🧬 biology

SCPInt: explicit disentanglement of biological and batch variation for single-cell proteomics integration

SCPInt is a deep-learning framework that integrates heterogeneous single-cell proteomics datasets by explicitly disentangling biological and batch variations, thereby outperforming existing methods in preserving biological structure while providing quantitative insights into technical biases for scalable atlas construction.

Original authors: Yadong Wang, Yuzhi Sun, Renjie Liu, Jichong Mu, Yachen Yao, Liyuan Zhang, Tianyi Zhao

Published 2026-07-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yadong Wang, Yuzhi Sun, Renjie Liu, Jichong Mu, Yachen Yao, Liyuan Zhang, Tianyi Zhao

Original paper licensed under CC BY 4.0 (https://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 you are trying to build a massive, perfect library of every single type of worker in a human body. To do this, you need to look at individual cells (the workers) and see what tools they are carrying (their proteins).

However, there's a problem: The tools are hard to count.

Unlike counting apples in a basket (which is like counting RNA in cells), measuring proteins in a single cell is like trying to weigh individual grains of sand using a scale that is a bit wobbly. Different labs use different scales, different ways of preparing the sand, and different times of day. This creates "noise" or "static" that makes it look like the workers are different when they are actually the same, or makes it hard to see how they change over time.

This paper introduces a new tool called SCPInt to fix this mess. Here is how it works, using simple analogies:

1. The Problem: The "Noisy Room"

Imagine you are at a party where people from different cities are mixing.

The Biology: The people are the cells. Their "personality" (what job they do) is their biology.
The Batch Effect: The "city" they came from is the batch. People from City A might all wear red hats because that's the local fashion. People from City B wear blue hats.
The Issue: If you just look at the room, you can't tell if someone is a "Doctor" or a "Teacher" because you are too distracted by whether they are wearing a red or blue hat. Existing computer programs try to fix this by just "erasing" the hat color, but they often accidentally erase the person's job title too, or they fail to mix the groups properly.

2. The Solution: SCPInt (The "Smart Translator")

SCPInt is a smart computer program that acts like a translator who can separate the person from the hat.

It uses a special technique called "Explicit Disentanglement." Think of it like a magic sorting machine with two conveyor belts:

Belt A (Biological Embedding): This belt carries only the person's identity (their job, their health, their story). The program forces the computer to strip away all the "hat" information here.
Belt B (Batch Embedding): This belt carries only the hat information (the city, the scale used, the freezer vs. fresh sample). The program forces the computer to strip away the person's identity here.

By separating them into two distinct piles, the computer can:

Rebuild the party: It puts everyone on Belt A together, regardless of their hat, so Doctors can talk to Doctors, and Teachers to Teachers, even if they came from different cities.
Analyze the hats: It keeps Belt B separate so scientists can study why the red hats are different from the blue hats (e.g., "Oh, the red hats are from a freezer, and that changes the fabric slightly").

3. How It Handles the "Sand" (The Math)

Most computer programs assume data is like counting apples (discrete numbers). But protein data is like measuring the weight of sand (continuous numbers) and often has holes where sand is missing.

The Analogy: SCPInt doesn't try to count the sand grains. Instead, it uses a "Two-Component Gaussian Mixture." Imagine it knows that the sand is actually a mix of two types of piles: one pile of "real sand" and one pile of "empty space." It models this mix perfectly so it doesn't get confused by the missing grains.

4. What They Found (The Results)

The authors tested SCPInt on real data from different labs and found:

It builds better maps: When they combined data from different studies about the human brain, SCPInt successfully drew a smooth map showing how brain cells grow from babies to adults. Other tools either broke the map or mixed up the different stages.
It finds hidden states: They looked at immune cells (macrophages). Some were "calm," and some were "angry" (activated by infection). SCPInt could separate the calm from the angry cells perfectly, even when they came from different labs. Other tools either mixed them all up or erased the "angry" signal.
It measures the "Static": Because SCPInt keeps the "hat" information in a separate pile, they could actually measure how much the "freezing" of a sample changed the data compared to "fresh" samples. They found that freezing affects some cell types (like cartilage) much more than others.

Summary

SCPInt is a new way to clean up and combine messy protein data from single cells. Instead of just trying to "fix" the errors, it splits the data into two parts: the true biological story and the technical noise. This allows scientists to build larger, more accurate maps of the human body and also understand exactly how their experiments (like freezing samples or using different machines) change the results.

Key Takeaway: It doesn't just hide the noise; it separates the signal from the noise so you can study both of them clearly.

Technical Summary: SCPInt – Explicit Disentanglement of Biological and Batch Variation for Single-Cell Proteomics Integration

Problem Statement

Mass spectrometry-based single-cell proteomics (SCP) offers a direct measurement of cellular functional states, overcoming limitations of sequencing-based methods like CITE-seq that capture less than 1% of the proteome. However, MS-based SCP data is inherently low-throughput, necessitating the integration of multiple datasets to support atlas-scale studies. Current integration methods (e.g., Harmony, scVI, Scanorama) are primarily designed for count-based sequencing data (scRNA-seq) and fail to adequately model SCP data, which exhibits continuous distributions, heterogeneous missing structures, and substantial technical variation arising from diverse platforms and sample preparation protocols. Furthermore, existing tools lack the capability to explicitly disentangle and quantify the complex batch effects specific to MS-based SCP, hindering comprehensive downstream analysis and the construction of unified proteomic atlases.

Methodology: The SCPInt Framework

SCPInt is a deep-learning framework designed to learn a unified biological representation while explicitly separating technical variation. It addresses two core challenges: modeling the specific distribution of MS proteomic data and disentangling biological signals from batch effects.

1. Architecture and Components

The model operates on an input protein expression matrix and batch labels, generating two distinct latent embeddings for each cell: a biological embedding and a batch embedding.

Encoder: A shared multilayer perceptron (MLP) extracts global features from the input expression matrix. These features are projected into two separate heads: a biological head and a batch head, producing the respective embeddings.
Batch-Conditioned Gaussian Mixture (GMM) Decoder: Unlike standard autoencoders that assume a single Gaussian distribution, SCPInt utilizes a two-component Gaussian mixture model tailored to the zero-inflated and continuous nature of SCP data. The decoder takes the concatenation of the biological embedding and the one-hot batch label vector as input. It reconstructs the protein expression matrix by predicting the mixture weight ( $\pi$ ), mean ( $\mu$ ), and variance ( $\sigma$ ) for each protein. This design ensures the biological embedding is not confounded by batch information, as the reconstruction relies on the explicit batch input.
Adversarial Disentanglement:
- Batch Classifier: A classifier predicts the batch label from the batch embedding, encouraging this embedding to fully capture batch-associated information.
- Batch Discriminator: A discriminator attempts to predict the batch label from the biological embedding. Crucially, a Gradient Reversal Layer (GRL) is inserted between the biological head and the discriminator. During backpropagation, the GRL reverses the gradient, forcing the encoder to minimize the discriminator's ability to predict the batch from the biological embedding. This adversarial process explicitly removes batch-specific information from the biological representation.

2. Loss Function

The total loss is a weighted sum of four components:

GMM Reconstruction Loss: Maximizes the likelihood of the observed data under the two-component Gaussian mixture model.
Mean Squared Error (MSE) Loss: An auxiliary loss applied to the mixture mean to preserve the global protein expression profile.
Batch Classification Loss: Cross-entropy loss on the batch embedding to ensure it retains batch information.
Adversarial Loss: Cross-entropy loss on the biological embedding (via the discriminator) to remove batch information.

Key Contributions

Distribution-Aware Modeling: SCPInt is the first integration method specifically designed for MS-based SCP data, utilizing a two-component Gaussian mixture model to address the continuous distribution and zero-inflation inherent to proteomics, rather than relying on count-based assumptions.
Explicit Disentanglement: The framework moves beyond implicit batch correction. By generating separate biological and batch embeddings, it allows for the simultaneous preservation of biological structure and the explicit quantification of technical variation.
Quantitative Technical Phenotypes: The batch embeddings serve as "quantitative technical phenotypes," enabling the measurement of platform bias, cryopreservation effects, and cell-type-specific technical sensitivity.

Results

SCPInt was evaluated across six integration tasks comprising 18 datasets from eight independent studies, spanning multiple technologies, developmental stages, and sample conditions.

Benchmarking on Heterogeneous Datasets: In tasks involving multiple batches from single studies (e.g., Huffman et al.) and cross-study integration (e.g., Derks, Leduc, Specht), SCPInt consistently outperformed state-of-the-art methods (Harmony, scVI, Scanorama, LIGER, scProtein). It achieved a superior balance between biological conservation (preserving cell type separation) and batch correction (mixing cells from different batches).
Atlas Construction and Developmental Continuity: Applied to human brain SCP data across developmental stages (GW13, GW15, GW19), SCPInt successfully integrated the data while preserving continuous developmental manifolds. It enabled the reconstruction of proteomic developmental trajectories from radial glia to excitatory neurons, a feat where other methods failed to maintain continuity or cell-type separation.
Cellular State Identification: In macrophage datasets, SCPInt preserved distinct biological states (e.g., LPS-treated vs. untreated) while removing batch effects. It identified biologically meaningful "Normal" and "Inflammatory" states, with pathway enrichment analysis confirming that the identified states correlated with Type I IFN signaling and metabolic reprogramming.
Quantification of Technical Variation:
- Platform Bias: Using batch embeddings, SCPInt quantified the separation between three MS platforms (pSCoPE, SCoPE2, plexDIA). The results aligned with known technical principles, showing pSCoPE and SCoPE2 (TMT-based) were closer to each other than to plexDIA (DIA-based).
- Sample Preparation Effects: In fresh vs. frozen tissue data, SCPInt revealed that while platform differences accounted for the majority of variance, cryopreservation effects were cell-type-specific. For instance, chondrocytes showed a larger fresh-to-frozen shift than other cell types, and the direction of this shift varied across cell types, indicating structured technical variation rather than random noise.

Significance and Claims

The paper claims that SCPInt shifts the paradigm of single-cell proteomics integration from implicit batch-effect removal toward a unified biological representation coupled with explicit, interpretable, and quantitative analysis of technical variation.

By providing a foundation for scalable atlas construction, SCPInt addresses the fragmentation of current MS-based SCP datasets. The ability to generate quantitative technical phenotypes allows researchers to systematically evaluate experimental workflows, instrumentation, and sample preparation strategies. The authors assert that this capability supports more reproducible atlas construction and the optimization of SCP workflows, particularly in settings where fine-grained metadata is unavailable.

Limitations Acknowledged:
The authors note that SCPInt currently relies on predefined batch labels and does not identify the causal experimental factors behind technical shifts. Additionally, the framework is designed for expression-matrix integration and does not model spatial coordinates or imaging-derived features, limiting its direct application to spatially resolved SCP technologies like scDVP.