gneSeqCOO: a novel method for classifying diffuse large B-cell lymphoma cell of origin based on bulk tumor RNA sequencing profiles

Abstract

The cell of origin (COO) classification is an expression-based tumor algorithm identifying molecular subtypes of diffuse large B-cell lymphoma (DLBCL) with distinct prognostic characteristics. Traditional immunohistochemical methods for classifying COO subtypes have poor concordance and limited prognostic value in frontline DLBCL. In contrast, RNA-based metrics like the NanoString Lymphoma Subtyping Test (LST) define more robust subtypes with validated prognostic associations. This study introduces gneSeqCOO, an algorithm using bulk RNA Sequencing (RNASeq) profiles of individual tumor biopsies for COO classification based on a fixed reference. This method produced consistent per-sample results and was robust to variation in RNA quality and sequencing bias. Validation in >1000 DLBCL samples showed high concordance with the NanoString LST assay, even in cohorts containing only one COO subtype. gneSeqCOO presents a robust and versatile alternative to existing assays, potentially reducing the need for additional samples where RNASeq was already generated. The package is available at https://github.com/Genentech/gneSeqCOO.

Introduction

The cell of origin (COO) classification is a gene expression-based tumor algorithm that was developed in diffuse large B-cell lymphoma (DLBCL) in 2000 [1]. This classifier divides samples into two distinct subtypes, which are associated with the tumor’s COO. The two subtypes are termed germinal center B-cell-like (GCB) and activated B-cell-like (ABC) DLBCL, with an unclassified region separating them, and express genes characteristic of normal germinal center B cells or activated B cells, respectively. Since the discovery of these two molecular subtypes in 2000, distinct prognostic characteristics have repeatedly been observed among previously untreated patients with DLBCL treated with rituximab plus cyclophosphamide, doxorubicin, vincristine, and prednisone (R-CHOP), and a number of different therapy options have been investigated [2–6].

The most common methods for classifying COO subtypes are based on immunohistochemistry staining of patient samples. However, these methods have been reported to have poor concordance with each other and are of limited use in predicting the outcomes of patients with DLBCL in representative datasets [7–10]. In contrast, increasing evidence has demonstrated the utility of molecular subtyping as a tool to deliver precision medicine in DLBCL [11–14]. Tools based on RNA expression, such as NanoString, reverse transcriptase multiplex ligation-dependent probe amplification (RT-MLPA), and RNA Sequencing (RNASeq) have shown promise in effectively separating distinct prognostic subtypes and are commonly used in research applications. NanoString’s Lymphoma Subtyping Test (LST) is a Clinical Laboratory Improvement Amendments (CLIA)-validated assay for predicting subtypes based on a panel of 20 genes: 15 genes specifically expressed in GCB or ABC, plus five additional housekeeping genes. The assay produces a linear predictor score (LPS), which uses fixed cutoff points to classify samples as either GCB (LPS <1920) or ABC (LPS >2430). Samples falling between these cutoffs are categorized as unclassified.

In many research settings, it is common to run the NanoString algorithm alongside other molecular characterization tools, such as RNASeq. This approach requires a large amount of tumor material and tends to produce overlapping information regarding COO-related gene expression. To address this issue, other RNASeq-based COO classifiers have been developed [15]; however, these methods rely on understanding the distribution of expression profiles across a cohort before a COO subtype can be assigned, which complicates their use in a clinical setting where samples may only be received in small batches.

Previously, data from a historical clinical trial dataset was utilized to develop an RNASeq-based COO classifier [16]. Here, we provide a detailed description of the method, referred to as gneSeqCOO, which uses the RNASeq expression profile of a single sample to classify COO subtype based on a fixed reference, with the NanoString LST assay as the gold standard. This method produces results on a per-sample basis, reducing the need for large cohorts. Validation was conducted using >1000 DLBCL samples, to demonstrate high concordance with the gold standard, even in samples with low RNA quality.

Methods

Bulk RNASeq datasets

RNA was isolated and purified from pretreatment formalin-fixed paraffin-embedded (FFPE) biopsies of patient samples collected from four clinical trials (CAVALLI [NCT02055820] [17], ROBUST [NCT02285062] [5], POLARIX [NCT03274492] [18], and GOYA [NCT01287741] [19]), using Qiagen FFPE RNAEasy preparation kits. The purified RNA was used to create complementary DNA libraries, which were assayed using TruSeq (Illumina) RNASeq. The RNASeq data were analyzed with HTSeqGenie and aligned to the human reference genome (NCBI Build 38) using GSNAP [20,21]. To quantify gene expression levels, the number of reads mapped to the exons of each RefSeq gene was calculated.

Normalization of RNASeq data to a fixed reference

To ensure that the final COO classification was not biased by differences in RNA quality or library size, a schema was established to normalize the input RNASeq profile to a fixed reference. To enable single-sample normalization for future samples, the gneSeqCOO method reimplements DESeq2’s size factor calculation using a fixed reference sample; in this case, the geometric mean of all samples from the GOYA datasets. For a given sample $j$, the size factor $S_j$ was calculated as the median of the ratios of observed counts $k$, divided by the reference counts for each gene $i$:

$${\widehat{S}}_j=\underset{i}{\mathrm{median}}\frac{k_{ij}}{re{f}_i}$$

Of note, genes with low expression in the reference dataset were excluded (median [count] > 0 and mean [count] > 100), such that genes with high expression were predominantly used to estimate the size factor. Following this, the robust library size for a sample was calculated by multiplying the size factor by the sum of the reference counts:

$${\widehat{L}}_j={\widehat{S}}_j\sum _{i}re{f}_i$$

Samples were then normalized by the robust library size to obtain counts per million, which were used in all downstream applications. This approach generates normalized RNASeq profiles that are robust to library size variations and are generally consistent across differences in RNA quality. To demonstrate this, reference normalization was applied to the GOYA dataset. It is important to note that the minimum value (i.e. genes with zero reads) will always depend on the true library size, meaning that a classifier using low-expression genes may still be subject to technical biases.

While library size is a major source of sample variance, it does not capture all aspects of sample quality. In cases with low RNA quality, deep sequencing may yield similar library sizes but result in a large number of zero-count genes, leading to an increasingly skewed distribution. To explore these biases further, the degree of skewness (measuring the degree of non-normal distribution in gene expression) was calculated for all samples in the GOYA dataset.

Developing a robust RNASeq classifier

Given the challenges in completely eliminating batch effects, particularly those related to distribution skew, an effective prediction model should preferentially utilize genes that are least affected by the skew. To identify these genes, and to assess the robustness of the classifier relative to data quality, the dataset was divided into four cross-validation cohorts based on the skew quantiles established above. In each cross-validation run, three of the cohorts were combined to create a training dataset, while the fourth was used as a validation cohort. A penalized regression model (lasso) was built to predict the NanoString LPS score, which was then cut using the pre-defined COO cutoff points (GCB: LPS <1920; ABC: LPS >2430). A total of 165 features were considered, including COO-related features from previous literature, the five housekeeping genes from the NanoString LST assay, and a measure of sample skewness.

Evaluation of model performance on other datasets

To assess the performance of the developed linear model, three validation cohorts were employed; clinical and bulk RNASeq data were available for 140 patients with DLBCL in the CAVALLI cohort, 205 patients with DLBCL in the ROBUST cohort, and 641 patients with DLBCL in the POLARIX cohort. Notably, the ROBUST trial specifically aimed to recruit patients with the ABC subtype, with 204 of the 205 available samples classified as ABC by the NanoString LST assay.

Results

Normalization of RNASeq data to a fixed reference

Reference normalization applied to the GOYA dataset showed that the distribution of expression profiles remained largely stable regardless of input library size (Figure 1(A)). This stability was also observed for most genes used in the NanoString LST assay; however, for some genes, an association between gene expression and library size was observed (5/20 genes with p < 0.05; representative genes are shown in Figure 1(B)). The median and third quartile expression were generally uncorrelated with skewness, while the first quartile expression demonstrated a strong association (Figure 1(C)). This was also evident in the expression of individual genes (15/20 genes with p < 0.05; representative genes are shown in Figure 1(D)), although both the degree and the direction of correlation varied from gene to gene.

Figure 1.

Normalization of RNASeq data to a fixed reference. Expression quantiles for GOYA samples after fixed reference normalization. Points represent: min (red), Q1 (25%; orange), median (dark blue), Q3 (75%; light blue), and max (purple). Samples are ordered by total library size (A). Expression of select genes from the NanoString LST assay, bucketed by quartile of library size (red: 0–25%; orange: 25–50%; dark blue: 50–75%; light blue: 75–100%) (B). Expression quantiles for GOYA samples ordered by degree of skewness. Points represent: min (red), Q1 (25%; orange), median (dark blue), Q3 (75%; light blue), and max (purple) (C). Expression of select genes from the NanoString LST assay, bucketed by skewness (red: 0–25%; orange: 25–50%; dark blue: 50–75%; light blue: 75–100%) (D).

LST: Lymphoma Subtyping Test; Q: quartile; RNASeq: RNA Sequencing.

Developing a robust RNASeq classifier

All four cross-validated models had high accuracy, with correlations ranging from 0.97–0.99 in the holdout data (Figure 2(A)). Importantly, none of the datasets showed evidence of a consistent shift in prediction scores; errors were centered around 0 (Figure 2(B)), and misclassification of COO subtypes was generally balanced between GCB and ABC subsets (Figure 2(C)). The models also utilized a similar set of genes, with 16 genes used in all four models, five genes in three out of four models, and 20 genes used only once or twice. Importantly, the skewness parameter was not included in any of the four models and was therefore not considered for downstream analysis.

Figure 2.

Cross-validation model performance.* Predicted LPS scores were compared with the gold standard LPS scores from the NanoString LST assay (A). Beeswarm of the model error (predicted - gold standard) (B). COO classification determined based on a fixed cutoff (C).

*Samples from the GOYA cohort were split into four quartiles based on the degree of skewness, and the resulting quartiles were used for cross validation.

ABC: activated B-cell; COO: cell of origin; GCB: germinal center B-cell; LPS: linear predictor score; LST: Lymphoma Subtyping Test.

Based on these results, the 21 genes selected in at least three models were used to create a final classifier utilizing a simple linear model (without any regularization step). The selected genes were BATF, MAST2, HCK, BSPRY, SMARCA4, GNAT3, PIM2, CCDC50, TNFRSF13B, SACS, SERPINA9, IRF4, ITPKB, MME, MYBL1, CYB5R2, MAML3, CREB3L2, ASB13, LIMD1, and S1PR2. This set included 11 out of 15 COO-associated genes identified by the NanoString LST assay and did not include any housekeeping genes. The coefficients for the final model are reported in Table 1.

Table 1.

Genes and model weights used in the final RNASeq-COO classifier.

Gene (Intercept)	Coefficient 1310.3
BATF	−14.9
GNA13	−70.4
PIM2	145.2
CCDC50	26.7
MAST2	−25.0
TNFRSF13B	89.1
SACS	78.4
HCK	42.8
SERPINA9	−72.2
IRF4	83.9
ITPKB	−47.9
MME	−57.3
MYBL1	−105.8
CYB5R2	81.7
BSPRY	24.2
MAML3	−84.1
CREB3L2	105.2
SMARCA4	−71.0
ASB13	−84.6
LIMD1	92.7
S1PR2	−98.6

COO: cell of origin; RNASeq: RNA Sequencing.

Evaluation of model performance on other datasets

The raw counts from each validation dataset were normalized to the fixed reference, demonstrating that the expression of COO model genes was consistent with the GOYA reference (Figure 3(A)). All three validation cohorts exhibited high concordance with the gold standard NanoString LPS score classification (Figure 3(B)), ranging from 90.17–98.05% (Table 2). NanoString and predicted LPS were strongly correlated (Supplementary Figure 1A and 1B). Of note, most differences occurred along the decision boundaries, with samples switching to or from the unclassified subset. Only one case involved a patient with the GCB subtype being classified as ABC; there were no cases of patients with the ABC subtype being classified as GCB. These results suggest that the model’s performance was highly consistent, regardless of the distribution of the underlying data.

Figure 3.

Model validation. Samples from the external clinical cohorts compared with GOYA by PCA; samples with discordant COO calls are highlighted with a dark border (A). Predicted LPS scores compared with the gold standard LPS scores from the NanoString LST assay; colors are associated with NanoString COO call (B).

*In ROBUST, the only patient with the GCB subtype was correctly categorized.

ABC activated B-cell; COO: cell of origin; GCB: germinal center B-cell; LPS: linear predictor score; LST: Lymphoma Subtyping Test; PC: principal component; PCA: principal component analysis.

Table 2.

Model validation with the CAVALLI, ROBUST, and POLARIX cohorts.

	Predictions
NanoString	ABC	Unclassified	GCB
CAVALLI cohort (n = 145)
ABC	44	0	0
Unclassified	3	16	2
GCB	1	3	71
Unknown	0	2	3
ROBUST cohort (n = 205)
ABC	200	4	0
GCB	0	0	1
POLARIX cohort (n = 665)
ABC	212	2	0
Unclassified	33	59	0
GCB	0	28	307
Unknown	11	4	9

ABC: activated B-cell; GCB: germinal center B-cell.

Discussion

This study introduced a novel classifier that utilizes the RNASeq expression profile of a single sample to categorize the COO subtype based on a predefined reference set. The findings demonstrate that this new RNASeq-based classifier exhibits a high level of concordance with the gold standard NanoString LST assay. Across all three validation cohorts, the concordance rates ranged from 90.17–98.05%, with all but one of the misclassified samples switching to or from the unclassified subset.

In many studies of DLBCL, classification of COO subtypes has relied on the NanoString LST assay, which requires 200–300 ng of RNA. In contrast, this novel approach, which utilizes bulk RNASeq, requires only 100 ng of RNA. Notably, the field is witnessing the development of more prognostic signatures using bulk RNASeq data, including deconvolution methods that explore the complexities of the tumor microenvironment and their impact on treatment outcomes [22,23]. Given the challenges of obtaining biopsy samples of sufficient quantity for clinical trials, prioritizing methods that can assess multiple biomarkers becomes imperative. Consequently, the availability of a publicly accessible tool for determining COO using bulk RNASeq data, in conjunction with other signatures, presents researchers with a comprehensive understanding of a patient’s tumor biology. This innovative approach promises to enhance access to COO classification through gene expression profiling, thereby improving the diagnosis of patients with DLBCL.

In the present study, an RNASeq-based COO classification algorithm was validated using three large independent cohorts of samples collected through multiple global Phase 2 and 3 studies [5,17–19]. This represents the largest direct comparison with results obtained from the NanoString LST assay. Importantly, while other algorithms often rely on fitting the distribution of a cohort of DLBCL samples, this algorithm is designed to classify one sample at a time, allowing it to perform effectively even with highly skewed data. This was demonstrated using the ROBUST dataset, which consisted almost entirely of patients with ABC-DLBCL, where the algorithm maintained strong performance despite the imbalanced distribution.

While the overall accuracy of the classifier was high, consistent shifts in prediction scores were observed in the POLARIX dataset. Most misclassification occurred from GCB to unclassified and from unclassified to ABC. Although the exact cause of these shifts is unknown, they may reflect differences in sample processing or sequencing platforms used in the study. Additionally, this algorithm was trained using FFPE-derived RNASeq profiles and is designed to work with FFPE samples. While the algorithm can be applied to fresh frozen profiles, differences in RNA quantification tend to lead to lower overall accuracy (data not shown). This highlights a potential limitation of RNASeq-based classification methods; namely that different platforms may have varying exon capture rates, which can lead to slight but consistent shifts in gene expression levels within the classifier. Historically, the unclassified border was intended to provide a buffer against such changes, and the shifts observed in the POLARIX dataset did not result in misclassifications outside this region. However, future work will be required to evaluate the performance of this method on other datasets.

Code availability

The gneSeqCOO package is available at https://github.com/Genentech/gneSeqCOO.