Tumor subtype classification tool for HPV-associated head and neck cancers

Abstract

Background:

Molecular subtypes of HPV-associated Head and Neck Squamous Cell Carcinoma (HNSCC), named IMU (immune strong) and KRT (highly keratinized), are well-recognized due to distinct molecular features, tumor microenvironments, clinical outcomes, and potentially differing optimal treatment strategies. Currently, no standardized method exists to subtype a new HPV + HNSCC tumor. Our paper introduces a machine learning-based classifier and webtool to reliably subtype HPV + HNSCC tumors using the IMU/KRT paradigm and highlights the importance of subtype in HPV + HNSCC.

Methods:

We conducted RNA-seq on 67 HNSCC tumors from University of Michigan Health. Combining this with three publicly available datasets, we utilized a total of 229 HPV + HNSCC RNA-seq samples. The classifier was trained and tested using 84 subtype-labeled HPV + RNA-seq samples and validated with the remaining samples. We also tested the association of 37 clinicodemographic and molecular variables with subtype.

Results:

The classifier achieved 100% accuracy in the test set. Validation on two additional cohorts demonstrated successful separation by known features of the subtypes. Investigation of the relationship between subtype and the molecular and clinicodemographic variables revealed 21 significant associations, both confirming previous findings and revealing novel subtype associations.

Conclusions:

This study provides a reliable classifier for subtyping HPV + HNSCC tumors as either IMU or KRT based on bulk RNA-seq data and improves our understanding of the HPV + HNSCC subtypes.

Introduction

Cancer types are typically categorized according to cell of origin, but vast heterogeneity often exists within these groupings [1]. Research in breast [2], lung [3], pancreatic [4] and colon [5] cancers has uncovered distinct subtypes based on specific gene driver mutations or epigenetic signatures. Such subtypes often have clinical utility as prognostic biomarkers, aid physicians in therapeutic strategies, or are associated with treatment response [2]. As precision medicine and targeted therapies advance, the utility of defining more narrowly defined subtypes is amplified.

HPV-associated head and neck cancer continues to increase at an epidemic level. Approximately 30 % of HNSCC [6] can be attributed to human papillomavirus (HPV) with oropharyngeal (OPSCC) being the most common site associated with HPV [7]. HPV infection currently drives approximately 71 % and 52 % of OPSCC in the USA and UK, respectively [8], and typically confers a survival advantage, with 5-year survival rates averaging ~ 80 % [9]. While there is wide morphologic and epigenetic diversity within HPV + HNSCC [10], tumor subtyping is not yet widely used as a clinical indicator for this cancer population.

HPV + HNSCC molecular subtyping has been conducted by multiple groups, as reviewed by Qin et al [9], and most of these studies used gene expression levels [11–13]. Keck, et al. were first to define HPV + HNSCC subtypes, which they named IMS (immune strong), and CL (classical) [11]. IMS had prominent immune and mesenchymal features while CL was enriched for the putrescine (polyamine) degradation pathway. Zhang et al. re-identified the HPV + HNSCC subtypes as IMU (immune strong) and KRT (highly keratinized) using RNA-seq and copy number variations (CNVs) [12], discovering a strong association between KRT and HPV integration. Locati et al. further discriminated KRT tumors into high and low stromal groups [13,14], and demonstrated that IMU patients have better prognosis than either high or low stromal KRT. The subtype naming convention of IMU/KRT was adopted in Leemans et al [15], and these subtypes have been characterized using additional high-throughput technologies, including DNA methylation [16], showing stronger global hypomethylation in KRT, and DNA hydroxymethylation [17]. The IMU/KRT subtypes were also demonstrated to significantly associate with HPV E6 isoform expression, with KRT tending to have higher levels of the spliced E6* isoforms compared to E6 full length [18,19].

Although unsupervised methods such as clustering have identified cancer subtypes, the clusters and subtype assignments obtained naturally vary across studies. This inconsistency arises from differences in factors such as cohort attributes, sample quality, RNA preparation methodologies, technical variations between batches, and the specific clustering algorithm utilized [20]. Thus, a consistent, reproducible approach is required. To overcome the current limitations in subtyping HPV + HNSCC and standardize subtype classification of new tumors, we trained and built a robust machine learning (ML) classifier, including several steps to enhance rigor and reproducibility. We first used 84 HPV + HNSCC from two cohorts (18 from University of Michigan and 66 from TCGA) to train and test an ensemble classifier involving five ML models and three predefined gene sets as input features. We then applied our classifier to two additional cohorts of HPV + OPSCC and found results consistent with known subtype features and clustering results. We introduce a user-friendly webtool that streamlines and simplifies the process of subtyping HPV + HNSCC for future research. Lastly, we performed meta-analysis of the 219 subtyped HPV + HNSCC unique patient samples and identified 21 relevant pathways and clinicodemographic variables associated with subtype.

Methods

HNSCC cancer datasets

Four RNA-seq datasets were used. To train and test the classifier, we used two datasets with previously identified subtypes, 18 HPV + HNSCC cases from the University of Michigan (UM18) (GSE74956) and 66 HPV + TCGA-HNSC samples (n = 84: 33 IMU & 51 KRT) [12]. Two additional datasets, used for validation, were from the HPV Virome Consortium [21] (HVC) (n = 83 HPV+; EGAD00001004366) and a newly introduced University of Michigan (UM) OPSCC cohort (UM67) (n = 62 HPV +) from which we used RNA from formalin-fixed, paraffin-embedded (FFPE) blocks. Written informed consents were obtained and the study was approved by the University of Michigan Institutional Review Board (See Supplementary Methods). Ten samples (duplicated10) in UM67 had matched FF samples in UM18. Raw gene counts of all 229 samples from the four cohorts were converted to log2CPM values and then z-transformed for each gene (Non-PCA). We also performed principal component analysis on Z-scores covering 80 % of the total variance (PCA). Both Non-PCA and PCA values were used for training, testing and validating purposes. Individual participant information including tumor site is provided in Supplementary Table S1.

Feature selection

For training, we designed three varying-sized gene sets (Supplementary Table S2); the smallest one was derived from KECK (IMS/CL [11]) and the two larger sets were obtained from Zhang et al (IMU/KRT [12]). Each training gene set was selected from the most significantly differentially expressed genes between subtypes and balanced by IMU/KRT differentially expressed pathways (See Supplementary Methods).

Classifier model training

To improve the performance and robustness of our model, we used an ensemble approach by training five different ML models and applying majority voting on the 15 (5 ML methods × 3 input gene sets) individual models to make the final prediction. We tuned the ML models’ hyper-parameters by cross-validation (CV) (see Supplementary Methods). One ensemble model was trained for each format of input features (PCA and Non-PCA).

Validation of the ensemble ML subtype classifier

We applied the classifier on two additional independent HPV + OPSCC cohorts (UM67 and HVC) and checked whether 1) the classifier results for the duplicated10 samples matched the original subtype for those patients; 2) the assigned subtypes correlated well with the clinical characteristics, molecular features and pathway scores known to be associated with subtype.

Webtool development and usage

A user-friendly webtool was developed accepting a matrix of gene counts or log₂CPM values. We provide our UM18 cohort to assist in mitigating batch effects between training data and user input, and to assure accurate results for small numbers of samples (see Supplementary Methods).

Calculation of molecular variables

For all samples, HPV + HNSCC-relevant gene expression signature scores were calculated to characterize tumor immune microenvironment, differentiation state, HPV gene activity, and oxidation–reduction. These were generated as sample-wise pathway scores aggregated by the rank of log₂CPM gene expression levels. First, for each gene in the relevant pathway, we ranked the samples according to their expression levels. For each sample, the ranks of the genes were summed, and the resulting values were z-score transformed across samples. To calculate the epithelial-to-mesenchymal transition (EMT) score [22], negatively regulated genes were also ranked in descending order. The gene sets used in Fig. 3A–B for molecular features are in Supplementary Table S3. For additional variable calculations, see Supplementary Methods.

Fig. 3.

Subtype is central to biological and clinically relevant HPV + head and neck cancer characteristics. (A) Network of clinicodemographic and molecular variables significantly associated with subtype. Red edges represent positive associations with IMU and blue edges represent positive associations with KRT. Edges between subtype and other nodes are bold for ease of viewing. The width of the edges represents the strength of association by p-value – wider being more highly significant. The shape of the nodes represents the size of the cohort used. (B) Heatmap showing pathway and cell type associations with subtype, created with R package ComplexHeatmap using z-scores of gene expression corrected for cohort effect with linear regression. Gene set scores and deconvolution derived cell-type proportions were clustered with Euclidean distance and compete linkage clustering.

Calculation of associations among subtype, molecular, and clinicodemographic variables, and generation of network graph

To generate the association network, associations between variables were calculated including cohort as a covariate using logistic regression (categorical-categorical) or ANOVA tests (see Supplementary Methods).

Results

Ensemble classifier to subtype HPV + HNSCC bulk RNA-seq samples

Although IMU is typically characterized by a strong immune response and KRT by high levels of keratinization, we found that measures of immune infiltration (T cell activation scores or T cell proportion) and keratinization scores (see Methods) were inadequate to subtype tumors in UM18 and TCGA HNSC cohorts (Supplementary Fig. S1 A–B). This motivated us to develop a ML-based classifier, taking several steps to ensure its rigor and reproducibility (Supplementary TAble S4).

We designed three varying-sized gene sets (10, 50, and 148 genes) of top differentially expressed genes between subtypes to balance the risk of overfitting with including sufficient information. (see Methods, Figure1A, Supplementary Table S2, Supplementary Fig. S2A, see Methods). As a sanity check, we verified that these gene sets were able to effectively separate TCGA samples by subtype using standard PCA, indicating their potential as training features (Supplementary Fig. S2 B–D).

Fig. 1.

Ensemble classifier training, cross-validation (CV), and testing. (A) The schematic description of input gene set selection, pre-processing for training and testing, CV and testing, and implementation of the ensemble model. (B) The mean CV accuracy for each ML model (elastic net, Gaussian Naïve Bayes (gnb), k-nearest neighbors (knn), random forest (rf), and support vector machine (svm)) and input gene set size. (C) Confusion matrices from test results for the non-PCA and PCA based ensemble models.

We trained and tested the classifier using the 84 RNA-seq samples from TCGA and the University of Michigan that were previously subtyped as IMU or KRT. Training was performed on the UM18 cohort and 49 of the TCGA cohort (n = 67;18 + 49, Supplementary Table S1). To enhance the classifier robustness, we used an ensemble approach with five ML algorithms and three gene sets in two feature formats (PCA and Non-PCA) (Fig. 1A, see Methods). By comparing mean cross-validation (CV) accuracy between PCA and Non-PCA, we did not observe significant differences (Fig. 1B). We found that models trained using more genes (50 and 148 genes) tended to have ~ 0.96 mean CV accuracy compared to ~ 0.92 for the gene set of 10 (Fig. 1B), but noted this benefit of the larger gene sets may be offset by overfitting in other datasets. The fluctuations in mean CV accuracy across ML algorithms demonstrated that no single ML algorithm was overall optimal (Fig. 1B, Supplementary Fig. S3A–B), illustrating the value of an ensemble approach by reducing bias from a single gene set or ML algorithm.

We evaluated our ensemble model by testing it on the remaining TCGA samples (7 IMU; 10 KRT). Based on the confusion matrices (Supplementary Fig. S3C–D), individual model misclassifications occurred in both subtypes across models and gene sets, indicating no bias towards either subtype. Although single-model misclassifications occurred for six samples (Supplementary Table S5), the final ensemble model achieved 100 % accuracy for both the PCA and Non-PCA format (Fig. 1C).

Application of the subtype classifier to two additional HPV + OPSCC cohorts validates its accuracy and robustness

To further validate the robustness and generalizability of our subtype classifier, we applied it to two additional HPV + OPSCC cohorts with RNA-seq data (145 samples from UM67 and HVC, Supplementary Table S1) and assessed it from four perspectives: consistency among the ensemble machine learners, reproducibility between FFPE and fresh frozen (FF) samples, subtype-associated pathways, and robustness across cohorts and with missing values (Fig. 2A). Overall, 101/145 (70 %) of the samples had majority votes of 15 vs 0 or 14 vs 1, while 11 samples had votes of 6 vs 9 or 7 vs 8 in both PCA- and Non-PCA-based results (Supplementary Table S5). Only two samples (GS18070 and GS18034) (1.4 %) had inconsistent assignments between PCA and Non-PCA (Fig. 2B), likely due to their voting results being divided: either 6 versus 9 or 7 versus 8, with predicted IMU probabilities between 0.4 and 0.6 (Supplementary Table S5). Upon closer examination, both of these samples had mixed molecular signals in the key distinguishing pathways of IMU/KRT (Supplementary Fig. S4A). This demonstrates how the classifier can identify the rare cases that have molecular characteristics inconsistent with either IMU or KRT using the voting pattern. Secondly, we examined the results for the duplicated10 samples (see Methods) and found that all ten predictions for the FFPE samples were consistent with the originally-defined subtypes from their paired fresh frozen samples (Supplementary Table S5), confirming that the classifier is robust within patients and across sources of biospecimens (FFPE versus fresh frozen). We next examined the expression of 24 key IMU/KRT differential genes, which were selected in the original Zhang et al paper defining the IMU/KRT subtypes [12] to represent the five main differential pathways between subtypes, in all newly classified samples. Importantly, only 3 of the 24 genes (SFN, HLA-DQB2, BCL2) were used in training. For both cohorts, the Non-PCA and PCA-based classifier subtyped the samples consistently and in close agreement with expected changes in these genes (Fig. 2B, Supplementary Fig. S4B), while the unsupervised clustering (k = 2) based on these genes resulted in inconsistent clustering for both cohorts (Supplementary Fig. S4C–D), emphasizing the robustness of the classifier over unsupervised clustering. Lastly, we validated the robustness of the classifier by testing random subsets of the features (30 %, 50 % and 80 %), and found that 30 % led to only eight (5.5 %), 50 % led to five (3.4 %) and 80 % led to two (1.4 %) samples being misclassified (Supplementary Fig. S5A–B). These provide an estimate of accuracy for various levels of missing data. For all cohorts involved in this study, we consistently found that approximately 60 % of tumors were KRT and 40 % were IMU (Fig. 2C), confirming stability of the classifier.

Fig. 2.

Application of the classifier on two independent HPV + OPSCC cohorts validates the classifier accuracy and robustness. (A) The schematic for applying the classifier on two independent cohorts and evaluating the classifier performance. (B) Heatmap showing that 24 key genes in pathways separate the IMU/KRT subtypes for the HVC cohort, with PCA and Non-PCA predicted results ordered as annotation rows. (C) Proportion of the final subtypes for each cohort of patients.

Subtype is central to biological and clinically-relevant HPV + cancer characteristics

To illustrate the importance of molecular subtype in HPV + HNSCC research and translational studies, we tested for significant associations among subtype and 37 carefully selected clinical, demographic, and molecular variables. Of the variables tested, 22/37 (59.4 %) were available for all four cohorts. Overall, 21 variables were significantly associated with subtype (Fig. 3A). Known associations with subtype were reconfirmed including the association of IMU tumors with stronger EMT (p-value: 2.25×10⁻⁴), lower Chr16q copy number (p-value: 3.70×10⁻⁶), and heightened immune response as demonstrated by the significant associations with macrophages (p-value: 6.59×10⁻⁴), B cells (p-value: 1.73×10⁻⁴), B cell activation score (p-value: 8.67×10⁻⁹), dendritic cells (p-value: 1.58×10⁻¹¹), T cell activation score (p-value: 2.18×10⁻⁶), CD8 + T cells (p-value: 4.62×10⁻⁴), and CD4 + T cells (p-value: 5.48×10⁻⁶) (Fig. 3B). Also associated with IMU was the E6 full length (E6_FL) activation score (p-value: 1.61×10⁻¹²) [23] and the E6 full length ratio (calculated as E6_FL/E6_ALL) influence score [19] (p-value: 1.18×10⁻¹⁰), which estimate the activity level of the HPV oncogene E6 and the proportion of expressed E6 that is not an E6* isoform, respectively. Thus, KRT tumors had more E6* influence. We reconfirmed the associations of KRT tumors with heightened keratinization (p-value: 2.53×10⁻⁸), a high probability of expressed HPV integration (p-value: 3.53×10⁻⁶) and copy number gains in Chr3q (p-value: 0.011).

In addition to our confirmatory results, we discovered novel associations with subtype. We performed recurrence analysis with the UM67 cohort, which was the only cohort with recurrence information available, finding that KRT patients were more likely to recur using a Cox proportional hazards model controlling for tumor stage (p-value: 0.050; HR = 0.26; log-rank p-value: 0.081) (Fig. 4A). Examining the types of recurrence (local, regional, and distant), we found that participants with KRT tumors had a higher recurrence rate in all categories as opposed to being concentrated in one type (Supplementary Table S6). In line with this result, we found that KRT has higher estimated radiation resistance (p-value: 0.0050) (Fig. 4B) and higher AJCC clinical stage (p-value: 0.039) than IMU (Fig. 4C). We also found a novel association between sex and subtype (p-value: 0.0082) demonstrating females are more likely to be KRT than IMU (Fig. 4D). No association was found between subtype and p53 mutational status, drinking status, smoking status, packs per year smoked, genomic instability, N stage, T stage, respiratory electron transport chain, 3-year survival (or overall survival using Cox proportional hazards), ACE score, or oxidative phosphorylation (Supplementary Table S7).

Fig. 4.

(A) Kaplan-Meier curve of recurrence probability for IMU vs. KRT with log-rank test p-value. (B) Raincloud plot of subtype and radiation resistance corrected for cohort effect. P-values were calculated by t-test. (C) Bar plot of stage by subtype. (D) Bar plot of sex by subtype.

Implementation of the subtype classifier

The classifier is available as python-based models on GitHub (https://github.com/shengzhulst/IMUKRTclassifier), and as a webtool (https://hpv-hnscc-subtypeclassifier.dcmb.med.umich.edu/). The webtool can remove batch effects and display results as tables, PCA visualizations, and an interactive heatmap, assisting users in evaluating the classifier’s performance (see Methods).

Discussion

Consistently, HPV + HNSCC tumors have been characterized as either immune strong (IMU) or highly keratinized (KRT) [12]. These two subtypes have been further characterized based on mutations, CNVs, DNA methylation [16], DNA hydroxymethylation [17], and comprehensively reviewed in Qin et al [9]. However, unsupervised clustering cannot provide a standardized tumor classification. The IMU/KRT classification framework described here builds on existing knowledge of HPV + HNSCC phenotypes and provides a method for subtyping future samples which will aid in disentangling HPV + HNSCC tumor heterogeneity.

After classifying 219 tumors as IMU/KRT, we examined features significantly correlated with each subtype. Our association tests confirmed that KRT is more likely to have Chr3q gains, where the gene PIK3CA resides, and that IMU is more likely to have loss of Chr16q, where several tumor suppressor cadherin genes reside including E-cadherin and P-cadherin. This is consistent with previous findings that KRT is more likely to have activating PIK3CA mutations and that IMU is associated with an EMT signature with a switch from E-cadherin to N-cadherin. Our findings also revealed that females are more likely be KRT, and KRT tumors tend to be more radiation resistant and diagnosed at a higher stage. Although Locati et al found that patients with IMU-like tumors have better survival than KRT-like tumors (defined as CI1 (immune strong) vs. CI2 and CI3 (highly keratinized)) [13], we did not find a significant association with 3-year survival. However, we did find that KRT tumors are more likely to recur by approximately 75 % (HR = 0.26). The lack of association with overall survival in this cohort may, in part, be because 6 of the 13 recurrences (46 %) did not result in an observed death, and 4 of the 11 deaths (36 %) had no observed recurrence. HPV-negative HNSCC tumors are more likely to progress due to local invasion, whereas HPV + HNSCC tumors are more likely to progress due to distant metastasis [24]. Future work testing whether IMU tumors are more likely to lead to distant metastasis and KRT tumors to local spread would be well-motivated, given the closer overall resemblance of KRT to HPV-negative oropharynx tumors and the higher EMT signature of IMU. In addition, given the lower radiation resistance signature and high immune cell infiltration of IMU tumors, one may hypothesize that IMU patients with N0 nodal status may be candidates for de-escalation trials.

Unsupervised clustering applied to HPV + HNSCC RNA-seq identified generally reliable subtypes, but it lacked reproducibility [9,20]. To overcome this, as outlined in Supplementary Table S4, we implemented multiple ML algorithms, used multiple input gene sets to increase generality and robustness, trained on data from multiple cohorts, and tested the consistency between FFPE and FF samples. To minimize batch effects, we use Combat-seq [25] and provide a core set of samples to offset batch effects in new samples. Finally, we assessed the classifier in new cohorts and demonstrated classifier stability.

Sixty-seven (sixty-six non-Hispanic white) samples were used for training, which is relatively small, representing a limitation of this study. However, the steps to enhance robustness compensated for this, and we benefitted from the relative ease that the subtypes can be distinguished, as demonstrated by the large number of differentially expressed genes found by multiple studies and consistent rediscovery of the subtypes [11–13]. It will be useful to test these classifiers on well-annotated cohorts from other populations, especially understudied populations with survival disparities such as African American. For the validation cohorts (UM67, HVC), we lacked complementary assays that could support or expand our work, and clinical data was not available for HVC. Finally, 10–20 % of the samples could not be classified with a unanimous vote, however 100 % accuracy was achieved with the ensemble majority voting scheme, which is likely due to these tumors expressing a combination of IMU and KRT features or exhibiting a third rarer phenotype which cannot be easily disentangled.

Our use of bulk RNA-seq was effective but limiting in terms of studying within-tumor heterogeneity and its effects. Single-cell and spatial transcriptomics with clinical data will provide new opportunities to understand tumor subtype formation, heterogeneity, and within-tumor correlates [26,27]. We hope this classifier will facilitate future discoveries for HPV + HNSCC and lead to tailored avenues for treatment, prevention, and early detection.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.oraloncology.2025.107726.

Data sharing Statement

The UM67 cohort data, with aligned bam files are available at the European Genome-Phenome Archive as EGAS50000000893.