💻 bioinformatics

Integration of proteomic data from cell lines and tumors

The authors introduce ProtInt, a deep learning framework that successfully integrates proteomic data from cancer cell lines and patient tumors by overcoming missing value challenges, thereby outperforming existing methods and revealing key biological shifts needed to better align preclinical models with clinical reality.

Original authors: Ta, C. Q., Auth, J. M., Schilling, M., Klingmüller, U., Raue, A.

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ta, C. Q., Auth, J. M., Schilling, M., Klingmüller, U., Raue, A.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Cancer research relies heavily on a simple, powerful tool: the cancer cell line. These are groups of cancer cells grown in a laboratory dish, kept alive for decades, and used by scientists to study how tumors behave and to test new drugs. Because they are easy to grow and share, they form the backbone of preclinical studies. However, a stubborn problem has long plagued the field: findings from these dish-grown cells often fail to translate into successful treatments for actual patients. The cells in a petri dish lack the complex environment of a human body, and over time, they can drift genetically and chemically away from the tumors they were originally taken from. To bridge this gap, researchers need a way to compare the molecular makeup of these lab-grown cells directly with the molecular makeup of real patient tumors. While scientists have successfully compared the genetic instructions, or RNA, of these two groups, a similar comparison for proteins—the actual working molecules that drugs usually target—has remained out of reach. This is largely because protein data is notoriously messy, often missing large chunks of information due to the limitations of the machines used to measure them.

A team of researchers has now developed a new method called ProtInt to solve this specific puzzle. They applied this approach to a massive collection of data, combining protein measurements from 771 different cancer cell lines with data from 550 treatment-naïve patient tumors across 13 different tissue types. The goal was to create a unified map where the two distinct groups could be seen side by side, not as separate islands, but as parts of a single landscape. The researchers found that standard methods used to clean up data or align genetic information simply could not handle the unique gaps and missing values found in protein datasets. When they tried to force these older methods to work, the cell lines and tumors remained stubbornly separated. ProtInt, however, uses a sophisticated learning system that understands how proteins are measured and why data goes missing. It learns to fill in the blanks in a realistic way while simultaneously adjusting the data so that the cell lines and tumors can be compared directly.

The results showed that ProtInt successfully merged the two datasets. In the new, unified view, the artificial separation between the lab-grown cells and the patient tumors disappeared. Instead of two distinct clusters, the samples began to group together based on the type of tissue they came from, such as lung, breast, or blood. This alignment was not perfect for every single tissue type; for instance, cell lines from the central nervous system and smooth muscle still struggled to match their tumor counterparts, likely because these specific cells change significantly when grown in a dish. However, for many other tissues, the method worked remarkably well. When the researchers looked closely at what changed during this alignment, they saw a clear biological story emerge. As the cell lines were adjusted to look more like real tumors, their protein profiles shifted in predictable ways. They showed an increase in proteins related to the immune system, communication between cells, and interaction with the surrounding tissue structure. At the same time, proteins involved in basic cellular machinery, like copying genetic instructions and generating energy, decreased. This shift mirrors the reality that real tumors exist within a complex body environment, whereas lab cells are often isolated and focused purely on rapid growth.

The study also tested whether different mathematical strategies could achieve this same result. The researchers compared their method against others that rely on different learning techniques, such as those that try to force data to cycle back to its original state. They found that these alternative approaches did not perform as well; they either failed to mix the data effectively or introduced too much error. ProtInt proved superior because it specifically accounted for the way missing data behaves in protein experiments, treating the gaps not as random noise but as a pattern related to how abundant a protein is. This allowed the system to reconstruct missing values with high accuracy, ensuring that the final comparison was based on a complete picture. The researchers noted that while the method works well for the specific type of protein data they used, it is not yet ready for other common laboratory techniques that label proteins differently.

Ultimately, this work provides a new framework for making sense of complex biological data. By successfully integrating the proteomes of cell lines and tumors, ProtInt offers a way to identify which laboratory cell lines are the best matches for specific patient cancers. This could help scientists choose the most relevant models for their experiments, potentially reducing the number of failed clinical trials. The method does not claim to have solved the entire problem of cancer treatment, nor does it guarantee that every drug tested on these aligned cells will work in patients. Instead, it offers a clearer, more accurate lens through which to view the differences between the lab and the clinic. The researchers have made their code available to the scientific community, inviting others to use this tool to explore the vast, previously disconnected worlds of cell line and tumor biology.

Technical Summary: ProtInt – Integration of Proteomic Data from Cell Lines and Tumors

Problem Statement
Cancer cell lines are indispensable for preclinical research, yet findings from these models often fail to translate to clinical success due to biological discrepancies between in vitro cultures and patient tumors. While transcriptomic integration methods exist to align cell lines and tumors, comparable approaches for proteomic data are lacking. A primary barrier to proteomic integration is the prevalence of missing values, which arise from both missing-at-random (MAR) and missing-not-at-random (MNAR) mechanisms. Standard batch correction methods (e.g., ComBat, Harmony) are designed to remove technical variation between biologically similar datasets and fail to address the profound biological and technical differences separating cell lines from tumors. Furthermore, existing transcriptomic integration tools (e.g., Celligner, MOBER) do not explicitly model the intensity-dependent missingness patterns characteristic of label-free proteomics, often reconstructing missing values as zeros rather than plausible low-abundance estimates.

Methodology: ProtInt Framework
The authors introduce ProtInt, a deep learning framework designed to integrate label-free proteomic data from cell lines and patient tumors by simultaneously performing data integration and imputation. ProtInt combines three core components:

Conditional Variational Autoencoder (cVAE): The model encodes log-transformed protein intensities ( $x$ ) and a data source covariate ( $c$ , indicating cell line or tumor) into a latent embedding ( $z$ ). A decoder reconstructs the protein expression distribution. Unlike standard imputation, ProtInt follows the PIMMS approach by reconstructing only observed intensities, avoiding the introduction of noise into non-missing data.
Probabilistic Dropout Model: To address MNAR patterns, ProtInt incorporates a probabilistic dropout function ( $\Phi$ ) derived from msBayesImpute. This function models the probability of a protein being missing as a function of its intensity. The model minimizes a binary cross-entropy loss ( $L_{dropout}$ ) to align the decoder-estimated missingness with the observed missingness patterns, allowing for the generation of plausible estimates for missing values rather than defaulting to zero.
Adversarial Training: To remove the data source bias (distinguishing cell lines from tumors) while preserving biological variation (tissue of origin), ProtInt employs an adversarial neural network. This discriminator attempts to predict the data source from the latent embedding, while the encoder is trained to maximize this prediction error, thereby learning a data source-invariant latent space.

The authors evaluated alternative architectures, including cycle consistency (inspired by single-cell transcriptomic integration), but found that adversarial training outperformed cycle consistency in integrating proteomic data, particularly at higher weighting parameters, without compromising biological signal conservation.

Key Contributions

First Deep Learning Framework for Proteomic Integration: ProtInt is the first method specifically developed to integrate proteomic data across distinct biological contexts (cell lines vs. tumors) while explicitly modeling MNAR missingness.
Simultaneous Imputation and Integration: By integrating imputation within the adversarial training loop, ProtInt avoids the error propagation associated with sequential imputation followed by integration.
Benchmarking: The study provides a comprehensive benchmark comparing ProtInt against conventional batch correction (ComBat, limma, Harmony) and transcriptomic integration methods (Celligner, MOBER) on a pan-cancer proteomic dataset.

Results
The authors applied ProtInt to integrate label-free proteomic profiles from 771 cancer cell lines and 550 treatment-naïve tumor samples across 13 tissue types.

Integration Performance: ProtInt successfully removed the separation between cell lines and tumors in the joint embedding space, a feat unachieved by batch correction or transcriptomic-based methods. Quantitative metrics (graph iLISI) confirmed ProtInt achieved the highest level of dataset integration. While there was a slight reduction in biological signal conservation (graph cLISI) compared to the original data, this was attributed to the successful mixing of datasets rather than the loss of tissue-specific clustering.
Imputation Quality: The model generated plausible estimates for missing values, generally imputing lower intensities than detected values, consistent with the biological expectation that low-abundance proteins are more likely to be undetected.
Biological Insights:
- Tissue-Specific Alignment: Post-integration, cell lines from hematopoietic/lymphoid, large intestine, breast, and neuroectodermal tissues showed greater proteomic similarity to their tumor counterparts than those from CNS or smooth muscle tissues.
- Proteomic Shifts: Differential analysis of cell lines before and after "tumor projection" revealed recurrent changes. Proteins associated with immune reaction, cell-cell communication, and extracellular matrix interaction increased in abundance, reflecting the acquisition of microenvironment-like features. Conversely, proteins involved in transcription, post-transcriptional processing, and mitochondrial gene expression decreased, likely reflecting the reduced proliferation rates and altered metabolic states of tumors compared to rapidly dividing cell lines.
- Classification: When tumor samples were classified based on their nearest cell line neighbors in the ProtInt embedding, 38% were correctly assigned to their tissue of origin, with higher accuracy observed for specific tissue types (e.g., lung, breast).

Significance and Claims
The paper positions ProtInt as a framework to facilitate the joint analysis of proteomic datasets across distinct biological settings. The authors claim that ProtInt enables the identification of cell lines that best resemble patient tumors, which is critical for selecting models for clinically relevant preclinical studies.

The authors modestly note that while ProtInt improves the alignment of proteomes, the lack of matched drug response data in their current dataset prevents them from definitively proving that the aligned cell lines exhibit similar drug sensitivities to their neighboring tumors. They suggest this as a direction for future validation. Additionally, they acknowledge current limitations, noting that ProtInt is designed for label-free data and may require adaptation for label-based approaches (e.g., TMT) where normalization alters the relationship between intensity and missingness. The authors emphasize that ProtInt provides a foundation for future extensions, such as integrating single-cell proteomic data or modeling missingness patterns across different data modalities.

Technical Summary: ProtInt – Integration of Proteomic Data from Cell Lines and Tumors

More like this