Immune checkpoint inhibitors (ICIs) have significantly extended the survival of patients with advanced melanoma. However, approximately 40% of melanoma patients will have minimal benefit from ICIs and more than 50% will progress on therapy [1]. There is an urgent need to identify robust biomarkers predictive of ICI response to optimize outcomes for each patient. Multiple studies have considered clinicopathological correlates, genomic features and transcriptomic signatures [2,3,4] but these often fail to predict ICI response in independent cohorts. For instance, the IPRES transcriptome signature [3] was not predictive of ICI response in multiple melanoma cohorts [5, 6]. Tumour mutation burden and PD-L1 expression are commonly used ICI response biomarkers but show predictive value only in select cancer types such as non-small cell lung cancer and head and neck squamous cell carcinoma (reviewed in [7]). Attempts to integrate multiomic features can improve ICI predictive accuracy [6], but these approaches can be costly and are difficult to replicate and interpret.

Currently, immune-associated transcriptomic signatures (e.g. IFNγ, IMPRES, TIDE) remain the most cited and utilized predictive biomarkers of ICI response [8,9,10]. Interestingly, despite preclinical and clinical studies demonstrating that lower tumour volume is associated with better response to ICIs [11], tumour volume is not commonly considered when evaluating predictive signatures of ICI response.

In this study, we performed RNA sequencing on pre-treatment (PRE, 1-1162 days before treatment start) melanoma specimens from 32 patients treated with combination PD1 and CTLA4 inhibitors to identify transcriptomic features associated with ICI response. Eleven patients (11/32, 34%) also had early during treatment (EDT, 5–15 days after treatment start) tumour samples (Supplementary Table 1). Tumour volume was calculated by measuring the longest diameter of all measurable lesions and presented as either total tumour volume (sum of tumour diameters) or average tumour volume (total tumour volume divided by the number of metastases). Tumour volume was determined at PRE for all 32 patients, and at first progress computed tomography (CT) imaging for the 11 patients with EDT samples (Supplementary Table 2). Of the 32 patients, 24 were categorized as responders (irRECIST complete response (CR, n = 3), partial response (PR, n = 20), and stable disease > 6 months (SD, n = 1)) and eight were considered non-responders (irRECIST progressive disease (PD, n = 6) and PR with progression free survival (PFS) of < 6 months, n = 2).

Predictably, total PRE tumour volumes correlated with average PRE tumour volumes and the number of metastases (Supplementary Fig. 1A). Total PRE tumour volume was also associated with LDH status but not ICI response (Supplementary Fig. 1B-C).

At PRE, analysis of the Hallmark Interferon Gamma (IFNγ) Response gene set (single sample Gene Set Enrichment Analysis (ssGSEA) score) revealed no significant difference between responders and non-responders (AUC ROC curve = 0.6406, p = 0.2400, 63% sensitivity, 71% specificity; Fig. 1A). However, when IFNγ signalling was considered along with baseline total tumour volume (ssGSEA score divided by the total tumour volume) as predictors in a logistic model with ICI response as dependent variable, the model demonstrated a clearer ability to discriminate responders from non-responders (AUC = 0.7760, p = 0.0211, 88% sensitivity, 67% specificity; Fig. 1A).

Tumour volume-normalised immune gene sets predict ICI response. A) Receiver operator characteristic (ROC) curves measuring the ICI predictive performance of each indicated immune-associated transcriptome gene sets with and without tumour volume normalisation (ssGSEA scores divided by total tumour volume) in melanoma tumours prior to treatment with combination ICIs (n = 32). B) ROC curves of Hallmark Interferon Gamma (IFNγ) Response gene set with and without tumour volume normalisation in melanoma tumours prior to treatment with PD-1 ICI (n = 23).

The improved predictive performance of tumour volume-normalised IFNγ response gene set score was reproduced for other immune-related transcriptome gene sets derived from the Molecular Signatures Database [12], including PID IFNγ pathway, PID CD8 TCR pathway, Biocarta TCR pathway, Biocarta CTL pathway, and KEGG antigen processing and presentation (Fig. 1A). Importantly, we validated the superior predictive performance of tumour volume-normalised IFNγ response gene set score (AUC = 0.7833, p = 0.0282, 88% sensitivity, 60% specificity) in a separate cohort of 23 melanoma patients prior to treatment with PD1 inhibitor [6] (Fig. 1B, Supplementary Table 3), confirming that this approach is generalizable to other melanoma cohorts, including those treated with anti-PD1 based monotherapy.

Furthermore, the model underwent internal validation within each cohort using bootstrapping techniques to assess the potential for overfitting and to measure optimism in the estimated performance metrics [13]. The differences between the observed AUC for each cohort and the corresponding average AUCs from 500 bootstrap replications were calculated. The bootstrapping result (Supplementary Fig. 2) also supported the integration of tumour volume and immune-associated transcriptomic signatures as predictors for ICI response. The mean differences between the observed AUC and the corresponding average AUCs from bootstrap replications were 0.003 (95% CI, -0.173 to 0.159) for the initial combination immunotherapy-treated cohort (n = 32) and 0.002 (95% CI, -0.184 to 0.163) for the PD1 inhibitor-treated cohort (n = 23).

We also examined tumour volume and Hallmark Interferon Gamma (IFNγ) Response gene set in the 11 PRE-EDT matched tumour samples (six responders, irRECIST PR, n = 5 and SD of > 6 months, n = 1, and five non-responders, irRECIST PD, n = 3 and PR with PFS of < 6 months, n = 2). Tumour volumes in responding versus non-responding patients did not differ at PRE, but on first progress imaging, as expected, tumour volumes were significantly smaller in the responding patients (p = 0.0053, Kruskal-Wallis test with Dunn’s multiple comparison, Fig. 2A).

Comparison of tumour volume and transcriptomic signatures in 11 matched PRE-EDT tumour biopsies. A) Change in total tumour volume from baseline and first progress CT imaging in responders and non-responders. Black coloured circles indicate tumours with < 50 mm in sum diameter at baseline, red coloured circles indicate the two non-responders with initial PR but PFS < 6 months, Kruskal-Wallis test with Dunn’s multiple comparisons, p = 0.0053. B) Change in expression of the Hallmark IFNγ response gene set (left) and the tumour-normalised Hallmark IFNγ response gene set (right) in matched PRE-EDT samples (n = 11, n = 6 responders, n = 5 non-responders). Black coloured circles indicate tumours with < 50 mm in sum diameter at baseline, red coloured circles indicate the two non-responders with initial PR but PFS < 6 months, Kruskal-Wallis test with Dunn’s multiple comparisons, p = 0.0124. C) ROC curve analysis of IFNγ-related and immune-associated transcriptomic signatures with and without tumour volume normalisation (ssGSEA scores divided by total tumour volume) in melanoma tumours early during treatment (EDT) with combination ICIs (n = 16)

The expression of the Hallmark IFNγ gene set was upregulated, from PRE to EDT, in 8/11 patients (4/6, 73% responders and 4/5, 80% non-responders). This increase was not associated with ICI response or with change in tumour volume. For example, three responding patients with small PRE tumour volumes (< 50 mm in sum diameter) showed minimal change in IFNγ signature expression from PRE to EDT (mean fold change of 1.04, range 0.93–1.18, compared to an overall mean fold change of 2.02, range 0.93–7.33, Fig. 2B), and these tumours also showed high baseline IFNγ activity. Comparison of the IFNγ signature at EDT did not separate responders from non-responders, and the two groups were only stratified when IFNγ signature at EDT was normalised to tumour volume on first progress imaging (p = 0.0124, Kruskal-Wallis test with Dunn’s multiple comparison, Fig. 2B). Notably, all non-responding patients including the two patients who had initial PR but PFS < 6 months, had significantly lower expression of the IFNγ signature normalised to tumour volume compared to responders at EDT (Fig. 2B). We included five additional EDT samples that had no patient-matched PRE tumour, and ROC curve analysis of the 16 EDT samples with tumour volume data at first progress imaging confirmed superior predictive performance when tumour volume was integrated with immune-associated transcriptomic signatures at EDT (n = 16, Fig. 2C).