Specific predictive effect of metastatic-organ landscape for ICI therapy
To delineate the immunotherapeutic predictive effect of organ metastases, we explored the impact of metastatic-organ landscape on survival outcomes in the atezolizumab arm and the docetaxel arm from OAK respectively. In terms of OS, survival prospects were significantly different among patients with various metastatic organs in the atezolizumab-treated population (P = 0.0105; Fig. 2A). Nevertheless, OS was generally similar in the docetaxel-treated population regardless of metastatic organs (P = 0.3742; Fig. 2B). Pairwise comparisons among metastatic organs in the atezolizumab arm showed that adrenal gland metastasis and brain metastasis yielded the best long-term survival benefits, with significantly longer OS compared with metastases to other organs, including liver, bone, and pleural effusion (the lower triangle in Fig. 2C). With regard to the docetaxel arm, however, no significant difference of OS existed between any pair of metastatic organs (the upper triangle in Fig. 2C).
Consistent findings were observed in terms of PFS, where significant survival difference was seen among atezolizumab-treated patients with various metastatic organs (P = 0.0167; Fig. 2D) but not in docetaxel-treated populations (P = 0.8242; Fig. 2E). In agreement with this, there only existed significant differences of PFS in pair-wise comparisons in the atezolizumab arm (the lower triangle in Fig. 2F) but not in the docetaxel arm (the upper triangle in Fig. 2F).
Metastatic-organ landscape as a predictor in a PD-L1 dependent manner
Upon comparing the efficacy following treatment with immunotherapy and chemotherapy, we found an identifiable association between the metastatic-organ landscape and the clinical benefits of atezolizumab versus docetaxel in the total population from OAK (Fig. 3A). On account of the clinical practice of immunotherapy in patients who were PD-L1 positive (≥ 1%), we investigated the predictive effect (immunotherapy versus chemotherapy) of the metastatic-organ landscape stratified by PD-L1 status. Intriguingly, the predictive effect was observed exclusively in the PD-L1-positive population (TC/IC ≥ 1%; Fig. 3B) rather than in the PD-L1-negative population (TC/IC < 1%; Fig. 3C).
In general, the predictive significance varied across metastatic organs at different degrees or even conversely within the PD-L1-positive population, which defined two organ categories (Fig. 3). For category I organs, OS benefits of atezolizumab versus docetaxel were found in patients whose tumors metastasized to adrenal glands (HR 0.42, 95% CI 0.25–0.70), brain (HR 0.41, 95% CI 0.19–0.88), and liver (HR 0.57, 95% CI 0.36–0.89) (Fig. 3B). On the contrary, for bone, pleural, pleural effusion, and mediastinum, patients harboring metastasis to any of these category II organs did not benefit from atezolizumab relative to docetaxel (Fig. 3B). Synergistic predictive effect was found among the category I organ metastases (adrenal glands, brain, and liver) within the PD-L1-positive population (Additional file 4: Fig. S1A-F), where the OS benefits of atezolizumab versus docetaxel in patients with metastases to double category I organs (HR 0.32, 95% CI 0.14-0.76, P = 0.0013; Additional file 4: Fig. S1C) were even more pronounced in comparison to those with metastasis to a single category I organ (HR 0.59, 95% CI 0.41-0.85, P = 0.0048; Additional file 4: Fig. S1B).
We also examined the predictive versus prognostic effect of the metastatic statuses of category I organs (Additional file 5: Fig. S2A-D). Remarkably within the PD-L1-positvie population, the presence of the category I organ metastases was associated with decreased OS in the docetaxel arm (P < 0.0001; Additional file 5: Fig. S2B), indicative of the inherent unfavorable prognostic effect of metastases. In contrast, OS was not influenced by the metastatic number of the category I organs in the atezolizumab arm (P = 0.5978; Additional file 5: Fig. S2A), suggesting that the unfavorable prognostic effect was alleviated by the favorable predictive effect specific to immunotherapy. Conversely, the inherent unfavorable prognostic effect of metastases became discernible in PD-L1-negative patients treated with atezolizumab, since the immunotherapeutic predictive effect was absent in this population; consequently, we observed a negative correlation between OS and the number of the category I organ metastases (P = 0.0005; Additional file 5: Fig. S2C).
Incorporating metastatic-organ landscape to forecast ICI therapy
The synergistic value of different metastatic organs inspired us to incorporate the whole landscape for comprehensive assessment. Having identified metastatic-organ landscape as a determinant of both inherent prognosis and immunotherapeutic benefit, we believed that only through incorporating both the treatment-independent prognostic effect and the immunotherapy-specific predictive effect of the whole metastatic-organ landscape could we truly reflect its impact on ICI therapy.
To this end, we proposed a metastasis-based scoring system (METscore) to forecast survival outcomes of advanced-stage NSCLC patients treated with ICI agents using the PD-L1-positive OAK cohort, since the therapeutic impact was particularly witnessed in this stratum (Fig. 4A). On this basis, given a metastatic profile of a certain patient obtained from pretreatment radiologic assessments, two scaled points could be assigned to each metastatic organ, one of which delineated the prognostic effect and the other delineated the predictive effect. The total points of prognostic and predictive effects of the whole metastatic-organ landscape constituted the METscore in forecasting the survival benefits following ICI therapy.
A range of putative cut-points (0 ~ 5) for METscore was assessed in the atezolizumab population (Fig. 4B). There was a general pattern of prolonged OS in patients with lower METscore relative to those with higher METscore, and an enlarged OS difference could be expected if more stringent cut-points (4 and 5) were selected (Fig. 4B). For practical application, the optimal cut-point of METscore, corresponding to 3 achieved by the maximally selected rank statistics, was determined by weighing the survival benefits against the minimal proportion of each group. And patients were classified into METscore-High (METscore ≥ 3) and -Low (METscore < 3) categories accordingly. Thereupon, the METscore system and the threshold were eventually locked for performance evaluation throughout the study. The system has been translated into a web-based tool that is freely available to the public (Additional file 6: Fig. S3).
METscore enables identification of therapeutic benefit from checkpoint blockade
We first evaluated the discrimination performance of METscore in the PD-L1-positive OAK cohort. Patients with METscore-Low obtained significantly longer OS than METscore-High counterparts in the atezolizumab-treated population (HR 0.48, 95% CI 0.32–0.72, P < 0.0001; Fig. 5A). By contrast, there was no significant difference in terms of OS between METscore-Low and -High groups in the docetaxel-treated population (HR 0.73, 95% CI 0.49–1.07, P = 0.0805; Fig. 5B). Concurrently, upon direct comparison of survival prospects following treatment with immunotherapy and chemotherapy, OS was demonstrated to favor atezolizumab as compared to docetaxel within the METscore-Low group (HR 0.64, 95% CI 0.49–0.84, P = 0.0011; Additional file 7: Fig. S4A), whereas it was generally similar between the two arms within the METscore-High group (HR 0.98, 95% CI 0.64–1.48, P = 0.9121; Additional file 7: Fig. S4B). These results indicated that METscore enabled noninvasive identification of beneficiaries of ICI therapy, i.e., the METscore-Low patients.
Generalization of METscore in external clinical-trial and real-world cohorts
The METscore-based system was then externally validated in PD-L1-selected clinical-trial and real-world cohorts. In FIR, the use of METscore allowed stratification of patients into METscore-High and -Low groups with significant difference of OS (HR 0.57, 95% CI 0.33–0.97, P = 0.0350; Fig. 5C). In BIRCH, OS rates consistently favored METscore-Low over METscore-High (HR 0.53, 95% CI 0.40–0.71, P < 0.0001; Fig. 5). On top of that, the generalization performance of METscore was confirmed in NFyy, where patients with METscore-Low had prolonged OS than those with METscore-High (HR 0.60, 95% CI 0.39–0.92, P = 0.0181; Fig. 5E).
Considering that ICI has become the first-line treatment of advanced NSCLC [17], we evaluated the discrimination performance of METscore specifically in the first-line setting. It is noteworthy that survival prospects were significantly longer for METscore-Low patients referenced to METscore-High patients in the first-line FIR/BIRCH cohort in terms of both OS (HR 0.44, 95% CI 0.23–0.83, P = 0.0087; Fig. 6A) and PFS (HR 0.56, 95% CI 0.32–0.95, P = 0.0290; Fig. 6B). The significant survival advantage of METscore-Low over METscore-High groups were replicated in the first-line NFyy cohort according to OS (HR 0.51, 95% CI 0.29–0.90, P = 0.0182; Fig. 6C) and PFS (HR 0.45, 95% CI 0.25–0.80, P = 0.0045; Fig. 6D).