Unraveling the complex relationship between mRNA and protein abundances: a machine learning-based approach for imputing protein levels from RNA-seq data

Abstract

The correlation between messenger RNA (mRNA) and protein abundances has long been debated. RNA sequencing (RNA-seq), a high-throughput, commonly used method for analyzing transcriptional dynamics, leaves questions about whether we can translate RNA-seq-identified gene signatures directly to protein changes. In this study, we utilized a set of 17 widely assessed immune and wound healing mediators in the context of canine volumetric muscle loss to investigate the correlation of mRNA and protein abundances. Our data reveal an overall agreement between mRNA and protein levels on these 17 mediators when examining samples from the same experimental condition (e.g. the same biopsy). However, we observed a lack of correlation between mRNA and protein levels for individual genes under different conditions, underscoring the challenges in converting transcriptional changes into protein changes. To address this discrepancy, we developed a machine learning model to predict protein abundances from RNA-seq data, achieving high accuracy. Our approach also effectively corrected multiple extreme outliers measured by antibody-based protein assays. Additionally, this model has the potential to detect post-translational modification events, as shown by accurately estimating activated transforming growth factor β1 levels. This study presents a promising approach for converting RNA-seq data into protein abundance and its biological significance.

Introduction

The way organisms interact with their environment and undergo development and tissue regeneration is governed by their genomic DNAs. These DNA sequences are then transcribed into messenger RNAs (mRNAs), which are subsequently translated into proteins. Many of these proteins undergo further post-translational modifications to enhance their cellular functions. Based on this central dogma of molecular biology, it is widely accepted that protein levels are primarily determined by the abundance of their respective mRNAs. However, the context-dependent concordance between mRNAs and proteins has been discussed by many studies, yet still remains a point of debate (1–6). Over the past decade, RNA sequencing (RNA-seq) has become a powerful and widely used technology to detect and quantify mRNA molecules (7,8). This high-throughput approach has become the standard for analyzing transcriptomic dynamics and identifying potential biomarkers in developmental processes and diseases. Despite the extensive use of RNA-seq in investigating transcriptional changes, the critical question still remains: Can the transcriptional biomarkers and dynamic patterns observed through RNA-seq be translated into corresponding changes at the protein level?

There are two types of mRNA–protein correlations. The first type is the mRNA–protein correlation at the whole-genome scale: This type of correlation aims to understand whether high-abundance genes (genes with high mRNA expression levels) tend to have high-abundance proteins under the same experimental conditions. Most mRNA–protein concordance studies were trying to address this question and explore the overall agreement or discrepancy between mRNA and protein levels on a global scale (9). The second type of mRNA–protein correlation focuses on individual genes. It examines how changes in mRNA expression levels are related to corresponding changes in protein expression levels under varying conditions. The second type of mRNA–protein correlation focuses on individual genes and the relationship between changes in mRNA expression and the corresponding changes in protein expression levels under varying conditions. This second type of correlation, which examines the specific relationship between mRNA and protein levels for a particular gene under different experimental conditions, remains relatively underexplored. This type of correlation addresses questions such as whether changes in mRNA expression (upregulation or downregulation) can translate directly to corresponding changes in protein level or whether additional regulatory mechanisms are involved in the process. Understanding this type of correlation is crucial for deciphering the intricate and context-dependent nature of gene expression regulation.

Despite a wealth of RNA-seq and proteomics data available, particularly for humans and mice, many of these datasets either are unpaired, with RNA-seq and proteomics conducted in different contexts, or cover a limited set of conditions. Tackling this issue is technically challenging. Merging mRNA–protein data from diverse conditions across multiple studies can introduce significant batch effects (10). Since different mRNA and protein assays have different dynamic ranges and technical biases, combining multiple datasets might amplify the inconsistencies between mRNAs and proteins. However, producing a paired mRNA and protein dataset across a broad range of conditions, such as varying individuals and time points, is both time-consuming and expensive.

In the present study, using a recently developed canine volumetric muscle loss (VML) wound healing model (11), we conducted time-series experiments with RNA-seq to measure mRNA levels and Luminex™/ELISA (enzyme-linked immunosorbent assay) to measure 17 immune and tissue regeneration mediators at the protein level, aiming to evaluate the concordance between mRNA and protein expressions. We found that given a particular biopsy, the mRNA and protein abundances generally correlate well for these 17 mediators. However, a particular gene’s mRNA and protein levels at different time points are generally not correlated. We reasoned that this mRNA and protein abundance discrepancy is due to the complexity of translational control at the protein level across conditions.

Emerging evidence indicates that most of the transcript abundance can be computationally inferred by the expression levels of a small representative gene list (gene signatures) (12). For example, 81% human transcript abundance can be computationally imputed by the expression levels of 19% transcripts (12). Given that RNA-seq is the most commonly used high-throughput method, in the present study we asked whether the protein abundance can also be computationally imputed by a set of signature gene expression patterns in a specific context (e.g. wound healing). In the present study, we developed a machine learning model (support vector regression, SVR) to predict protein abundance from transcriptional expression data (RNA-seq). We successfully imputed all of the 17 protein abundances with high correlations with Luminex™ and ELISA measured protein abundance. We further demonstrated that the RNA-seq-based imputing model can potentially detect post-translational modification events, as shown by estimating activated transforming growth factor (TGF)-β1 levels.

Our study demonstrates that a combination of RNA-seq data and machine learning models can provide estimates for both protein levels and events related to post-translational modifications. This extends the application of RNA-seq beyond just analyzing transcriptomic alterations and allows for correlation with proteomic changes and post-translational modifications.

Materials and methods

Canine VML model

The in vivo studies described herein were approved by the Institutional Animal Care and Use Committee at the University of Pittsburgh. A VML defect was created in the quadriceps muscles of 10 female dogs. Tissue biopsy samples were collected at various intervals from days 1 to 42 following injury. The dogs were prepared for surgery by sedation, shaving of the surgical site, scrubbing with 70% ethyl alcohol, followed by a topical 10% povidone-iodine solution and induction of surgical plane anesthesia with 2% thiopental sodium, intubation and maintenance with 1–3% isoflurane. A 10 cm long × 4 cm wide area was defined, and an initial incision point at the greater trochanter that extended to the lateral epicondyle. The subcutaneous tissue was bluntly dissected, and the skin retracted to visualize the biceps femoris. The underlying muscle tissue and fascia were surgically excised to a depth of 2–3 cm. The depth of excision depended upon individual dog anatomy. This procedure results in the removal of between 50% and 70% of total muscle mass loss. The overlying skin was removed to leave an open wound that mimicked trauma-induced VML in the clinical setting. The upper leg and wound site were covered with sterile dressings and all animals were administered antibiotic prophylaxis for the first 5 days post-surgery (cephalexin, 25 mg/kg) and analgesic every 8–12 h for the first 5 days (buprenorphine, 0.002 mg/kg). The overall health of each dog, including temperature, appetite and activity levels, was monitored daily. The dogs were fed a high-energy, high-protein diet (Advanced Protocol High-Density Canine Diet; PMI Nutrition LLC, Henderson, CO, USA) and provided unlimited access to water. Additional details can be found in a separate report (11).

Biopsy samples were collected from the center, intermediate and edge of the wound bed (see Supplementary Figure S1 for details) at various days ranging from days 1 to 42 post-surgery. Detailed biopsy sample information used in this study can be found in Supplementary Table S1.

Ethical statement

The animal protocols were approved by the Institutional Animal Care and Use Committee through the Animal Research Protection Office (https://www.iacuc.pitt.edu/about). We hold AAALAC certification, possess Animal Welfare Assurance through OLAW [D16-00118 (A3187-01)] and are approved by the United States Department of Agriculture (USDA; Certificate No. 23-R-0016, Customer No. 289). Our work with canines was governed by both the Welfare Assurance and USDA regulations. Dr Deborah L. Chapman is the Chair of the Animal Protocol Committee. The reviewers of any individual protocol are anonymous.

RNA-seq of canine VML biopsy

Total RNAs were isolated from dog biopsies using TRIzol (Thermo Fisher, #15596018) and chloroform phase separations followed by the RNeasy mini protocol (Qiagen, #74106) with optional on-column DNase digestion (Qiagen, #79254). One hundred nanograms of total RNA was used to prepare sequencing libraries using the ligation-mediated sequencing protocol (7). RNAs were selected using the NEBNext Poly(A)+ Isolation Kit (NEB, #E7490S/L). Poly(A)+ fractions were eluted, primed and fragmented for 7 min at 85°C. First-strand complementary DNA (cDNA) synthesis was performed using SmartScribe Reverse Transcriptase (Takara Bio USA, #639538) and RNA was removed. cDNA fragments were purified with AMPure XP beads (Beckman Coulter, #A63881). The 5′ adapter was ligated, and 18 cycles of amplification were performed. These final indexed cDNA libraries were quantified, normalized, multiplexed and run as single-end reads on the HiSeq 3000 (Illumina, San Diego, CA, USA). RNA-seq reads were mapped to the dog genome and annotated protein-coding genes (version: canFam3) using Bowtie (v0.12.8) (13), allowing up to two mismatches. The gene expected read counts and transcripts per million (TPMs) were estimated by RSEM (v1.2.3) (14). The TPMs were further normalized by EBSeq (15) R package to correct the potential batch effect.

Protein assay to measure protein abundance

A set of inflammatory mediators was measured in wound tissue samples using a Luminex™ 100 IS apparatus (Luminex, Austin, TX, USA) and the canine-specific Luminex™ 13-plex bead set (Millipore), a separate 4-plex kit (Invitrogen/Themo Fisher) and ELISAs for detecting additional mediators as follows: GM-CSF, IFN-γ, IL-2, IL-6, IL-7, IL-8, IL-15, IL-10, CXCL10 (IP-10), IL-18, CCL2 (MCP-1) and TNF-α (Luminex™, Millipore); β-NGF and VEGFA (Luminex™, Invitrogen/Themo Fisher); IL-33 (ELISA, CusaBio); and BDNF and TGF-β1 (ELISA, R&D) (11,16). To activate latent TGF-β1 to immunoreactive TGF-β1 detectable by the Quantikine^® TGF-β1 immunoassay, 20 μl of 1 N HCl was added to each 40 μl of plasma, mixed and incubated for 10 min at room temperature. Next, the acidified samples were neutralized by adding 20 μl of 1.2 N NaOH/0.5 M HEPES, mixed and then diluted with a calibrator diluent prior to the assay. To perform the assay, 50 μl of diluent RD1-73 was added to each well followed by the addition of 50 μl of recombinant human TGF-β1 standard, control or activated sample per well and incubated for 2 h at room temperature. Following four aspirations/wash steps, 100 μl of TGF-β1 conjugate was added to each well and incubated for an additional 2 h at room temperature, washed and then 100 μl substrate solution was added to each well and incubated for 30 min at room temperature. Finally, 100 μl of stop solution was added to each well, and the optical density (wavelength set at 450 nm) was determined, as previously described (16).

SVR to impute the protein abundance from RNA-seq data

SVR (17) is a type of machine learning algorithm that falls under the category of support vector machines. It is primarily used for regression tasks, where the goal is to predict continuous outcomes. We chose SVR as our machine learning model to impute protein abundance from RNA-seq data because SVR is one of the most flexible and robust prediction algorithms (17–19). SVR aims to find a function that approximates the relationship between input variables and output variables with a certain tolerance (ϵ). Unlike traditional regression, SVR does not minimize the error between the predicted and actual values; instead, it focuses on fitting the error within a certain threshold. The primary goal of SVR in the context of using RNA-seq data to impute protein abundance is to develop a function that can accurately map the relationship between the RNA-seq gene expression data (input) and the corresponding protein abundance (output) within a certain tolerance (ϵ). This is achieved by fitting a hyperplane or a set of hyperplanes in a high-dimensional space. The key to SVR’s adaptability to handle nonlinear correlations is the use of the kernel function, such as the radial basis function (RBF) kernel. The RBF kernel function, represented as , is to transform the input data into a higher-dimensional space. Here, x and are input vectors, and γ is a parameter that defines the kernel’s behavior. This transformation facilitates the fitting of a linear hyperplane in this new space, enabling the handling of nonlinear relationships. SVR also utilizes support vectors, which are data points most critical to defining the hyperplane’s position. These points are the ones that lie closest to the hyperplane, having the most influence on its orientation and position. The regularization parameter in SVR, denoted as C, plays a vital role in balancing the model’s complexity and its performance on unseen data. A higher value of C implies giving more importance to minimizing errors, potentially leading to a more complex model, while a lower C might result in a simpler model with higher bias. SVR does not just aim to minimize the prediction error; it seeks to fit the prediction error within a predefined threshold, making it suitable for biological data, which often contains inherent variability and noise. Specific technical details include the following:

• (a) SVR algorithm: SVR is a machine learning algorithm to solve nonlinear regression problems. Given a training dataset , where x_i is an input vector (e.g. features) and y_i is the observed value, SVR mapped x into a higher-dimensional feature space F via a nonlinear mapping , and then linear regression is performed in this space. In SVR, the goal is to find a function f(x) that has at most ϵ deviation from the actual target values y_i for all the training data, and at the same time is as flat as possible. This is achieved by minimizing the following objective function:

  

Here, w represents the weights of the hyperplane and  is the regularization term. It is used to prevent overfitting by penalizing large values of the weight vector w. Minimizing this term helps to ensure that the model is as simple as possible, thereby reducing the risk of overfitting. The second term, , involves a sum over a loss function for each data point in the training set. C is the regularization parameter, and  and are the slack variables that measure the degree of a misfit. This is where the concept of ϵ-insensitive loss comes into play in SVR. The loss function  measures the deviation of the predicted values from the actual values, but only if this deviation is greater than a specified threshold ϵ. The constant C is a regularization parameter that balances the trade-off between the flatness of the model (minimizing ) and the amount up to which deviations larger than ϵ are tolerated. The optimization is subject to the constraints

where  is the feature transformation function (applied by the kernel function), b is the bias term and is the margin of tolerance. In the context of using RNA-seq data to impute protein abundance, the optimization problem in SVR with a kernel function is transformed into a quadratic programming problem through the introduction of Lagrange multipliers (20,21):

where K is the kernel function . The RBF kernel was used in this study:

Here, x represents a new data point for regression and x_i represents a data point from the training set. The function  represents the decision function in the SVR model, where and are the Lagrange multipliers associated with each data point in the training set. These multipliers play a crucial role in determining the support vectors, which are the critical data points that define the decision boundary, or hyperplane, in the SVR model. The kernel function K is at the heart of the SVR’s ability to handle nonlinear relationships. In this case, the RBF kernel is used, defined as , with γ being a positive parameter. The gamma () parameter controls the shape of the decision boundary, with a high value of gamma resulting in a more complex and wiggly boundary. This kernel function effectively measures the similarity between any two data points and in the input space, with γ controlling the width of the kernel and, consequently, the degree of nonlinearity in the model. In the specific scenario of predicting protein abundance from RNA-seq data, the gene expression values (obtained from RNA-seq) are the input features x, while the protein abundance is the target output. The RBF kernel’s ability to handle the complex and nonlinear relationships that often exist between gene expression and protein levels makes it particularly suitable for this task.

All parameters, including γ and C, were tuned by R package ‘caret’ with function ‘train control’ via searching an optimized parameter combination set to maximize the prediction accuracy.

• (b) Feature selection: In our feature selection process, we started with the 15 388 protein-coding genes annotated in the Ensembl canine genome (version: canFam3). Our first step is to exclude any genes that are not expressed under any conditions. After this, we focused on identifying and addressing potential multicollinearity issues within our features (genes). Multicollinearity, where several independent variables in a model are correlated, is a potential problem in regression models. When independent variables are correlated, it means that changes in one variable are associated with shifts in another variable. It becomes difficult for the model to estimate the relationship between each independent variable and the dependent variable independently because the independent variables tend to change simultaneously. To solve this problem, we performed a pairwise Spearman’s correlation between the expressed genes. The multicollinearity-removing procedure starts with identifying gene pairs that exhibit the highest correlation, as determined by Spearman’s correlation coefficient (Rho). Among these pairs, the gene that shows the greatest correlation (positively or negatively) with any other gene is eliminated. For instance, consider two genes, Gene A and Gene B, which have an absolute correlation coefficient (Rho) of 0.99, indicating a strong correlation. However, if Gene A also shows a higher correlation (e.g. Rho of 0.95) with another gene, like Gene C, then Gene A is considered redundant and is removed from the selection. This is because Gene A’s expression pattern is more closely mirrored by Gene C than by Gene B. On the other hand, if Gene B does not show such a high correlation with any other genes, it remains in the selection pool. The procedure is iteratively conducted until the remaining genes in the pool have pairwise absolute Spearman’s correlation coefficients (Rho) <0.75. This approach was performed by the R package ‘caret’. We selected genes having a mutual Spearman’s absolute Rho of ≤0.75, resulting in the selection of 4598 nonredundant gene patterns for subsequent modeling.

• (c) Model evaluation: We applied 5-fold cross-validation to impute protein abundance from RNA-seq data, ensuring a robust evaluation of the model without overlapping training and testing sets. This approach segments the dataset into five equal ‘folds’. The model undergoes five iterations of training and testing, each utilizing a different fold as the validation set and the remaining four for training. This procedure enhances the model’s generalizability and provides a more resilient estimate of its predictive capacity on unseen data. The imputed protein abundance was subsequently validated against measurements obtained from antibody-based protein assays using Spearman’s rank correlation analysis.

• (d) Feature importance ranking: Feature importance in machine learning is a method employed to gauge the impact or role of individual features in a model’s ability to make predictions. In the context of our SVR model, the significance of different genes is evaluated by ranking them according to the changes in performance when the gene expression of that particular gene is permuted. This ranking is carried out through the ‘VarImp’ function in the R package ‘caret’.

Protein–protein interaction enrichment analysis of model-selected features

We conducted an analysis to determine whether the top 100 features selected by a model for protein-level imputation are more likely to interact with the imputed proteins as opposed to other background genes. Specifically, for each imputed protein, we assessed the enrichment of genes whose translated proteins directly bind to that imputed protein among the top 100 model-selected features (based on their contribution to the model). We used the STRING database to obtain the protein–protein interaction network.

For each imputed protein, we constructed a 2-by-2 table as follows:

• (a) Number of top 100 model-selected features that directly connect or bind to the imputed protein.

• (b) Number of top 100 model-selected features that do not directly connect or bind to the imputed protein.

• (c) Number of background genes (those with Spearman’s absolute Rho <0.1 compared to observed protein levels) that directly connect or bind to the imputed protein.

• (d) Number of background genes (with the same Spearman’s criterion) that do not directly connect or bind to the imputed protein.

We utilized the Fisher’s exact test to calculate the statistical significance of this enrichment analysis.

Gene Ontology analysis

The Gene Ontology (GO) enrichment analysis was performed by the R package ‘allez’ (22). The P-values were further adjusted by Benjamini–Hochberg multiple-test correction (23). P-values <0.05 were considered statistically significant.

Results

The relationship between mRNA levels and protein abundance

We utilized an experimental canine VML wound healing model to explore the link between mRNA levels and corresponding protein abundance for a specific set of 17 genes: CSF2, IFN-γ, IL-2, IL-6, IL-7, IL-8, IL-15, IL-10, CXCL10, IL-18, CCL2, TNF-α, IL-33, TGF-β1, BDNF, NGF-β and VEGFA. This study involved 77 biopsy samples taken from seven canine individuals at seven different time points (post-injury days 1, 2, 3, 4, 7, 14 and 42) and from three distinct wound zones (center, edge and intermediate) [see the ‘Materials and methods’ section and the canine VML model (11)]. These samples were analyzed using paired RNA-seq and Luminex™/ELISA.

Initially, we examined whether the protein abundance of the 17 genes was rank correlated with their mRNA levels. We calculated Spearman’s rank correlations between mRNA levels and protein abundance for each biopsy. The median Spearman’s correlation coefficient (Rho) across all 77 biopsies was found to be 0.67, suggesting a positive relationship between mRNA and protein levels under the given experimental conditions (Figure 1A). Specific examples of individual cases are also presented in Figure 1B–D, further illustrating this correlation.

Figure 1.

Correlations between mRNA expression and protein levels for 17 genes (CSF2, IFN-γ, IL-2, IL-6, IL-7, IL-8, IL-15, IL-10, CXCL10, IL-18, CCL2, TNF-α, IL-33, TGF-β1, BDNF, NGF-β and VEGFA) within a biopsy. (A) Spearman’s rank correlation coefficient (Rho) distribution across 77 biopsy samples. (B–D) Examples of mRNA and protein correlations for the 17 genes within an individual biopsy.

Subsequently, we investigated the correlation between mRNA and protein levels for each gene under different conditions. Surprisingly, only two genes (IL-6 and MCP-1) showed a moderate concordance of abundance between their mRNAs and proteins, with Spearman’s rank correlation coefficients (Rho) >0.4 (Figure 2). The remaining 15 genes demonstrated either weakly correlated patterns (e.g. IL-8 and VEGFA; Rho > 0.3) or no correlations at all (Figure 2). These results indicate that mRNA and protein levels are largely inconsistent across different conditions. These results further suggest that transcriptomic changes identified through RNA-seq may not necessarily reflect corresponding changes at the protein level. The relationship between mRNA and protein abundances appears to be complex and can vary depending on the gene and experimental conditions.

Figure 2.

Spearman’s rank correlation between mRNA expression and protein levels for each gene under different conditions. Among 17 gene–protein pairs, only IL-6 and MCP-1 show moderately correlated mRNA–protein levels with Spearman's correlation coefficient (Rho) >0.4.

Imputing protein abundance from RNA-seq data

Recent studies indicate that despite the vast number of protein-coding genes in the human genome (∼20 000), the mRNA expression patterns of any given gene can be predicted by the expression patterns of a limited set of signature genes (12). Building upon this observation, we hypothesized that protein abundance could also be imputed using a group of mRNA expression patterns (referred to as mRNA signatures), though not necessarily the exact mRNA responsible for translating into its ultimate protein. To test this hypothesis, we developed an SVR model, trained on mRNA expression levels estimated from RNA-seq data (see the ‘Materials and methods’ section), to predict protein abundance.

Figure 3 demonstrates that the abundance of all 17 proteins can be imputed from RNA-seq data. Notably, the median Spearman’s correlation coefficient (Rho) between the protein levels measured by Luminex™/ELISA and those imputed from RNA-seq data is 0.94 by 5-fold cross-validation. Among the 17 proteins, IP-10 (gene name: CXCL10) protein levels were predicted with the highest accuracy (Spearman’s Rho = 0.99), while IFN-γ protein was predicted with the lowest prediction accuracy (Spearman’s Rho = 0.78). These results suggest the potential of using RNA-seq data in conjunction with a machine learning model to impute protein abundance. The high correlation coefficients between imputed protein levels and experimental protein measurements suggest that this approach holds promise for protein abundance prediction and opens up exciting possibilities for future research in this field. To assess whether the alterations in estimated protein abundance exhibit similarity to the changes observed in protein abundance measured through Luminex™/ELISA, we conducted a principal component analysis (PCA) to identify key factors, such as post-injury time or wound zones, responsible for the overall variations. As demonstrated in Supplementary Figure S2, there are minimal differences between different wound zones, with significant overlap, whereas the post-injury time predominantly dominates the overall trends in protein abundance (PCA performed based on Luminex™/ELISA measured protein abundance). This implies that the dynamic changes in protein levels are primarily influenced by post-injury healing time rather than wound zones. Consequently, we grouped different wound zones at the same time points to calculate the fold changes relative to the average expression of day 1. As shown in Supplementary Figure S3, the fold changes computed on imputed data largely mirror the fold change based on Luminex™/ELISA.

Figure 3.

Spearman’s rank correlation between RNA-seq imputed and Luminex™/ELISA measured protein abundance. The imputed protein abundance was based on the SVR model via 5-fold cross-validation.

Imputed protein abundance is robust to outliers

Imputed protein abundance is based on patterns of multiple genes rather than relying on a single gene/protein measurement. Hence, we reason that it may inherently resist the influence of stochastic outliers, because the effect of extreme values (outliers) in one gene can be mitigated by considering the expression patterns of other genes when combining them for protein abundance imputation. To evaluate this hypothesis, we investigated whether the protein outliers (defined as measured protein abundance exceeding a 5-fold deviation from the mean of replicate measurements) measured by Luminex™/ELISA can be mitigated by RNA-seq-based imputing model. Out of 17 proteins, 7 (BDNF, GM-CSF, IL-6, IL-10, IL-15, MCP-1 and TNF-α) had at least one outlier according to Luminex™/ELISA measurements. Interestingly, these outliers were all corrected when assessed through the RNA-seq-based imputation model. For example, TNF-α (Figure 4A), GM-CSF (Figure 4B) and MCP-1 (Figure 4C) contained multiple outliers in the Luminex™ assay measurements, all of which were successfully corrected by our imputed model (Figure 4). These findings indicate that our imputing model is robust to technical variations. However, we acknowledge that outliers in protein data are not inherently ‘incorrect’. Therefore, even though our RNA-seq imputing model indicates that these outliers might be artifacts, we will approach them with caution and further explore their biological relevance and potential significance.

Figure 4.

RNA-seq imputed protein abundance is robust to extreme outliers. (A–C) Examples of RNA-seq corrected extreme outliers.

The predominant features of the imputing model can be partially attributed to potential physical interactions

To assess whether the model-selected features for protein-level imputation are more likely to interact with the imputed proteins than other genes, we conducted a protein–protein interaction enrichment analysis for each imputed protein. We examined the ratio of predominant features (top 100 major contributors to the model) that exhibit direct binding evidence with their translated proteins to the imputed protein, utilizing data from the STRING database. We then compared this ratio to other background genes showing direct binding evidence to the imputed protein. Among the 17 imputed proteins, 5 of them, namely IL-2, IL-6, GM-CSF, TNF-α and MCP-1, exhibited significant enrichment P-value (P-value <0.05, Fisher’s exact test). As an example, we found that 32 out of the top 100 predominant features utilized for imputing IL-6 protein levels were direct binding proteins based on the STRING database (Figure 5A). In contrast, only 920 out of 5479 background genes (whose mRNA levels did not correlate with IL-6 protein expression) were direct IL-6 binding proteins. This means that there was a 2-fold increase of the ratio of IL-6 binding proteins among the top 100 predominant features used for imputation compared to background genes. The enrichment was statistically significant with an enrichment P-value of 1.68 × 10⁻⁴, as determined by Fisher’s exact test. Figure 5B shows the IL-6 local protein–protein interaction network that includes the 32 genes selected by the model to impute the IL-6 protein abundance. The detailed enrichment fold changes and P-values for IL-2, IL-6, GM-CSF, TNF-α and MCP-1 can be found in Supplementary Table S2. This indicates that physical interaction is a potential factor partially contributing to the protein-imputing model.

Figure 5.

Using IL-6 as an illustrative example to show that features selected by the protein-level imputing model exhibit a higher likelihood of interaction with the imputed proteins compared to other genes. (A) Protein–protein binding enrichment analysis. The top 100 features ranked by relative contribution to the imputing model are statistically enriched in proteins with potential binding evidence to the imputed protein (IL-6). (B) Among the top 100 model-ranked features, 32 (32%) are predicted to bind directly to the IL-6 protein. Panel (B) shows the local protein–protein interaction network of these 32 features, as derived from the STRING database.

Independent dataset testing

To further confirm that our protein abundance imputing model is robust, we performed two extra canine surgery experiments (two canines with a total of 27 biopsy samples) and performed paired RNA-seq and protein assay (Luminex™ and ELISA) as a new hold-out dataset. The feature selection and model training were based on the previous 77 biopsy samples. The 27 newly acquired biopsy samples were used as an independent dataset for hold-out testing. As shown in Supplementary Figure S4, all of the 17 proteins in the 27 hold-out biopsy samples are successfully imputed from RNA-seq data, with Spearman’s Rho > 0.5 between RNA-seq imputed protein abundance and measured protein abundance. Among them, 10 out of the 17 proteins showed very high prediction accuracy with Rho > 0.8. This hold-out dataset further confirmed that our RNA-seq-based protein abundance imputing model is robust.

Imputing abundance of post-translationally activated TGF-β1 protein from RNA-seq data

After mRNA is translated to proteins, many proteins undergo post-translational modifications, which significantly modify the protein activity, function and interactions with other molecules within the cell. We hypothesize that post-translational modifications could initiate a cascade response at the transcriptomic level, leaving discernible imprints on the mRNA expression patterns. As a result, the RNA-seq data can be potentially used to infer and investigate these post-translational modification events.

TGF-β1 is a cytokine and a member of the TGF-β superfamily, which plays a crucial role in various cellular processes, including cell growth, differentiation, apoptosis and immune regulation. Initially, TGF-β1 is produced as a precursor molecule, known as the ‘large latent complex’, which consists of TGF-β1 associated with its propeptide called latency-associated peptide (LAP). This complex prevents TGF-β1 from interacting with its receptors and signaling pathways. To activate TGF-β1, the large latent complex must undergo specific cleavage events. Once the LAP is cleaved off, the active TGF-β1 is released and can bind to its specific receptors on the cell surface. Once activated, TGF-β1 can initiate intracellular signaling cascades, ultimately regulating gene expression and leading to various cellular responses.

As depicted in Figure 6A, a moderate correlation is observed between the abundance of the active form of TGF-β1 and the total abundance (comprising both active and latent forms) with a Spearman’s Rho value of 0.53. This finding suggests that relying solely on the total abundance as a measurement may not accurately reflect the regulatory role of TGF-β1. To test the hypothesis that changes in transcriptomic footprint resulting from post-translational modifications could be used to infer these modification events, we trained an SVR model by paired RNA-seq data and ELISA measured the TGF-β1 active form abundance data. Our model suggests that the RNA-seq imputed TGF-β1 active form abundance is highly correlated with the ELISA measured abundance (Figure 6B; Spearman’s Rho = 0.95; 5-fold cross-validation). There are nine GO terms (molecular function) enriched (adjusted P-value <5%) in features selected by the model (Figure 6C). Most enriched GO terms are linked to phosphatase activity. It is well established that the activation of TGF-β signaling primarily hinges on a phosphorylation cascade, originating from the receptor and spanning to the Smad proteins (24). Additionally, a significant number of protein phosphatases have been recognized as crucial controllers of TGF-β signaling, operating at both the receptor and Smad levels (24). This suggests that the model chose genes associated with TGF-β1, encompassing both upstream regulators and downstream-regulated targets, as features to estimate the abundance of the active form.

Figure 6.

Imputing TGF-β1 active form abundance from RNA-seq data. (A) Moderate correlation between the total and the active TGF-β1 abundance. (B) Correlation between ELISA measured and RNA-seq imputed active TGF-β1 form abundance. (C) Enriched GO terms for imputed model-selected features.

Discussion

The correlation between mRNA abundance and protein abundance has been a subject of long-standing debate. Some studies highlight a strong relationship between the two (25), while others indicate no such correlation (6). This disagreement arises from the complexities of gene expression and regulation, as well as the diversity of methodologies employed in different investigations. Key factors contributing to this debate include but are not limited to post-transcriptional regulation, translation efficiency, protein degradation, experimental techniques and conditions, and biological variability (e.g. organisms may exhibit varied correlations between mRNA and protein levels, reflecting their unique biological characteristics). In this study, we observed a general concordance between mRNA and protein levels for 17 genes relevant to wound healing under the same experimental conditions. However, for specific genes and under varying conditions, this correlation was absent. One potential reason could be that the variance between genes within the same condition is substantially greater than the variation in any single gene across different conditions. The relationships between mRNA and protein levels are predominantly affected by genes that display substantial variations. This observation is supported by the differences in mRNA levels from gene to gene within the same biopsy (as seen in Figure 1), in contrast to the variations between biopsies that correspond to differing wound healing times, wound zones and individuals (as illustrated in Figure 2). Such observations might lead to the question of how much change at the mRNA level is required to alter the protein level. Answering this question is technically challenging, as the answer could heavily depend on specific conditions, tissues or timing. In the present study, we propose an alternative approach to tackle this issue by using RNA-seq data to infer protein levels. The protein amounts imputed from RNA-seq strongly correlate with the levels measured in protein assays. These findings suggest that we can employ RNA-seq data to estimate protein abundance using a machine learning model. Additionally, the model showed the capability to predict TGF-β1 activation events based on RNA-seq data, suggesting that the method may have even broader applications.

Results of this study show the potential to extend the utility of RNA-seq from simply investigating transcriptomic changes to also investigating proteomic alterations and post-translational modifications. This capability represents a significant advancement in the field, broadening the scope of RNA-seq applications to include a more comprehensive understanding of protein activity and modifications.

It is acknowledged that for generalized use, one must train the model via a comprehensive biological context with more proteins. Future efforts will focus on gathering a larger dataset encompassing a wider range of biological conditions and protein variations. These expanded studies will refine the model and improve prediction accuracy. We intend to continuously validate the findings reported herein with experimental results to ensure the model’s robustness. By expanding and refining this methodology, we will provide a reliable tool for researchers to estimate protein levels from RNA-seq data.

Data availability

We have submitted the RNA-seq data and the gene expression data (TPMs and mapping counts) to Gene Expression Omnibus (accession number: GSE242973).

Supplementary data

Supplementary Data are available at NARGAB Online.

Funding

Defense Advanced Research Projects Agency (DARPA) [HR00111950027].

Conflict of interest statement. None declared.