Validation of a tissue-based predictive RNA test for immunotherapy benefit in NSCLC: the PREDAPT study

Abstract

Lung cancer remains the leading cause of cancer mortality worldwide. Immune checkpoint inhibitors (ICI) targeting PD-1/PD-L1 have significantly improved outcomes in a subset of patients. OncoPrism, a clinical test employing a multidimensional predictive RNA-based immune biomarker, was evaluated for predicting immune checkpoint inhibitor benefit in non-small cell lung cancer (NSCLC) patients. This study evaluated OncoPrism in 1487 patients across four NSCLC cohorts: one PD-L1 inhibitor cohort (n = 195), one PD-1 inhibitor cohort (n = 89), and two non-ICI cohorts (n = 193 and n = 1010). In the PD-L1 inhibitor cohort, OncoPrism predicted progression-free survival (p < 0.0001) and overall survival (p = 0.043). In the PD-1 inhibitor cohort, an observational clinical trial, PREDAPT, enrolling patients from 17 healthcare systems, OncoPrism predicted overall response rate (p = 0.008), progression-free survival (p = 0.004), and overall survival (p = 0.011). PD-L1 Tumor Proportion Score (TPS) was not predictive of response, progression-free survival, or overall survival. OncoPrism did not predict overall survival across two non-ICI NSCLC cohorts (p = 0.54, p = 0.73), suggesting the test is specifically predictive of ICI benefit rather than being prognostic with more limited clinical utility. Overall, the data show OncoPrism high patients have two to three-fold greater overall response rate, progression-free survival, and overall survival compared to those in other OncoPrism groups. These results underscore the impact of OncoPrism to address the current unmet need for ICI response prediction in NSCLC. This trial was registered with ClinicalTrials.gov, number NCT04510129, on 10 August 2020.

Introduction

Non-small cell lung cancer (NSCLC) affects over 230,000 individuals annually in the United States, resulting in around 127,000 deaths^1,2. The nearly half of all patients diagnosed at stage IV have a dire prognosis^3,4. The five-year overall survival (OS) rate for NSCLC is around 27%, with the 5-year relative survival rate for patients with distant metastases being just 9%^2,5.

Current treatment options for late-stage NSCLC include surgical intervention, radiation therapy, chemotherapy, and targeted drug therapies. The introduction of immune checkpoint inhibitors (ICI), such as anti-PD-1 and anti-PD-L1 antibodies, has transformed the therapeutic landscape, approximately doubling five-year OS versus platinum-based chemotherapy in some contexts^6–8. However, only a minority of ICI-treated patients achieve durable benefit, underscoring the urgent need for precision medicine approaches that identify the subset of NSCLC patients most likely to benefit from ICI.

Despite considerable effort over the past decade, progress toward effective ICI predictive biomarkers in NSCLC has been limited^9,10. PD-L1 Tumor Proportion Score (TPS) remains the only predictive clinical biomarker consistently recommended to guide ICI treatment decisions in NSCLC patients across leading treatment guidelines (NCCN, ASCO, ESMO)^11–13. Unfortunately, PD-L1 TPS has only modest positive predictive value, often resulting in escalation of treatment to ICI plus chemotherapy regimens. Despite significant efforts in the field, other potential biomarkers have shown limited clinical utility. Mutations in genes such as STK11 and KEAP1 are associated with poor response to ICI, but their utility as predictive rather than prognostic markers is not clear¹⁴. High specificity predictive biomarkers positively associated with ICI response are lacking. Microsatellite instability-high (MSI-H) is rare in NSCLC (<1%)¹⁵. Despite initial promise, tumor mutational burden (TMB) has been diminished from national guidelines due to modest correlation with response, issues of assay standardization, and inconsistent TMB-high cutoffs¹⁶. As a result, clinicians are left with PD-L1 TPS as the sole tool to predict response to ICI.

To make precision medicine a reality for NSCLC patients facing ICI therapy decisions, new predictive biomarkers must improve accuracy with balanced improvement in both sensitivity and specificity beyond PD-L1 TPS. New biomarkers must also demonstrate robust analytical performance across commonly encountered pre-analytic variables in clinical testing. Furthermore, such tests must be compatible with limited formalin-fixed, paraffin-embedded (FFPE) tissue specimens to work with existing clinical workflows and enable widespread adoption. Finally, predictive biomarkers of ICI response must show clinical specificity rather than simply being prognostic¹⁷. Predictive biomarkers are specific to the benefit from one class of therapy whereas prognostic biomarkers are predictive of outcome regardless of therapy. Unlike prognostic biomarkers, predictive biomarkers are clinically actionable to improve patient outcomes^18–20.

Previously, we reported the clinical and analytical validation of OncoPrism-HNSCC, a test that addresses the ICI-predictive diagnostic needs detailed above in head and neck cancers. OncoPrism-HNSCC is a robust, novel ICI predictive diagnostic employing RNA, machine learning, and an immune-based multidimensional biomarker capable of predicting clinical benefit from ICI in recurrent or metastatic head and neck squamous cell carcinoma (RM-HNSCC)^21,22. This test uses RNA expression data from the immune microenvironment of FFPE tumor samples to categorize patients based on likelihood of disease control in response to ICI with high sensitivity and specificity. OncoPrism is a unique predictive classifier that, at its core, is built from expression signatures of immune-related genes with observed expression across numerous cancers. Therefore, we hypothesized that OncoPrism could predict ICI response in indications beyond RM-HNSCC. We sought to evaluate whether the OncoPrism biomarker predicts response in NSCLC. Here, we report the tuning and validation of an NSCLC-specific test, OncoPrism-NSCLC, using the same underlying biomarker described previously.

Results

Establishing an NSCLC ICI prediction model

To establish the OncoPrism biomarker presence and differences across cancer types, OncoPrism scores were calculated using the previously published OncoPrism-HNSCC biomarker in TCGA patient data across six cancer indications^21,22. Mirroring the varied composition of tumor-immune microenvironments across tumor types, as well as differences in response rates to ICI, the population distribution of OncoPrism scores varied across cancer indications (Fig. 1A). These differences in population score distributions, patient response rates, and clinical practice in each indication suggested that a uniform approach is not the optimal diagnostic strategy for maximal predictive accuracy across cancer indications. Therefore, a new distribution was created using 1010 NSCLC patient samples to tune the biomarker underlying OncoPrism-HNSCC for use in NSCLC patients. This NSCLC-specific distribution accounts for the population-level characteristics of the NSCLC tumor-immune microenvironment. Thresholds were set to divide the lower, middle, and high scoring tertiles (i.e., one-third each) of this NSCLC patient population. This resulted in OncoPrism low (0–40), medium (41–57), and high (58–100) groups (Fig. 1B). These score thresholds and resulting OncoPrism groups were used for all subsequent work to interrogate the correlation between OncoPrism group and clinical benefit from ICI.

Fig. 1. The OncoPrism score distribution varied by indication.

A The distribution of OncoPrism scores, scaled relative to the entire dataset rather than the specific indication, is shown for TCGA data from six indications: lung adenocarcinoma/lung squamous cell carcinoma (LUAD/LUSC); kidney renal clear cell carcinoma (KIRC); head and neck squamous cell carcinoma (HNSC); colon adenocarcinoma (COAD); triple negative breast cancer (BRCA-TNBC); and bladder carcinoma (BLCA). B OncoPrism-NSCLC scores were scaled using TCGA LUAD and LUSC data. The 33rd and 67th percentile values were used as the thresholds (dashed lines) between low (0–40), medium (41–57), and high (58–100) OncoPrism groups. C OncoPrism predicted progression-free survival (PFS) in patients treated with anti-PD-L1 therapy (OAK ICI cohort, n = 195). D OncoPrism predicted overall survival (OS) in patients treated with anti-PD-L1 therapy (OAK ICI cohort, n = 195).

Validation of OncoPrism-NSCLC in PD-L1 inhibitor cohort

Having established an NSCLC-specific OncoPrism model, we evaluated its performance in a cohort of 195 Stage IIIB or IV NSCLC patients treated with anti-PD-L1 ICI after one or two previous chemotherapy regimens in the OAK trial (NCT02008227)²³. OncoPrism high patients had significantly longer progression-free survival (PFS; median = 4.7 months) than OncoPrism low (median = 2.8 months) and medium (median = 1.5 months) patients (Fig. 1C, p < 0.001). The univariate Cox proportional hazard ratio for OncoPrism high PFS compared to low plus medium was 0.52 (95% confidence intervals [CI] 0.37–0.72, p < 0.001). Likewise, OS was longer for OncoPrism high patients (median = 20.4 months) than low (median = 12.7 months) and medium (median = 10.5 months) patients (Fig. 1Dp = 0.043). The univariate Cox proportional hazard ratio for OncoPrism high OS compared to low plus medium was 0.61, 95% CI 0.41–0.90, p = 0.014).

Validation of OncoPrism-NSCLC in the multicenter PREDAPT PD-1 inhibitor cohort

Given the predictive value of OncoPrism in the OAK anti-PD-L1 cohort, we proceeded to evaluate the performance of OncoPrism-NSCLC in a multicenter, retrospective clinical trial, PREDAPT (NCT04510129). A total of 176 patients were enrolled from 17 sites at the time of this readout. Samples from 89 patients across 12 sites met all eligibility and quality control criteria and were included in the validation study (Fig. 2 and Table S1). Patient clinical characteristics are shown in Table 1, alongside the 154 patients in the ICI arm of the KEYNOTE-024 trial that compared outcomes in pembrolizumab vs. chemotherapy-treated patients²⁴.

Fig. 2. Study overview.

1574 patient samples were screened or used in this study. TCGA patient samples (n = 1010) were used to normalize the data and set NSCLC thresholds. The test was validated using 1487 patient samples across four NSCLC cohorts: PD-L1 inhibitor-treated patients (OAK ICI cohort, n = 195), PD-1 inhibitor-treated patients (PREDAPT ICI cohort, n = 89), and two non-ICI cohorts (OAK and TCGA, n = 193 and n = 1010, respectively). Of 176 samples assessed for inclusion in the PREDAPT trial, 46 were excluded for eligibility reasons and 41 failed quality control (QC).

Table 1.

Patient clinical characteristics

Characteristic	Levels	PREDAPT	KEYNOTE-024
Age (median)	Age	67	64.5
Sex	Male	51 (57.3%)	92 (59.7%)
	Female	38 (42.7%)	62 (40.3%)
ECOG	Not available	10 (11.2%)	-
	0	19 (21.3%)	54 (35.1%)
	1	55 (61.8%)	99 (64.3%)
	2	5 (5.6%)	2 (0.6%)
Smoking status	Used to smoke cigarettes	50 (56.2%)	115 (74.7%)
	Currently smoke cigarettes	22 (24.7%)	34 (22.1%)
	Not known	11 (12.4%)	-
	Never smoked cigarettes	6 (6.7%)	5 (3.2%)
Histology	Non-squamous	63 (70.8%)	125 (81.2%)
	Squamous	26 (29.2%)	29 (18.8)
Therapy	ICI plus chemo	63 (70.8%)	-
	Mono ICI	26 (29.2%)	154 (100%)
PD-L1 TPS	<1	37 (44.6%)	-
	1–49	18 (21.7%)	-
	50+	28 (33.7%)	154 (100%)

Because score thresholds between OncoPrism groups were set using the tertile scores in the reference dataset, the expected distribution of patients was one-third of patients in each OncoPrism group. In the validation cohort, 28% of patients were in the low group, 40% of patients were in the medium group, and 31% of patients were in the high group. Median length of follow-up was 9.3 months.

A significant trend towards higher overall response rate (ORR) was observed from the low to medium to high OncoPrism groups (p = 0.0083; Fig. 3A). The OncoPrism high ORR was 71%, compared to just 36% for both OncoPrism low and medium. The overall study ORR was 47%. Similarly, patients in the OncoPrism high group also had significantly longer PFS (median = 18.6 months) than patients in the low and medium groups (median = 7.0 and 5.5 months, respectively; p = 0.0038, Fig. 3C). The univariate Cox proportional hazard ratio for OncoPrism high PFS compared to low plus medium was 0.37 (95% CI 0.2–0.678, p = 0.001; Fig. S1A). Likewise, the higher ORR and PFS corresponded with significantly longer OS in OncoPrism high patients (median = 22.9 months) relative to low (median=13.0) and medium (median = 9.2) patients (p = 0.011, Fig. 3E). The univariate Cox proportional hazard ratio for OncoPrism high OS compared to low plus medium was 0.39 (95% CI 0.389, CI 0.206–0.734, p = 0.004; Fig. S1B). To assess the robustness of the validation results, we performed a bootstrap analysis with 1000 iterations. The PFS and OS log-rank tests were statistically significant in 87% and 82.6% of iterations, respectively.

Fig. 3. OncoPrism group, but not PD-L1 TPS, is correlated with clinical benefit.

OncoPrism-NSCLC (n = 89, left) and PD-L1 TPS (n = 83, right) were evaluated for their ability to predict clinical benefit. A, B OncoPrism-NSCLC group was significantly correlated with overall response rate (ORR), while PD-L1 was not. The ORR for the entire study population is shown as a dashed line. C, D OncoPrism-NSCLC, but not PD-L1 TPS, significantly predicted progression-free survival (PFS). E, F OncoPrism-NSCLC, but not PD-L1 TPS, significantly predicted overall survival (OS).

OncoPrism-NSCLC vs. PD-L1 TPS predictive measures

Next, we compared the performance of OncoPrism-NSCLC to the performance of PD-L1 in the PREDAPT study. PD-L1 TPS results were not available for 6 patients, so they were excluded from the PD-L1 analysis. OncoPrism-NSCLC results for the remaining 83 patients are shown in Fig. S2. The commonly used TPS categories of TPS < 1, TPS 1–49, and TPS ≥ 50 were used to group patients. Unlike OncoPrism-NSCLC (Figs. 3A, C, E and S2), PD-L1 TPS was not predictive of ORR, PFS, or OS (Fig. 3B, D, F) in these 83 patients. OncoPrism-NSCLC also had numerically higher accuracy, sensitivity, specificity, positive predictive power, and negative predictive value than PD- L1 TPS (Table S2). In the 23 patients treated with monotherapy ICI with PD-L1 TPS data, OncoPrism-NSCLC had significantly higher specificity than PD-L1 TPS (McNemar test, p = 0.04) and numerically higher accuracy, positive predictive power, and negative predictive power. PD-L1 TPS had numerically higher sensitivity (Table S3).

To examine the cohort results at the individual patient level, we plotted all 89 patients sorted by their OncoPrism group, response label, PD-L1 status, treatment regimen and disease subtype (Fig. 4A). We tested for significant associations among these variables using the chi-squared test of independence. There was a significant association between OncoPrism group and response (p = 0.008), with an enrichment of responders in the OncoPrism high group. While the majority of these OncoPrism high true responders are TPS ≥ 50, both TPS < 1 and TPS 1–49 patients were also represented among these OncoPrism high true responders (Fig. 4B). Overall, there was a significant but modest association between OncoPrism-NSCLC and PD-L1 TPS (p = 0.006). PD-L1 TPS category was also associated with treatment type with more monotherapy patients among TPS ≥ 50 patients (p < 0.001). Finally, OncoPrism group was associated with disease histology with fewer squamous cancers among OncoPrism high patients (p = 0.028).

Fig. 4. Associations among clinical features.

A Patient-level data for OncoPrism group, response, PD-L1 TPS category, therapy regimen, and histology type. Categorical variables with significant associations are noted (*p < 0.05, chi-squared test of independence). B Number of PREDAPT patient samples in each PD-L1 TPS category (TPS < 1, 1–49, ≥50, and unknown) by OncoPrism group.

Given these associations, we used multivariate Cox proportional hazards models to confirm that OncoPrism-NSCLC independently predicts PFS and OS in the 83 patients with complete data (Tables S4 and S5). Across the variables OncoPrism group, treatment type, PD-L1 TPS and histology, only OncoPrism high vs. medium plus low had a significant hazard ratio for PFS (HR = 0.394, 95% CI 0.201–0.772, p = 0.007) and OS (HR = 0.407, 95% CI 0.204–0.81, p = 0.011).

Table S6 compares predictions and outcomes for OncoPrism-NSCLC and PD-L1 TPS. While OncoPrism-NSCLC and PD-L1 TPS were concordant for 58/83 results, for discordant results PD-L1 had both more false negatives (9 vs. 5) and more false positives (8 vs. 3) than OncoPrism-NSCLC. Responders were treated as the positive class and non-responders were treated as the negative class. Of the 22 responders in the OncoPrism-NSCLC low and medium groups (i.e., false negatives), only 3 were treated with monotherapy ICI. Likewise, only 2 of 26 PD-L1 TPS false negatives were treated with monotherapy ICI, suggesting that response to chemotherapy may be driving many of the false negatives for both OncoPrism-NSCLC and PD-L1 TPS.

Assessing predictive vs. prognostic characteristics of OncoPrism-NSCLC

To determine if OncoPrism-NSCLC was predictive of response to ICI or prognostic of outcome generally, we evaluated the performance of OncoPrism-NSCLC in patients who were not treated with ICI. OncoPrism-NSCLC did not predict PFS (HR = 1.07, 95% CI 0.78–1.46, p = 0.685) or OS (HR = 1.08, 95% CI 0.7603–1.533, p = 0.669) in patients treated with docetaxel (OAK non-ICI, n = 193, Fig. 5A, B). Likewise, OncoPrism-NSCLC was not predictive of OS (HR = 0.9056, 95% CI 0.7294–1.124, p = 0.369) in the TCGA data, composed of patients treated with a variety of non-ICI therapies (TCGA non-ICI, n = 1010, Fig. 5C)²⁵. Together, these data indicate that OncoPrism-NSCLC is not broadly prognostic and may be predictive of ICI specifically.

Fig. 5. OncoPrism-NSCLC does not predict survival in non-ICI-treated patients.

OncoPrism group was not correlated with (A) progression-free survival (PFS) or (B) overall survival (OS) in docetaxel-treated NSCLC patients (n = 193). C OncoPrism-NSCLC was not predictive of OS in non-ICI treated patients from TCGA NSCLC data (n = 1010).

Characterizing robustness through assessment of analytical variation

To evaluate the analytical performance of OncoPrism-NSCLC, each of two operators processed three replicates from seven different biological samples with a range of OncoPrism scores (23–75) on different days using two different reagent lots (Fig. S3). The overall variation was low, with a pooled standard deviation of just 0.615 (95% confidence interval: 0.504–0.790) and a categorical call concordance of 38/40 (95%). Both discordant replicates were from a single biological sample that straddled the low to medium threshold. Two separate processed replicates did not pass QC and were not included in this analysis.

Discussion

Using the previously published OncoPrism biomarker capable of predicting ICI response in HNSCC patients (OncoPrism-HNSCC), we optimized and validated OncoPrism for NSCLC (OncoPrism-NSCLC)²¹. OncoPrism-NSCLC maintains the clinical utility for predicting response to ICI treatment beyond standard PD-L1 IHC testing (Fig. 3), along with the robust analytical performance (Fig. S3) displayed by OncoPrism-HNSCC^21,22.

OncoPrism-NSCLC predicted PFS and OS in the OAK anti-PD-L1 ICI cohort, as well as ORR, PFS, and OS in the 17-site PREDAPT anti-PD-1 ICI cohort (Figs. 1C, D and 3A, C, E). By contrast, PD-L1 TPS failed to predict ORR, PFS, or OS in the PREDAPT anti-PD-1 cohort (Fig. 3B, D, F). Importantly, while OncoPrism-NSCLC predicted PFS and OS in the two ICI-treated patient cohorts, it did not predict PFS or OS in two non-ICI cohorts (Fig. 5). This finding was consistent with results from the OncoPrism-HNSCC study, where PFS or OS were not predicted for four non-ICI cohorts, indicating that the biomarker underlying the OncoPrism test is predictive of ICI benefit rather than prognostic of outcome regardless of treatment²¹. Being predictive rather than prognostic augments the clinical utility of the OncoPrism test. Lack of predictive value in non-ICI patients also suggests that the predictive value of OncoPrism-NSCLC in chemo-immunotherapy treated patients may be derived from the ICI component of the regimen. Among relevant clinical factors, OncoPrism-NSCLC group (high vs. low plus medium) was the only significant variable in multivariate Cox proportional hazard models for PFS and OS, reflecting OncoPrism-NSCLC’s independent predictive value (Tables S4 and S5).

While this study met its primary endpoint of predicting ORR, there are several limitations and opportunities for future directions. First, the retrospective design of the study may have introduced patient selection bias and unmeasured clinical covariates, particularly in the context of variable treatment regimes. A prospective study would strengthen the conclusions of this validation. Second, while performance in the OAK cohort and the PREDAPT cohort were similar, the size of the PREDAPT cohort (n = 89) limits the precision of the performance estimates. A larger cohort would potentially narrow the confidence intervals on the univariate and multivariate Cox proportional hazard models and possibly allow for identification of additional significant variables. Recruitment of additional patients for the PREDAPT study is ongoing. Third, this study currently lacks a stand-alone monotherapy ICI cohort, although OncoPrism-NSCLC did have a significant hazard ratio for both PFS and OS in monotherapy patients, as well as other subgroups such as PD-L1 TPS ≥ 50 (Fig. S1A, B). The ongoing recruitment of larger cohorts for the PREDAPT study will allow for more meaningful subgroup analyses in the future.

Because the PREDAPT cohort includes patients treated with ICI as a single agent and in combination with chemotherapy, there is a risk that treatment regimen biases the outcome, particularly ORR, across OncoPrism groups. Specifically, chemo-immunotherapy patients have a higher ORR than monotherapy patients (51% for chemo-immunotherapy therapy vs. 38% for monotherapy). However, there is no significant association between OncoPrism group and treatment regimen, indicating that the high ORR in OncoPrism high patients is not driven by treatment regimen (Fig. 4A). Conversely, PD-L1 TPS is associated with treatment regimen, with an enrichment in monotherapy-treated patients for those with TPS ≥ 50 (Fig. 4A). This association is unsurprising since PD-L1 TPS ≥ 50 is used to select for monotherapy treatment, and it likely explains the high sensitivity and low specificity for PD-L1 TPS in monotherapy patients (Table S3). PD-L1 TPS 1-49 and TPS ≥ 50 have similar ORRs, which may be explained in part by the higher proportion of PD-L1 TPS 1–49 patients who received chemo-immunotherapy. The inclusion of chemo-immunotherapy patients and the association of treatment with PD-L1 TPS also may confound comparisons between the performance of OncoPrism-NSCLC and PD-L1 TPS.

While significant, there is a relatively modest correlation between OncoPrism-NSCLC group and PD-L1 TPS category (Fig. 4A, B). Further, the OncoPrism high group was associated with improved PFS and OS in PD-L1 TPS ≥ 50 patients (Fig. S1A, B). Together, these results suggest potential complementary use of OncoPrism-NSCLC together with PD-L1 to improve predictive accuracy relative to either test alone. Similarly, future work will interrogate any relationship between OncoPrism-NSCLC and mutations in genes shown to be correlated with poor response to ICI (KEAP1 and STK11) and whether there is opportunity to further improve performance¹⁴.

In summary, OncoPrism-NSCLC is a laboratory developed test (LDT) that fulfills the current need for better predictive tests for NSCLC patients being considered for ICI. As reported here, OncoPrism-NSCLC is: (a) compatible with limited tissue from FFPE tumor biopsies; (b) robust against variation in the context of testing variables; (c) highly specific with good sensitivity, especially compared to PD-L1 TPS; and (d) predictive rather than prognostic, providing opportunity for greater clinical utility. The test can be implemented into standard clinical workflows with clinically actionable turnaround times similar to existing NGS molecular diagnostics using routinely collected biopsy material. These characteristics support OncoPrism-NSCLC’s real-world feasibility as a predictive diagnostic.

Methods

Datasets

To construct the TCGA non-ICI Cohort (n = 1010), RNA-seq counts and clinical data for TCGA disease codes LUAD and LUSC were downloaded using TCGAbiolinks²⁵. We filtered for primary tumor samples and filtered out technical replicates such that only one sample per patient was included in the final dataset. This resulted in a final set of n = 1010. Access to the OAK trial (NCT02008227) clinical sequencing data and clinical outcome data was granted by Genentech²³. The OAK non-ICI Cohort (n = 193) used for prognostic evaluation represented the arm of the trial treated with docetaxel 75 mg/m² (non-ICI) every 3 weeks until progressive disease or unacceptable toxicity. Two-thirds of the data was randomly selected for use, with one-third withheld for potential future analysis. This resulted in a final set of n = 193. The OAK ICI anti-PD-L1 Cohort (n = 195) represented the arm of the trial treated with atezolizumab 1200 mg (ICI/anti-PD-L1) every 3 weeks until progressive disease or unacceptable toxicity. Two-thirds of the data was randomly selected for use, with one-third withheld for potential future analysis. This resulted in a final set of n = 195.

Biomarker distribution and tertile score determination (OncoPrism-NSCLC)

The distribution of OncoPrism scores across six cancer types was determined using TCGA datasets for LUAD plus LUSC, KIRC, HNSC, COAD, the triple-negative breast cancer subset of BRCA, and BLCA (Fig. 1A). The distribution of OncoPrism scores varied across cancer types. Because we sought to evenly divide the NSCLC population into three groups, an NSCLC-specific score distribution was determined using a cohort of 1010 TCGA (LUAD plus LUSC) patient samples and divided into tertiles using the 33rd and 67th percentile score as thresholds for each group.

OncoPrism-NSCLC was built using the same biomarker as the previously published OncoPrism-HNSCC but tuned for use in NSCLC²¹. Specifically, the thresholds were adjusted to an NSCLC patient population as described above, the OncoPrism groups were divided into tertiles rather than 25%, 25%, and 50% for the low, medium, and high groups, respectively, and the test was built to predict ORR rather than disease control rate due to the higher response rates in NSCLC. OncoPrism-NSCLC was tuned using the 1010 TCGA (LUAD plus LUSC) dataset that acts as a reference distribution, with scaling of OncoPrism-NSCLC samples using the ICI-treated dataset collected via the PREDAPT trial. The PREDAPT cohort was not used for feature selection, model training, threshold selection or parameter optimization. To build the reference distribution, Log2 CPMs were calculated using protein coding genes in the TCGA dataset. Genes were scaled using the formula Z = (x − u)/s, where x is the normalized gene expression value of that gene in a given sample, u is the mean expression value of that gene across the entire dataset, and s is the standard deviation of the gene expression across the dataset. The transformed gene expression values were used as input to the previously published OncoPrism biomarker (OncoPrism-HNSCC), which generated values between 0 and 100²¹. The 33rd and 67th percentile OncoPrism scores in this NSCLC OncoPrism score distribution were used as the thresholds between low (0–40), medium (41–57), and high (58–100) OncoPrism groups. These score thresholds derived from the TCGA data were used as the categorical breakpoints for all subsequent datasets and evaluations.

The PREDAPT RNA-seq data (n = 89) was scaled the same way as the TCGA data, but independently. The IO data was then transformed using a min-max scalar fit from the TCGA data. Values that were greater than 100 or less than 0 were set to 100 and 0, respectively. This scaled value for a given sample is its OncoPrism score, which classified it into an OncoPrism group (low, medium, or high) based on the thresholds established in the TCGA data. To enable a locked-down model, the PREDAPT data are used for scaling not only the PREDAPT dataset, but also all subsequent test samples run through OncoPrism-NSCLC.

PREDAPT study design and participants

Patients for the PREDAPT ICI anti-PD-1 Cohort (n = 89) were recruited from the following 17 academic and community study sites across the United States, with the aim of a representative sample of the affected population: Appalachian Regional Healthcare (Hazard, KY), Cancer Care Northwest (Spokane, WA), Cox Medical Centers (Springfield, MO), Gundersen Medical Foundation (La Crosse, WI), John B Amos Cancer Center (Columbus, GA), Lahey Hospital and Medical Center (Burlington, MA), MultiCare Institute for Research and Innovation (Tacoma, WA), Ochsner Lafayette General Medical Center (Lafayette, LA), Providence Regional Cancer System (Lacey, WA), Touro Infirmary LCMC Health (New Orleans, LA), Valley Cancer Associates (Harlingen, TX), William Beaumont Army Medical Center (Fort Bliss, TX), Brooke Army Medical Center (Fort Sam Houston, TX), Decatur Memorial Hospital (Decatur, IL), Holy Cross Hospital (Fort Lauderdale, FL), Northwest Oncology and Hematology (Hoffman Estates, IL), Sharp Clinical Oncology Research (San Diego, CA).

Patients were enrolled from 2021 to 2025 in a retrospective, observational study. Research was conducted in accordance with the Declaration of Helsinki. No patient-level study data were reported to patients or physicians, and the patients and public were not involved in the study design. Patients were enrolled following the inclusion and exclusion criteria outlined below. Eligible patients were adults with stage IV non–small cell lung cancer (NSCLC) at diagnosis who had received at least two doses of anti–PD-1 immunotherapy (pembrolizumab) as first-line single-agent treatment (monotherapy) or in combination with chemotherapy (chemo-immunotherapy). Patients were required to have an Eastern Cooperative Oncology Group (ECOG) performance status of 0–2 prior to initiation of anti-PD-1 therapy as a single agent or 0-1 prior to initiation of anti-PD-1 therapy in combination with chemotherapy; if ECOG status was unavailable, documentation of self-care capacity and/or ambulatory activity for at least 50% of waking hours was accepted for monotherapy patients, or documentation of no requirement for assistance with daily living or activities for chemo-immunotherapy therapy patients. A pretreatment tumor biopsy was required, with at least 45 µm FFPE tissue available for analysis. Evaluation of tumor response to therapy by imaging and/or clinical assessment must have been performed, typically 1–4 months after initiation of therapy.

Key exclusion criteria included a history of stage III or later non-NSCLC malignancies, recurrent non-NSCLC cancers, progression from early-stage (stages I–III) NSCLC, receipt of fewer than two cycles of anti-PD-1 therapy or known targetable EGFR or ALK mutations. Patients who had previously received non-hormonal systemic therapy (e.g., platinum-based chemotherapy, targeted therapy, or immunotherapy) were excluded; prior hormonal therapy such as tamoxifen was permitted.

Tissue specimens analyzed in the study were collected from pre-treatment tumor samples originally processed as FFPE specimens using standard histologic protocols. De-identified, pre-treatment FFPE tumor biopsy specimens were provided to Cofactor Genomics for OncoPrism-NSCLC and PD-L1 IHC analysis. Following treatment, each patient’s tumor response to immunotherapy was evaluated using RECIST, PERCIST, or other clinical criteria as appropriate in standard of care to determine disease control. Patients with insufficient tissue for analysis or less than 10% neoplastic cellularity as determined by a study pathologist (EJD) were excluded from the study. Primary or metastatic tumor specimens were accepted, but metastatic tumors from brain, liver, or bone were not included. Length of follow-up ranged from 47 days to 68 months. The study protocol, “A Multicenter Cancer Biospecimen Collection Study” was registered on 10 August 2020 as “NCT04510129—PREDicting immunotherapy efficacy from Analysis of Pre-treatment Tumor biopsies (PREDAPT)” on clinicaltrials.gov. The study protocol was approved by institutional review boards at either the study (Advarra, Inc. [00051202] or WCG IRB [20201975]) or site level (Lahey Clinic, Inc. Institutional Review Board [20233025]; San Antonio Institution Review Board [977378]; Sharp Center for Research Institutional Review Board [2110806]; or Brooke Army Medical Center Human Research Protections Office [972580]), as appropriate. All patients provided signed informed consent to participate, or consent was waived for deceased patients according to study protocol. Independent data monitoring was conducted by the study clinical research organization Curebase, Inc. (San Francisco, CA).

RNA extraction

RNA was extracted using RNAstorm (Biotium, Fremont, CA) or the Zymo Quick-DNA/RNA FFPE Miniprep Kit (Irvine, CA) according to the manufacturer’s instructions for RNA isolation. RNA quantity was assessed by the High Sensitivity RNA Qubit assay (Thermo Fisher Scientific, Waltham, MA). A predefined yield of 20 ng FFPE RNA was used as the minimum QC threshold. Quality of the RNA was assessed using a bioanalyzer (Agilent Technologies, Santa Clara, CA), and a DV200 of at least 21% was required for all samples.

Library preparation and sequencing

Libraries were prepared using the QuantSeq 3′ mRNA-Seq Library Prep Kit FWD for Illumina with the QuantSeq-Flex Targeted RNA-Seq Module (Lexogen, Inc., Greenland, NH) following the manufacturer’s instructions. Library RNA input was 20–40 ng. UMI Second Strand Synthesis Module for QuantSeq FWD (Lexogen, Inc., Greenland, NH) replaced Second Strand Synthesis Mix 1 in the workflow. All samples were processed with two OncoPrism positive controls and a No Template Control. The positive (high or medium scoring) controls were RNA extracted from RM-HNSCC FFPE samples²³. Final libraries were sequenced to a minimum depth of 5 million single-end 75 base pair reads on a NextSeq500 (Illumina, San Diego, CA), following the manufacturer’s protocols.

Immunohistochemistry

PD-L1 staining and TPS were performed by NeoGenomics Laboratories (Fort Myers, FL) using the PD-L1 22C3 FDA for NSCLC stain. H&E staining was performed by NeoGenomics as part of the PD-L1 22C3 test or at Cofactor Genomics using xylene substitute Slide Brite (Newcomer Supply, Middleton, WI), as detailed by manufacturers. Neoplastic cellularity was estimated by a trained pathologist (EJD).

Processing of RNA sequencing data

FASTQ files were preprocessed with trim_galore/cutadapt version 0.4.1 to remove adapter sequences, reads with PHRED quality scores less than 20, and reads shorter than 20 base pairs. The trimmed reads were aligned to the human genome GRCh38 with STAR version 2.5.2a using the two-pass method as previously described²⁶. Read counts were generated using htseq-count version 0.9.1 and annotation from Gencode version 22²⁶. The data was normalized as counts per million (CPM) and log2 transformed using unique reads aligning to protein coding regions. Samples were required to have a minimum of 30% exonic alignment and 800,000 unique deduplicated counts to be included in the study.

Validation of PREDAPT cohort performance

Clinical validation of the OncoPrism-NSCLC test was performed using a cohort of 89 unique patient samples. Samples were processed in the Cofactor Genomics College of American Pathologists (CAP)-accredited, Clinical Laboratory Improvement Amendments (CLIA)-certified laboratory using strict quality controls. The primary validation metric was ORR in each OncoPrism group. ORR was calculated by dividing the sum of patients with RECIST 1.1-defined categories of partial response and complete response as initial response by the total number of patients in each group. RECIST label was determined one to four months after initiation of ICI treatment when possible, but one patient was evaluated at six months due to treatment regimen and availability for follow-up imaging. Operators were blinded to the RECIST label when processing samples and generating OncoPrism scores. The RECIST labels for each patient were determined independently from the OncoPrism group and were used to determine the ORR for each group in the validation set. Patients with complete response or partial response were treated as the positive class (“responders”). Patients with stable disease or progressive disease were treated as the negative class (“non-responders”).

Statistics

The primary endpoint of this study was overall response rate. A two-sided Cochran-Armitage test for trends was used to test the significance of the trend of increasing proportions for the ORRs of OncoPrism groups. Power analysis was performed using a range of potential AUCs of 0.68–0.72 and assuming a response rate of 50%. Under these assumptions, minimum cohort sizes of 60–100 were calculated for 80% power at p < 0.05 for the primary endpoint (Cochran-Armitage Test for trend in proportions for ORR). For PREDAPT patients PFS was defined as the time from start of ICI treatment to progression or death. OS was defined as the time from start of ICI treatment to death by any cause. Patients were censored if they had not progressed at last follow-up. PFS and OS figures and analysis were done using the “survminer” and “survival” packages, and significance was determined using log-rank methods^27,28. To test robustness, 1000 bootstrap datasets were generated. For each bootstrap dataset, survival analysis was performed. Hazard ratios were calculated for the OncoPrism high vs. medium plus low groups unless otherwise noted. 95% confidence intervals for model performance metrics were calculated using a non-parametric bootstrap resampling method. Pairwise associations among categorical variables were tested using a chi-squared test of independence. In all cases, a p-value of less than 0.05 was considered significant.

Author contributions

Conception and design: K.C.F., J.E., J.H., R.L.W., M.M.P., K.P., J.A.H., H.L.McL., E.J.D. and J.I.G. Recruitment of patients and acquisition of samples and data: M.M.P., J.I.G., O.M., K.W., A.W.P., K.D., X.S., W.L.A., J.S., S.J.S., T.D.S., S.D.B., S.A.S., G.A., B.B., S.E.K., P.G. and J.W.III. Execution of the research: K.C.F., J.E., J.H., R.L.W., M.M.P. and K.P. Analysis and interpretation of data: K.C.F., J.E., J.H. and J.I.G. Drafting of the manuscript: K.C.F., J.E., J.H. and J.I.G. Guarantor: J.I.G.

Data availability

All the relevant data supporting the findings of this study are provided within the main manuscript and its supplementary files. TCGA sequencing data is available from the TCGA consortium (https://www.cancer.gov/tcga). OAK sequencing data is obtained by request from Genentech through the European Genome-Phenome Archive (https://ega-archive.org/datasets/EGAD00001007703). PREDAPT sequencing data are derived from clinical samples. Due to IRB, hospital contracts, and HIPAA requirements we are not authorized to share individualized patient genomic data. Cofactor Genomics is dedicated to collaboration and the advancement of science. Reasonable data requests outside the scope of this manuscript should be addressed to the corresponding author.

Competing interests

K.C.F., J.E., J.H., R.L.W., M.M.P., K.P., and J.I.G. are employed by and have stock interests in Cofactor Genomics. J.A.H. has participated on advisory boards for Daiichi Sankyo and Boehringer Ingelheim. H.L.McL. consults for and has stock interest in Cofactor Genomics. E.J.D. consults for and has stock interest in Cofactor Genomics and received speaker fees from Blueprint Medicines and Syndax Pharmaceuticals. He also consults for and receives royalties from Caris Life Sciences and consults for Geneoscopy and Bruker Life Sciences. All others declare no competing financial or non-financial interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41698-026-01632-6.