Experimental vs. Author Statement vs. No Evidence) to allow researchers to filter by data reliability.

B. Experimental Setup

The authors evaluated a comprehensive suite of 11 models on CAPSUL:

Sequence-Based: DeepLoc 2.1, ESM-2 (650M), ESM-C (600M).
Structure-Based: CDConv, GearNet-Edge, FoldSeek.
Advanced/Extended Models: Graph Transformer, Graph Mamba, Graph Diffusion, CDConv with Contrastive Loss, and Fusion models (ESM-C + CDConv).
Optimization Strategies: Investigated reweighting schemes (Inverse frequency, Focal loss) and single-label classification to address class imbalance.

3. Key Contributions

First Comprehensive 3D SL Benchmark: CAPSUL is the first dataset to integrate high-quality 3D structural data (from AlphaFold2) with fine-grained, experimentally validated subcellular localization labels for human proteins.
Validation of Structural Necessity: Demonstrated that 3D structural information is critical for SL prediction, outperforming sequence-only baselines in many scenarios, particularly for underrepresented classes.
Interpretability via Attention: Showcased the ability of structure-based models to identify biologically meaningful patterns. A case study on the Golgi apparatus revealed that the model consistently attended to $\alpha$ -helix transmembrane domains, aligning with known biological mechanisms.
Strategic Optimization: Proposed and validated reweighting and single-label classification strategies to mitigate the severe class imbalance inherent in SL tasks.

4. Key Results

Performance Comparison:
- Structure-based models (specifically CDConv and GearNet-Edge) generally outperformed sequence-only models (like ESM-2) on the CAPSUL benchmark.
- ESM-C (600M) achieved the best overall sequence-based performance, highlighting the value of large-scale pre-training, but structure-based models showed superior capability in capturing spatial dependencies.
- Ablation Study: Randomly sampling $C_\alpha$ coordinates (destroying true 3D structure) caused a significant drop in performance, confirming that accurate 3D geometry is decisive for the task.
Handling Imbalance:
- Reweighting: Applying Focal Loss or inverse frequency reweighting significantly improved performance on minority classes (e.g., Lipid Droplets, Peroxisomes) for structure-based models.
- Single-Label Classification: Training separate binary classifiers for rare classes improved F1-scores for specific compartments where multi-label classification failed.
Interpretability Case Study:
- On Golgi apparatus prediction, the model identified specific $\alpha$ -helical residues (e.g., in proteins MFNG, B3GALT2, GIMAP1) with high attention scores.
- These findings matched prior experimental evidence regarding transmembrane domain topology, proving the model learns biologically causal patterns rather than just sequence correlations.

5. Significance and Future Impact

Bridging Biology and AI: CAPSUL provides a unified platform to test how structural features drive biological function, moving beyond sequence-based heuristics.
Data-Driven Discovery: The benchmark enables the discovery of novel localization patterns (e.g., specific helix orientations) that might be missed by traditional experimental screening.
Foundation for Future Research: The authors suggest future directions including:
- Aligning different structural dimensions (sequence, $C_\alpha$ , 3Di) for better representation.
- Exploring causal links between 3D structure and localization.
- Extending the pipeline to other species and context-dependent (dynamic) localization scenarios.

In conclusion, CAPSUL establishes a new standard for subcellular localization prediction, proving that incorporating 3D structural data is not just beneficial but essential for accurate, interpretable, and biologically grounded AI models in cell biology.