Differences in the risk of immune-related pneumonitis between PD-1 and PD-L1 inhibitors: a meta-analysis according to the new mirror-principle and PRISMA guidelines

Purpose To compare the risk of immune-associated pneumonitis between PD-1 and PD-L1 inhibitors, the meta-analysis was designed. Method The difference in risk of immune-associated pneumonitis between PD-1 and PD-L1 inhibitors was assessed by two different meta-analysis methods, the Mirror-pairing and the PRISMA guidelines. Results A total of eighty-eight reports were used for meta-analysis, while thirty-two studies were used for the Mirror-pairing. Both PD-1 and PD-L1 inhibitors (used alone or combined with chemotherapy) increased the risk of developing immune-related pneumonitis (P < 0.00001; P < 0.00001). Based on indirect analyses results (subgroup analyses), the risk of PD-L1-induced pneumonitis was weaker than that of PD-1 inhibitors when the control group was chemotherapy (OR = 3.33 vs. 5.43) or placebo (OR = 2.53 vs. 3.19), while no obvious significant differences were found (P = 0.17; P = 0.53). For the Mirror-pairing-based meta-analysis, the risk of PD-1-induced pneumonitis was significantly higher than that of PD-L1 inhibitors (OR = 1.46, 95%CI [1.08, 1.98], I2 = 0%, Z = 2.47 (P = 0.01)). However, this difference was not significant, when they were combined with chemotherapy (OR = 1.05, 95%CI [0.68, 1.60], I2 = 38%, Z = 0.21 (P = 0.84)). Conclusion Both PD-1 and PD-L1 inhibitors increased the risk of immune-related pneumonitis, while the risk of PD-1-induced pneumonitis was significantly higher than that of PD-L1 inhibitors. Supplementary Information The online version contains supplementary material available at 10.1007/s00262-024-03736-z.


Introduction
Many clinical trials have confirmed that programmed death-1 (PD-1) or Programmed cell death ligand 1 (PD-L1) inhibitors have excellent clinical efficacy for malignant tumors .Due to their unique immune mechanism, many immune-related side effects have been reported as part of clinical trial results .Of immunotoxic reactions, pneumonitis was mentioned and evaluated by clinicians for lung cancer patients [1-11, 13, 15, 16, 36-49, 67-71].Due to the increasing diversity of drug combinations based on PD-1 or PD-L1 inhibitors, assessing the risk factors for pneumonitis have become much more difficult.There were no clinical trials involving a direct comparison between PD-1 and PD-L1 , which further increased the difficulty of directly comparing the differences in toxicity reactions between PD-1 and PD-L1.However, the report of the Mirror-pairing meta-analysis has made it possible for us to solve this dilemma [88,89].
The Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) reporting guideline would also be followed [90].The results of subgroup analyses were used to test whether the analysis results of the Mirror-pairing were consistent with the PRISMA meta-analysis results.In this study, the differences in the incidence risk of pneumonitis between PD-1 and PD-L1 inhibitors were evaluated by the above two analysis methods (PRISMA and Mirrorpairing).Furthermore, the applicability and reliability of the Mirror-pairing analysis method were further validated [88][89][90].

Method
The classic PRISMA analysis method was followed and prioritized for the subsequent analyses [90].

Search strategy and screening
The searching process for relevant literature in PubMed was carried out according to the PICOS (participants, interventions, comparisons, outcomes, and study design) references [90].The searching keywords were not only limited to PD-1 or PD-L1, but also included specific product names and common names of related drugs.All clinical trials without a control group, meaning the single-arm clinical trial, would be excluded first.Randomized and controlled Phase III clinical trials would be prioritized, while other randomized controlled trials were considered as alternatives.
The time frame for all literature was just limited to the past ten years (August 2, 2013-August 2, 2023).The literature searching was completed by four participating authors, and the searching results would be checked by each other.In case of duplicated clinical trials, only one containing the most complete data could be used for the final analysis.

Mirror principle pairing
To increase the similarity and minimize heterogeneity and inconsistency between groups [88,89], the Mirror-pairing criteria were listed as follows: (1) Tumor type: Due to the significant differences among different tumor types, this is the primary criteria for the Mirror-pairing; ( 2 The two paired groups have the same order of magnitude; (7) All clinical trials data could only be used once for the best Mirror-pairing; (8) Results choice: The analyses result involving single drug regimens will be prioritized, while the analyses results of combination therapy regimens are just considered as reference; (9)Other: While the above factors have been confirmed by subgroup analysis to be not the factor causing differences in subgroup analysis results, this factor can be moderately adjusted during pairing.

Evaluation of study quality and publication bias
Egger's test was used to test the symmetry of funnel plots [91,92], while funnel plots and Harbor's test were used for publication bias evaluation [92,93].The Newcastle-Ottawa scale (NOS), recommended by Cochrane Collaboration, was used for quality assessments [94,95].The assessments contents were listed as follows: random sequence generation (selection bias), allocation concealment (selection bias), blinding of participants and personnel (performance bias), blinding of outcome assessment (detection bias), incomplete outcome data (attrition bias), selective reporting (reporting bias) and others.The evaluation method was to verify the original data of the reported clinical trials.P < 0.05 was considered to be statistically significant.

Exposure of interest, assessment of heterogeneity, and statistical analyses
The basic characteristics of clinical trials that met the inclusion criteria were collected and summarized in a separate table.This study focused on the incidence risk of pneumonitis in all grades.
Heterogeneity was assessed by Cochrane's Q statistic test [94,95], including the Mantel-Haenszel method and I 2 values suggested by Higgins and colleagues [90,95].According to the different I 2 values, heterogeneity was divided into three different levels: low (I 2 < 25%), medium (I 2 = 25-50%), and high (I 2 > 50%) [90,95].The software Review Manager 5.3 was used for all the following analyses.Due to the inevitable existence of intergroup heterogeneity in the real world, random effects (RE) models were used for calculating odds ratio (OR) and 95% confidence interval (CI) [96].The fixed effects (FE) model would just be used for funnel plot evaluations.All P values were calculated by two sides.P < 0.05 was deemed to be of statistical significance.Subgroup analysis was conducted based on PD-1 or PD-L1 types.When obvious heterogeneity was discovered, more detailed subgroup analyses would be conducted based on the specific situation.If heterogeneity was considered to be mainly caused by the data itself, further processing of the data would not be carried out, and the original data analysis results would still be adopted.

Discussion
With the increasing use of PD-1/PD-L1 inhibitors in clinical practice, the complex and diverse forms of immune-related toxic side effects are increasingly reported and valued by clinical doctors .Pneumonitis, as an important clinical event of pulmonary toxicities, requires rapid identification and management.Once suspected, the scope of differential diagnosis between infectious and vegetative processes might make the physician's diagnostic process challenging [97].A comprehensive assessment of the incidence risk of immune-related pneumonitis would have important guiding significance for physicians.However, due to the lack of clinical trials comparing PD-1 and PD-L1 head to head, it was difficult to determine the differences in risk of pneumonitis occurrence between the two.To address this dilemma, this study was designed [88,89].
A literature searching was conducted according to PRISMA guidelines and PICOS principles [90], and a total of 77 clinical trial data were collected (Fig. 1; STable 1) .Seventy-seven clinical trials were taken into account for a more comprehensive and detailed analysis by grouping in more ways, which increased the possibility of obtaining more Mirror pairings and reduced the possibility of bias due to insufficient data [88,89].We carefully reviewed the data of all enrolled clinical trials and conducted a comprehensive systematic evaluation of random sequence generation (selection bias), allocation consideration (selection bias), blinding of participants and personnel (performance bias), blinding of outcome assessment (detection bias), incomplete outcome data (attrition bias), selective reporting (reporting bias), and others.After the comprehensive evaluation, data with any kind of high risk biases would be excluded, and only the high-quality and complete clinical trial data were retained, ensuring the reliability and authenticity of our analyses results (SFigure 1; STable 1).The previous Mirror-pairing principles had been improved [88.89] , which would make the pairing much more accurate.After a detailed analysis of clinical trials using the Mirror-pairing principle, 16 Mirror pairings were obtained, which was the largest number of PD-1/PD-L1 related Mirror pairings first reported so far (Table 1; SFigure 3) [1, 2, 4, 6-8, 10, 12, 14, 15, 17, 19, 21-24, 36-38, 42, 43, 45, 48, 49, 51, 53, 57, 63, 71, 78, 79].These further strengthened the innovation of our research.
Due to the inevitable existence of intergroup heterogeneity in the real world, RE models were used for OR and 95% CI calculations [96].Although no highly heterogeneous results were found, we conducted sufficient subgroup analyses and speculated on the source of the corresponding heterogeneity (SFigure 4, SFigure 5, SFigure 6, and SFigure 7) .There were no data found that affected the analysis results.Furthermore, no significant bias was found through the corresponding funnel plot (SFigure 2 and SFigure 3), which confirmed the authenticity and reliability of the above analysis results.
Based on the subgroup analysis results (Fig. S3A, SFigure 4, SFigure 5, and SFigure 6), we found that the risk of pneumonitis in UC patients receiving PD-1 inhibitors was the highest among all tumor types.This meant that special attention should be paid to the risk of immune-related pneumonitis for PD-1 inhibitor use in UC patients.
By comparing the subgroup analysis results of the PRISMA meta-analysis with the results of the Mirror-pairing analysis, we found that the risk trend of the analysis results was basically consistent, while the analysis results of the Mirror-pairing seemed to be much more sensitive (Figs.2A,2C; S3A).It indicated that when mild differences in subgroup analysis was found, the Mirror-pairing analysis could be conducted to clarify the significance of these differences.Furthermore, this would be beneficial for clinicians to determine the choice of drugs (PD-1 or PD-L1) based on the degree of toxicities, as well as whether PD-1 was needed to be replaced by PD-L1.

Conclusions
Both PD-1 and PD-L1 inhibitors increased the risk of immune-related pneumonitis, while the risk of PD-1-induced pneumonitis was significantly higher than that of PD-L1 inhibitors.

The limitations of the study
The Mirror-pairing analysis is an indirect paired comparison of existing clinical trials while minimizing heterogeneity.Its reliability still needs to be validated with more head-to-head clinical trial data in the real world.
) Pathological type: Tumors occurring in the same organ need to be distinguished based on specific pathological types.(3) Treatment regimen: In the combination treatment regimens including PD-1 or PD-L1 inhibitors, it is necessary to keep consistent in drug composition of the Mirror-pairing groups; (4) Treatment line; (5) Phase stage; (6) Number of participants:

Table 1
Basic characteristics of Mirror-pairing clinical trials