RNA-Seq is crucial for understanding cancer at the genetic level. By analyzing gene expression, it allows researchers to categorize cancers into different molecular subtypes. These subtypes can provide important information about the disease’s behavior, prognosis (how the disease might progress), and potential treatment options. But, there’s a catch—how the RNA-Seq data is prepared (or “preprocessed”) before it’s analyzed can impact the results and, ultimately, the accuracy of subtype classification.

A recent study be researchers at the Hospital del Mar Research Institute (HMRI) focused on how different methods of preprocessing RNA-Seq data affect the classification of bladder cancer subtypes. The researchers tested several popular methods for processing RNA-Seq data, such as aligners (which map the RNA sequences to the correct parts of the genome), quantifiers (which calculate how much RNA comes from each gene), and techniques to normalize or transform the data to make it comparable across samples. They wanted to find out how these choices could influence which subtype a sample is classified into, and how reliable the results are.

The study found that log-transformation—a common technique used to adjust data—plays a very important role when using some classification methods, particularly centroid-based classifiers like consensusMIBC and TCGAclas. These classifiers try to group cancer samples into subtypes based on gene expression patterns. When the data was not log-transformed, the results showed low classification accuracy, meaning the subtype predictions were less reliable.

Heatmap of consensusMIBC classification using benchmarked tools

Each column shows the results for the different RNA-Seq combinations of aligner, quantifier, of normalization and transformation method. Each row represents a sample from several public datasets (GSE148079, GSE179440, GSE186610 and TCGA). Colors of the heatmap indicate the predicted molecular subtype. Column on the left represents the reference for each sample (i.e. the most commonly predicted subtype). Luminal papillary (LumP), luminal unstable (LumU), luminal non-specified (LumNS), stroma rich (Stroma-rich), basal squamous (Ba/Sq) and neuroendocrine-like (NE-like). Kallisto, Salmon, STAR + featureCounts (STAR + featC), STAR + HTSeq, STAR + StringTie (STAR + StringT), Hisat2 + featureCounts (Hisat2 + featC), Hisat2 + HTSeq, Hisat2 + StringTie (Hisat2 + StringT) and Reference. Barplots show the mean and standard deviation of the correlation values for each methodology

On the other hand, distribution-free algorithms like LundTax were much more robust to variations in data preprocessing. This means LundTax could accurately classify subtypes even when different preprocessing methods were used. Interestingly, LundTax also performed better overall, consistently distinguishing between bladder cancer subtypes more effectively than the other methods, regardless of how the data was preprocessed.

One important thing to note is that the study didn’t have a true reference to compare the results against—meaning they didn’t have a gold standard to say for sure how accurate the classifications were. This limits the ability to make firm conclusions about which preprocessing method is absolutely best for subtype classification. The researchers also suggest that more work will be needed to understand how their findings could be applied to larger and more diverse datasets.

This study emphasizes that preprocessing of RNA-Seq data is not a step to be overlooked—it’s crucial for ensuring accurate and reliable results, especially when classifying cancer subtypes. By understanding how different preprocessing methods affect results, researchers can choose the best approach to classify bladder cancer accurately, leading to better treatment decisions and outcomes for patients.