Atezolizumab versus pembrolizumab in the treatment of advanced triple-negative breast cancer: a Bayesian network meta-analysis of ITT and PD-L1 positive populations

Abstract

Background

Immune checkpoint inhibitors (ICIs) combined with chemotherapy (CT) are a standard treatment for advanced or metastatic triple-negative breast cancer (mTNBC). This study aims to evaluate the comparative efficacy and ranking of ICIs + CT regimens in treating advanced TNBC.

Methods

A Bayesian network meta-analysis of phase III randomized controlled trials (RCTs) comparing ICI + CT versus CT alone was conducted in advanced TNBC patients. Outcomes included objective response rate (ORR), disease control rate (DCR), landmark mortality probability ratios, and landmark progression probability ratios across intention-to-treat (ITT) and PD-L1-positive populations. Indirect comparison was performed using R software following PRISMA 2020 guidelines.

Results

The network analysis included four RCTs (six publications; n = 2,776). Atezolizumab (A) + CT achieved higher ORR than chemotherapy alone, especially in the PD-L1-positive population, whereas DCRs were comparable across groups. At 36 months, landmark mortality probability ratios in the ITT population were 0.68 (95% CrI 0.60–0.76) vs. 0.67 (95% CrI 0.58–0.78) for A + CT and pembrolizumab (P) + CT, respectively. In the PD-L1-positive subgroup, these ratios were 0.73 (95% CrI 0.61–0.88) vs. 0.83 (95% CrI 0.53–1.18) for A + CT and P + CT, respectively. The 36-month landmark progression probability ratios favored P + CT in the ITT cohort (0.47 vs. 0.62 for A + CT) and were comparable in the PD-L1-positive population (0.51 vs. 0.47 for A + CT). SUCRA ranking showed A + CT as having the highest probability of effectiveness for short-term outcomes in the PD-L1-positive patients. Conversely, P + CT demonstrated superior long-term survival stability in the ITT population, reinforced by restricted mean survival time (RMST) analysis. The certainty of evidence via GRADE was low to moderate.

Conclusion

Both ICI regimens provide clinical benefit, as reflected by lower landmark mortality and progression probability ratios compared with chemotherapy. A + CT demonstrates superior initial response probabilities in PD-L1-positive patients, whereas P + CT exhibits more durable long-term outcomes in the ITT population. Regimen selection should be guided by specific clinical priorities, biomarker status, and the desired temporal profile of the therapeutic effect.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-026-16207-8.

Background

Triple-negative breast cancer (TNBC) is a subtype of breast cancer with a disproportionately poor prognosis [1], and accounts for 10–25% of all breast cancer cases [2]. The five-year overall survival (OS) rate for TNBC is 61.4%, significantly lower than the 79.1% observed in hormone receptor-positive subtypes [3]. In the metastatic setting (mTNBC), the clinical feature is particularly challenging with a median survival of 8–13 months, and a five-year relative survival of approximately 11–12% [4].

The lack of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) renders TNBC unresponsive to conventional endocrine and HER2-targeted treatments. This lack of therapeutic targets often results in rapid disease progression and limited treatment options [5]. While TNBC is considered an immunologically hot tumor, which is characterized by a high tumor mutational burden (TMB) and dense immune cell infiltration, the clinical responses to single-agent immunotherapy have been largely unsatisfactory [6].

A primary challenge to effective treatment lies in the tumor microenvironment (TME), where TNBC cells exploit the PD-L1/PD-1 axis to evade immune surveillance, thereby attenuating the efficacy of checkpoint blockade. In response, immune checkpoint inhibitors (ICIs) such as pembrolizumab (anti-PD-1) and atezolizumab (anti-PD-L1) were developed. Based on the KEYNOTE-355 trial, the FDA approved pembrolizumab combined with chemotherapy as a first-line treatment for advanced TNBC patients with a PD-L1 combined positive score (CPS) ≥ 10 [7].

However, the regulatory landscape for atezolizumab has been marked by significant volatility, creating clinical uncertainty. Atezolizumab initially received FDA approval following the IMpassion130 trial, which demonstrated a significant improvement in progression-free survival (PFS) (7.4 vs. 4.8 months) in the PD-L1-positive population treated with atezolizumab plus nab-paclitaxel compared to chemotherapy alone [8]. However, in 2021, based on the findings of the IMpassion131 trial, which failed to meet its primary endpoint of PFS and changes in the treatment landscape that led the FDA to conclude that continued approval was no longer appropriate [9], as well as the results of the study based on the IMpassion132 trial, which had not succeeded in showing optimal outcomes [10].

In the absence of direct head-to-head randomized controlled trials (RCTs), comparative evidence regarding the relative efficacy of atezolizumab versus pembrolizumab regimens remains limited. Therefore, it is hypothesized that the failure to achieve consistent clinical outcomes across these trials may stem from evaluating unselected intention-to-treat (ITT) populations without adequate consideration of PD-L1 expression levels. Consequently, this study employs a Bayesian network meta-analysis to perform an indirect comparison of ICI efficacy across ITT and PD-L1-positive advanced TNBC populations, providing much-needed clarity on regimen-specific performance.

Methods

Study design and registration

This Bayesian network meta-analysis study was conducted based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. It was conducted in compliance with the PRISMA declaration standards [11]. (See Supplementary Information File S1). This study was prospectively registered through PROSPERO (www.crd.york.ac.uk/prospero/) with registration number CRD420251118179.

Database search and strategies

Studies published in the English language from March 21st, 2020, to April 21st, 2025, were comprehensively searched from the databases of PubMed (pubmed.ncbi.nlm.nih.gov), Cochrane Library (www.cochranelibrary.com), and Science Direct (www.sciencedirect.com). The search terms used were: “Immune-Checkpoint Inhibitors” OR “ICIs” OR “Atezolizumab” OR “Pembrolizumab”, combined with “Triple-Negative Breast Cancer” OR “TNBC”, “Locally-Advanced” OR “Metastatic”, and “Prognosis” OR “Overall Survival” OR “Progression-Free Survival” OR “Clinical Outcome”.

Study selection

Phase III RCTs enrolling patients with advanced or metastatic triple-negative breast cancer were included. Eligible trials had to compare an ICI plus chemotherapy versus placebo plus chemotherapy, reporting outcomes for both the intention-to-treat (ITT) and PD-L1-positive populations. Inclusion required reporting of at least one major clinical outcome: ORR, DCR, PFS, or OS. Additional requirements included documentation of adequate organ function, no prior targeted therapy for advanced disease, and completion of any previous curative treatment ≥ 12 months before enrolment. Trials were specifically excluded if they involved patients with active central nervous system metastases or mandated the use of chronic corticosteroid therapy.

Sourced publications identified from the databases were imported into the Covidence platform (www.covidence.org/) for systematic screening, which was conducted by 2 independent reviewers.

Data extraction

Following independent screening of titles/abstracts and full texts, the extracted articles were comprehensively evaluated by two independent reviewers. It was carried out using a predefined table that includes the author’s name, publication year, trial ID, country, study design, sample size, TNBC type, treatment regimen, control group, and clinical outcomes.

To prevent duplication of data from publications from the same trial, the IMpassion130 study was treated as a single parent trial, with data specifically selected based on outcome, analysis population, and maturity of follow-up. ORR and DCR data were extracted from the initial publication by Schmid et al. (2018) because these outcomes were fully reported there [8] landmark mortality probability data were entirely extracted from Emens et al. (2021) as the final results report of IMpassion130 [12], Landmark progression probability data in the ITT population were extracted from Schmid et al. (2019) [13], while landmark progression probability data in the PD-L1-positive population were extracted from Emens et al. (2021) because the final report provided more mature results for that subgroup [12]. This approach ensured that for each outcome and each analysis population, data from IMpassion130 were included only once, thus avoiding double-counting of participants in the analysis. Since this analysis indirectly compares the ITT and PD-L1-positive populations, the use of different publications from the same trial was done only to complement different outcomes or different populations, not to re-include the same samples into the same analysis.

Risk of bias assessment

The Cochrane Risk of Bias 2 (RoB2) tool was utilised to evaluate potential bias in selected studies [14]. Three independent reviewers conducted the assessments, and disagreements were resolved by discussion.

Certainty of evidence

The quality of evidence for each outcome was evaluated using the GRADE (Grading of Recommendations Assessment, Development, and Evaluation) approach [15]. Evidence quality was categorized as high, moderate, low, or very low based on risk of bias, inconsistency, indirectness, imprecision, and publication bias.

We explicitly assessed the transitivity assumption by comparing the distribution of clinical and methodological effect modifiers across the included trials. It was determined that transitivity is only partially met due to inherent differences in PD-L1 assay methodologies (SP142 vs. CPS), varying chemotherapy backbones, and disparate patient recruitment criteria (e.g., the early-relapse population in IMpassion132). Consequently, particular attention was given to the indirectness domain during GRADE rating. Certainty was evaluated separately for the ITT and PD-L1-positive populations across all specified time points (6, 12, 24, and 36 months).

Statistical analysis

Network meta-analysis model

Comparative effectiveness was performed using a random-effects Bayesian network meta-analysis in R software (v.4.4.0) utilizing the Bugsnet package. Categorical outcomes (ORR, DCR) were expressed as odds ratios (ORs), reported with 95% CrI. For landmark probability outcomes (OS, PFS), the primary effect measure was the ratio of event probabilities (death or progression) at each landmark time point, derived from the complement of the survival function (6, 12, 24, and 36 months). These estimates are derived as landmark summaries of the underlying survival curves, rather than independent Cox proportional hazard ratios, to capture the time-varying nature of the treatment effect. Values < 1 indicate reduced event risk (mortality or progression) for the ICI regimen compared with chemotherapy.

Model estimation and convergence

Parameter estimates were derived using the Markov Chain Monte Carlo (MCMC) algorithm with non-informative priors. We employed four chains with a 1,000-sample burn-in and 100,000 iterations. Convergence was confirmed by a potential scale reduction factor (Rhat) < 1.05; iterations were increased if necessary to meet this threshold. Treatment hierarchies were established via the Surface Under the Cumulative Rank (SUCRA) curve; however, given the transitivity limitations, these are presented as exploratory descriptors of relative ranking within the evidence structure rather than definitive comparative evidence.

Heterogeneity and inconsistency

Model fit was assessed using the Deviance Information Criterion (DIC). Inconsistency between direct and indirect evidence was assessed graphically by comparing the posterior mean deviations. Points aligning with the identity line indicated consistency, while significant deviations (absolute contribution > 1) were investigated for potential outliers or influential studies [16]. Clinical relevance was interpreted in relation to established minimal clinically important differences (MCID).

Sensitivity analysis

To address the dilution effect and methodological heterogeneity, a leave-one-out sensitivity analysis was performed. Specifically, the IMpassion132 study was excluded to evaluate its impact on the stability of the network, as its early-relapse population represented a significant clinical departure from the other included Phase III trials.

Restricted Mean Survival Time (RMST)

To provide a robust quantification of survival where proportional hazards assumptions may fail, RMST analysis was performed at a 24-month horizon (𝜏=24). Individual patient data (IPD) were reconstructed from published Kaplan-Meier curves through coordinate extraction using WebPlotDigitizer and the IPDfromKM package. Pooled data for atezolizumab, pembrolizumab, and placebo arms were compared using the survRM2 package. Visualizations, including RMST curves and shaded Kaplan-Meier plots, were generated via the survminer package.

Results

Study selection

A total of 6,006 articles were identified through database searches. Following the removal of 22 duplicates and applying database-specific filters, 5,300 studies were excluded, leaving 684 articles for title and abstract screening. After excluding 669 for breaking the inclusion criteria, 15 articles were assessed. Then, one study was excluded due to a lack of a full paper, and 8 studies were excluded due to inclusion of subgroup analyses, incomplete reporting, or irrelevant outcomes (e.g., pCR or monotherapy-focused trials) [17–24]. Ultimately, six studies met all inclusion criteria and were included in the study (Fig. 1).

Fig. 1.

PRISMA flowchart for the systematic literature review and meta-analysis

Study characteristics

The final analysis included four phase III RCTs (three IMpassion trials, one KEYNOTE trial) from six published studies [7–10, 12, 13] (Table 1). Where multiple publications were available for a single trial, the most recent and complete data for each outcome were prioritized. Specifically, three follow-up reports from IMpassion130, at which the initial report by Schmid et al. (2018) was added manually,⁸ were included to capture time-to-event outcomes at different follow-up milestones. Patients with unresectable locally advanced or metastatic TNBC were enrolled, with sample sizes ranging from 595 to 902 participants.

Table 1.

Study characteristics of included studies

Author, Years	Study Protocol	Country	Study Design	Sample Size	Median Age (years)	TNBC Type	Treatment	Control	PD-L1 Detection Methods	Chemotherapy Backbone	Steroid Use
Schmid et al., 2018 [8]*	IMpassion130	Multiple Countries	Phase III RCT	902	ITT: 55, Control: 56 PD-L1-pos: 53, Control: 53	Unresectable locally advanced or metastatic	Atezolizumab + paclitaxel	Placebo + Paclitaxel	VENTANA SP142	Nab-paclitaxel	> 10 mg prednisone equivalents in the 2 weeks prior to randomization
Schmid et al., 2020 [13]*	IMpassion130	Multiple Countries	Phase III RCT	902	ITT: 55, Control: 56 PD-L1-pos: 53, Control: 53	Unresectable locally advanced or metastatic	Atezolizumab + paclitaxel	Placebo + Paclitaxel	VENTANA SP142	Nab-paclitaxel	> 10 mg prednisone equivalents in the 2 weeks prior to randomization
Emens et al., 2021 [12]*	IMpassion130	Multiple Countries	Phase III RCT	451	ITT: 55, Control: 56 PD-L1-pos: 53, Control: 53	Unresectable locally advanced or metastatic	Atezolizumab + paclitaxel	Placebo + Paclitaxel	VENTANA SP142	Nab-paclitaxel	> 10 mg prednisone equivalents in the 2 weeks prior to randomization
Miles et al., 2021 [9]	IMpassion131	Multiple Countries	Phase III RCT	651	ITT: 54, Control: 53 PD-L1-pos: 55, Control: 53	Unresectable locally advanced or metastatic	Atezolizumab + paclitaxel	Placebo + Paclitaxel	SP142	Paclitaxel	Not standardized, steroid administration is only given for allergy prophylaxis
Dent et al., 2024 [10]	IMpassion132	Multiple Countries	Phase III RCT	595	ITT: 49, Control: 49 PD-L1-pos: 48, Control: 48	Unresectable locally advanced or metastatic	Atezolizumab + CT	Placebo + CT	SP142	Gemcitabine-carboplatin or capecitabine	pre-randomization systemic corticosteroid/immunosuppressive protocol was limited, no explicit steroid information was given
Cortes et al., 2022 [7]	KEYNOTE-355	Multiple Countries	Phase III RCT	847	ITT: 53, Control: 53 PD-L1-pos: 52, Control: 52 PD-L1-pos (CPS ≥ 10): 52, Control: 55	Unresectable locally advanced or metastatic	Pembrolizumab + CT	Placebo + CT	CPS	Nab-paclitaxel, paclitaxel, or gemcitabine/carboplatin	Not standardized, steroid administration is only given for allergy prophylaxis.

Abbreviation: CT  Chemotherapy, ITT  intention-to-treat population, PD-L1-pos  PD-L1 positive, RCT  Randomised controlled trial, TNBC  triple negative breast cancer, Nab-paclitaxel  nanoparticle albumin-bound paclitaxel, CPS Combine positive score

*These studies relate to the same clinical trial

Amongst the recruited population were individuals from multiple countries, and the reported median age range was 48–53 years. Three studies evaluated a combination of atezolizumab and chemotherapy compared to a group receiving a placebo and chemotherapy. Meanwhile, one study compared the treatment effect of pembrolizumab plus chemotherapy with that of chemotherapy alone.

Risk of bias assessment

No significant risk of bias was found in the included studies (Fig. 2) [7–10, 12, 13].

Fig. 2.

Assessment of risk of bias in the included studies

Overall Response Rate (ORR) and Disease Control Rate (DCR)

The network meta-analysis for ORR and DCR integrated data from six publications across four Phase III RCTs (IMpassion130, IMpassion131, IMpassion132, and KEYNOTE-355), representing a network of 2,776 patients for ORR and 2,396 patients for DCR in the Intention-to-Treat (ITT) population, and 1,650 patients for ORR and 1,296 patients for DCR in the PD-L1-positive population. The number of patients varies for each parameter because there is one study by IMpassion130 that reported several outcomes in different published studies. Certainty of evidence per the GRADE framework is summarized in Table 2.

Table 2.

GRADE assessment on objective response rate and disease control rate for the treatment effect of Atezolizumab-Chemotherapy vs. Pembrolizumab-Chemotherapy for advanced TNBC

Summary of findings:
Clinical Outcomes of Atezolizumab-Chemotherapy and Pembrolizumab-Chemotherapy for Triple-Negative Breast Cancer
Outcome	Population	№ of participants (№ of studies)	Anticipated relative effects (95% CrI)	Certainty	What happens
			Odds Ratio (OR)
ORR	ITT	2776 (4)	A + CT vs. PL + CT: 1.33 (0.95 to 1.80)	⨁⨁⨁◯ Moderate	A + CT probably increases ORR compared to CT, while P + CT may result in little to no difference in ITT patients.
			P + CT vs. PL + CT: 1.17 (0.70 to 1.98)	⨁⨁◯◯ Low
	PD-L1+	1650 (4)	A + CT vs. PL + CT: 1.63 (1.05 to 2.51)	⨁⨁⨁◯ Moderate	A + CT probably increases ORR compared to CT, while P + CT may result in little to no difference in PD-L1 + patients.
			P + CT vs. PL + CT: 1.29 (0.64 to 2.59)	⨁⨁◯◯ Low
DCR	ITT	2396 (4)	A + CT vs. PL + CT: 1.17 (0.66 to 2.02)	⨁⨁◯◯ Low	Both regimens demonstrated comparable efficacy with no statistically significant difference in DCR.
			P + CT vs. PL + CT: 0.81 (0.37 to 1.76)	⨁⨁◯◯ Low
	PD-L1+	1296 (4)	A + CT vs. PL + CT: 1.22 (0.63 to 2.21)	⨁⨁◯◯ Low	Both regimens demonstrated comparable efficacy with no statistically significant difference in DCR.
			P + CT vs. PL + CT: 0.88 (0.39 to 1.97)	⨁⨁◯◯ Low

GRADE Working Group grades of evidence. High certainty: We are very confident that the true effect lies close to the estimated effect. Moderate certainty: We are moderately confident in the effect estimate: the true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different. Low certainty: our confidence in the effect estimate is limited; the true effect may be substantially different from the estimate of the effect. Very low certainty: we have very little confidence in the effect estimate; the true effect is likely to be substantially different from the estimated effect

Abbreviations: A+CT Atezolizumab-Chemotherapy, CrI Credible interval, CT Chemotherapy, DCR Disease Control Rate, ITT Intention-To-Treat, ORR Overall Response Rate, OR Odds Ratio, P+CT Pembrolizumab-Chemotherapy, PL+CT Placebo-Chemotherapy

In the ITT population, both A + CT and P + CT demonstrated a higher ORR compared to chemotherapy alone. The A + CT group showed a stronger pooled effect (OR 1.33; 95% CrI 0.95–1.80) and achieved the highest SUCRA score (Fig. 3; Tables 4 and 5). This benefit was more pronounced in the PD-L1-positive population, where A + CT reached the most substantial numerical effect (OR 1.63; 95% CrI 1.05–2.51) and the highest SUCRA ranking, though the comparative efficacy between the two ICIs did not reach statistical significance.

Fig. 3.

Restricted mean survival time for OS in the ITT population. A Pooled RMST by treatment group. B Kaplan-Meier analysis by treatment group

Table 4.

Summary of findings in ITT AND PD-L1 + population

Table 5.

Summary of mortality and progression probability ratio findings in ITT AND PD-L1 + population after excluding the impassion132 trial

For the DCR, numerical improvement in disease stabilization was observed across both populations. A + CT was associated with higher DCR odds than both P + CT and control in the ITT and PD-L1-positive cohort (Table 2). SUCRA analysis ranked A + CT highest for DCR in both groups (Tables 4 and 5), suggesting a relatively better probability of achieving disease control compared to P + CT. The consistent trends and higher rank probabilities in the PD-L1-positive network support the role of PD-L1 expression as a predictive biomarker for ICI-mediated cytostatic benefit (Table 4).

Mortality and progression probability outcomes

The analysis of mortality and progression probability outcomes showed a complex interplay between the regimen, population, and follow-up time. The certainty of evidence with a detailed GRADE assessment is provided in Table 3.

Table 3.

GRADE assessment on mortality and progression probability ratios (landmark) for the treatment effect of Atezolizumab-Chemotherapy vs. Pembrolizumab-Chemotherapy for advanced TNBC

Summary of findings:
Efficacy of Atezolizumab-Chemotherapy and Pembrolizumab-Chemotherapy for Advanced Triple-Negative Breast Cancer
Outcome	Population	№ of participants (№ of studies)	Timepoint	Anticipated relative effects (95% CrI)	Certainty	What happens
				Probability Ratio (Landmark)
Mortality Probability Ratio	ITT	2780 (4)	6-months	A + CT vs. PL + CT: 0.84 (0.69 to 1.03)	⨁⨁⨁◯ Moderate	Both regimens demonstrated comparable efficacy with no statistically significant difference in mortality probability.
				P + CT vs. PL + CT: 0.91 (0.65 to 1.30)	⨁⨁◯◯ Low
			12-months	A + CT vs. PL + CT: 0.84 (0.73 to 0.97)	⨁⨁⨁◯ Moderate	Both regimens demonstrated comparable efficacy with no statistically significant difference in mortality probability.
				P + CT vs. PL + CT: 0.84 (0.67 to 1.05)	⨁⨁◯◯ Low
			24-months	A + CT vs. PL + CT: 0.81 (0.72 to 0.90)	⨁⨁⨁◯ Moderate	A + CT and P + CT probably reduce the mortality probability.
				P + CT vs. PL + CT: 0.76 (0.64 to 0.90)	⨁⨁⨁◯ Moderate
			36-months	A + CT vs. PL + CT: 0.68 (0.60 to 0.76)	⨁⨁⨁◯ Moderate	A + CT and P + CT probably reduce the mortality probability.
				P + CT vs. PL + CT: 0.67 (0.58 to 0.78)	⨁⨁⨁◯ Moderate
	PD-L1+	1651 (4)	6-months	A + CT vs. PL + CT: 0.84 (0.63 to 1.11)	⨁⨁◯◯ Low	Both regimens demonstrated comparable efficacy with no statistically significant difference in mortality probability.
				P + CT vs. PL + CT: 1.15 (0.73 to 1.85)	⨁⨁◯◯ Low
			12-months	A + CT vs. PL + CT: 0.88 (0.72 to 1.08)	⨁⨁◯◯ Low	Both regimens demonstrated comparable efficacy with no statistically significant difference in mortality probability.
				P + CT vs. PL + CT: 0.96 (0.71 to 1.29)	⨁⨁◯◯ Low
			24-months	A + CT vs. PL + CT: 0.84 (0.71 to 0.99)	⨁⨁⨁◯ Moderate	A + CT and P + CT probably reduce the mortality probability.
				P + CT vs. PL + CT: 0.78 (0.63 to 0.98)	⨁⨁⨁◯ Moderate
			36-months	A + CT vs. PL + CT: 0.73 (0.61 to 0.88)	⨁⨁⨁◯ Moderate	A + CT probably reduces the mortality probability, while P + CT may result in little to no difference.
				P + CT vs. PL + CT: 0.83 (0.53 to 1.18)	⨁⨁◯◯ Low
Progression Probability Ratio	ITT	2568 (4)	6-months	A + CT vs. PL + CT: 0.88 (0.77 to 1.01)	⨁⨁⨁◯ Moderate	A + CT and P + CT probably reduce the progression probability.
				P + CT vs. PL + CT: 0.80 (0.66 to 0.98)	⨁⨁⨁◯ Moderate
			12-months	A + CT vs. PL + CT: 0.82 (0.72 to 0.92)	⨁⨁⨁◯ Moderate	A + CT and P + CT probably reduce the progression probability.
				P + CT vs. PL + CT: 0.77 (0.65 to 0.92)	⨁⨁⨁◯ Moderate
			24-months	A + CT vs. PL + CT: 0.51 (0.45 to 0.57)	⨁⨁⨁◯ Moderate	A + CT and P + CT probably reduce the progression probability.
				P + CT vs. PL + CT: 0.54 (0.49 to 0.61)	⨁⨁⨁◯ Moderate
			36-months	A + CT vs. PL + CT: 0.62 (0.53 to 0.74)	⨁⨁⨁◯ Moderate	A + CT and P + CT probably reduce the progression probability.
				P + CT vs. PL + CT: 0.47 (0.42 to 0.54)	⨁⨁⨁◯ Moderate
	PD-L1+	1561 (4)	6-months	A + CT vs. PL + CT: 0.70 (0.57 to 0.87)	⨁⨁⨁◯ Moderate	A + CT probably reduces the progression probability, even though P + CT may result in little to no difference.
				P + CT vs. PL + CT: 0.78 (0.57 to 1.06)	⨁⨁◯◯ Low
			12-months	A + CT vs. PL + CT: 0.76 (0.64 to 0.89)	⨁⨁⨁◯ Moderate	A + CT and P + CT probably reduce the progression probability.
				P + CT vs. PL + CT: 0.74 (0.60 to 0.91)	⨁⨁⨁◯ Moderate
			24-months	A + CT vs. PL + CT: 0.67 (0.55 to 0.80)	⨁⨁⨁◯ Moderate	A + CT and P + CT probably reduce the progression probability.
				P + CT vs. PL + CT: 0.69 (0.57 to 0.84)	⨁⨁⨁◯ Moderate
			36-months	A + CT vs. PL + CT: 0.47 (0.39 to 0.56)	⨁⨁⨁◯ Moderate	A + CT and P + CT probably reduce the progression probability.
				P + CT vs. PL + CT: 0.51 (0.41 to 0.62)	⨁⨁⨁◯ Moderate

GRADE Working Group grades of evidence. High certainty: We are very confident that the true effect lies close to the estimated effect. Moderate certainty: we are moderately confident in the effect estimate: the true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different. Low certainty: our confidence in the effect estimate is limited; the true effect may be substantially different from the estimate of the effect.Very low certainty: we have very little confidence in the effect estimate; the true effect is likely to be substantially different from the estimated effect

Abbreviations: CrI Credible interval, CT Chemotherapy, A+CT Atezolizumab-Chemotherapy, ITT Intention-To-Treat, P+CT Pembrolizumab-Chemotherapy, PD-L1+PD-L1-positive, PL+CT Placebo-Chemotherapy

Bolded values indicate statistical significance (the 95% CrI does not cross the null value of 1.0)

In the initial 6 months follow-up, the landmark mortality probability ratios were comparable between A + CT and P + CT in the ITT (0.84 (95% CrI 0.69–1.03) vs. 0.91 (95% CrI 0.65–1.30), respectively), with low-to-moderate certainty. In the PD-L1-positive populations at 6 months, A + CT showed landmark mortality probability ratios of 0.84 (95% CrI 0.63–1.11), while P + CT was estimated at 1.15 (95% CrI 0.73–1.85). During extended follow-up, both regimens demonstrated moderate certainty in reducing the probability of death at 12–36 months across the ITT population. By 36 months, landmark mortality probability ratios in the ITT group reached 0.68 for A + CT and 0.67 for P + CT. Notably, in the PD-L1-positive population at 36 months, A + CT was associated with a reduction in the probability of death (0.73; 95% CrI 0.61–0.88; moderate certainty), while the P + CT estimate showed a wider interval crossing the null effect (0.83; 95% CrI 0.53–1.18; low certainty), achieving the highest SUCRA rank.

At 6 months, P + CT achieved a moderate-certainty landmark progression probability ratio of 0.80 in the ITT population, while A + CT was 0.88 (Fig. 3). In the PD-L1-positive subgroup at 6 months, A + CT demonstrated a significant reduction in the probability of progression (0.70; 95% CrI 0.57–0.87; moderate certainty), whereas the P + CT estimate (0.78; 95% CrI 0.57–1.06) was of low certainty (Tables 4 and 5). By the 36-month landmark, P + CT demonstrated a numerical lead in the ITT population (0.47 vs. 0.62 for A + CT), both with moderate certainty. In the PD-L1-positive population at 36 months, both A + CT (0.47) and P + CT (0.51) showed similar magnitudes of effect, both supported by moderate certainty. SUCRA analysis was relatively consistent and favored P + CT at 12 months in both groups and at 24 and 36 months in the PD-L1-positive population.

In the ITT population, A + CT achieved the highest 24-month RMST for OS (17.02 months), followed by P + CT (16.26 months) and placebo (15.98 months). Kaplan-Meier curves remained closely aligned, though the A + CT trajectory remained descriptively higher throughout the 24-month observation period (Fig. 3). In the PD-L1-positive subgroup, the 24-month OS RMST for A + CT was 16.69 months, compared to 16.43 months for P + CT and 15.67 months for placebo. Visual analysis demonstrated a clearer separation between the active therapy and placebo groups, while atezolizumab remained descriptively superior (Fig. 4).

Fig. 4.

Restricted mean survival time for OS in the PD-L1 positive population. A Pooled RMST by treatment group. B Kaplan-Meier analysis by treatment group

For PFS, P + CT demonstrated the longest 24-month RMST in the ITT population (10.07 months), compared to A + CT (8.25 months) and placebo (7.57 months). Kaplan-Meier curves showed a more significant separation than the OS analysis, primarily because P + CT consistently outperformed A + CT and placebo after the first few months of follow-up (Fig. 5). This pattern was consistent in the PD-L1-positive subgroup, where P + CT achieved a 24-month PFS RMST of 10.42 months (vs. 9.18 months for A + CT and vs. 7.37 months for placebo), maintaining the top-ranked position throughout the 24-month horizon (Fig. 6).

Fig. 5.

Restricted mean survival time for PFS in the ITT population. A Pooled RMST by treatment group. B Kaplan-Meier analysis by treatment group

Fig. 6.

Restricted mean survival time for PFS in the PD-L1 positive population. A Pooled RMST by treatment group. B Kaplan-Meier analysis by treatment group

Exclusion of the IMpassion132 trial was conducted to account for the potential dilution effect of its early-relapse population. This analysis did not result in substantial changes to the pooled network estimates (Table 5), suggesting the primary survival findings are robust across the included trial data.

GRADE assessment

Evidence for ORR was graded as moderate certainty. It showed a consistent benefit for A + CT in both the ITT (n = 2,776) and PD-L1-positive patients (n = 1,650). One level downgrade for indirectness was due to variation in chemotherapy backbones and PD-L1 assay cut-offs across studies.

Evidence for DCR was assessed as low certainty. Although A + CT trended favorably, certainty was limited by indirectness (same as above) and imprecision. The wide CrI for comparisons involving A + CT and P + CT often crossed the null effect (OR = 1), limiting confidence in the magnitude of benefit.

All landmark mortality and progression probability ratios were graded as low-to-moderate certainty. Downgrading for inconsistency was applied because treatment rankings (SUCRA) were unstable over time, with benefits at 24 months failing to maintain consistency by 36 months. Downgrading for imprecision was required because many landmark mortality and progression probability ratios clustered near 1, with CrIs ranging from modest benefit to potential harm. Additional downgrading for indirectness was noted as the evidence was predominantly confined to metastatic TNBC treated with taxane-based regimens, and long-term data reflected highly selected survivor cohorts.

Bayesian assessment suggested no evidence of significant publication bias. No formal downgrading was applied, though small-study effects cannot be entirely excluded. Detailed assessments are provided in Tables 2 and 3.

Discussion

This Bayesian network meta-analysis evaluated the differential magnitude and durability of clinical benefit between pembrolizumab plus chemotherapy (P + CT) and atezolizumab plus chemotherapy (A + CT) in advanced TNBC. By utilizing time-dependent outcomes and RMST analysis, this study reveals a longitudinal evaluation of immune checkpoint inhibitor (ICI) performance that extends beyond the static conclusions of individual trial reports.

Our findings demonstrate that ICIs provide clinically meaningful, albeit variable, benefits. In the ITT population, P + CT achieved superior longitudinal stability in landmark progression probability outcomes compared to A + CT and CT alone across nearly all intervals, except for 24 months. It was reinforced by RMST analysis, which identified P + CT as the regimen with the longest PFS duration. Similarly, A + CT exhibited a landmark progression probability outcome nearly comparable to P + CT, and showed better landmark progression probability outcomes than P + CT at 24 months. Furthermore, both regimens generally showed consistent results in reducing the probability of progression across time points. While RMST indicated a marginal OS advantage for A + CT in the ITT population, the overall data suggest that pembrolizumab maintains a more consistent PFS advantage across heterogeneous TNBC cohorts.

Among the PD-L1-positive population, both ICI regimens showed a consistent superior landmark progression probability outcome advantage over chemotherapy, reinforcing PD-L1’s role as a validated enrichment biomarker. Interestingly, while A + CT demonstrated a robust initial benefit at 6 months, there is a pattern in which A + CT shows a decrease in landmark progression probability outcome at 12 months, slightly worse than P + CT, then it improves again until it peaks at 36 months. Furthermore, a more consistent pattern is shown by P + CT. However, at the beginning of the landmark progression probability outcome observation at 6 months, it shows a worse landmark progression probability outcome than A + CT, but the increase in landmark progression probability outcome for P + CT occurs consistently without any fluctuation pattern of decreasing efficacy as observed in A + CT. A similar pattern was also observed in the landmark mortality probability outcomes for both A + CT and P + CT in the ITT and PD-L1-positive populations. This aligns with findings from KEYNOTE-355, where pembrolizumab benefit increased at higher CPS thresholds, particularly at CPS ≥ 10, and IMpassion130, which demonstrated the efficacy of atezolizumab in the PD-L1-positive cohort [7, 8].

The observed variability between A + CT and P + CT likely reflects trial-specific heterogeneity rather than solely divergent drug pharmacology. Key drivers include biomarker methodology, chemotherapy backbones and steroids, and patient selection. The SP142 assay (IMpassion) is restricted to immune cells, whereas CPS (KEYNOTE-355) integrates tumor and immune cells. Their populations are not directly comparable, and the consistency of pembrolizumab may be tied to the broader CPS capture [7, 18, 19]. The use of nab-paclitaxel (IMpassion130) avoids the corticosteroid premedication required for solvent-based paclitaxel (IMpassion131). Steroids may inhibit T-cell activation, potentially explaining the lack of benefit in the latter. This likely explains the lack of benefit in IMpassion131 with additional support from real-world data [9, 13, 25, 26]. Furthermore, IMpassion132 recruited patients with early relapse (< 12 months), a biologically aggressive group often refractory to ICIs, which likely diluted the pooled long-term survival estimates for atezolizumab [10, 27]. These survival patterns suggest a time-dependent benefit of immunotherapy, with less favorable initial outcomes but gradual improvement over longer follow-up, with atezolizumab appearing more dependent on PD-L1 expression and chemotherapy synergy, while pembrolizumab appears more consistent [28]. The decreased long-term benefit in both regimens is likely related to the evolution of TNBC towards a more immunosuppressive microenvironment, including the influence of myeloid suppression, pro-tumor secretome, and inhibitory mechanisms via liver, lung, and bone metastases, which is the majority of patients in each included trial [29–37].

Although the exclusion of the IMpassion132 study in the sensitivity analysis reduced statistical noise, it did not eliminate the fundamental differences in PD-L1 assay type and baseline chemotherapy regimens across trials. Consequently, the transitivity assumption was only partially satisfied, and our indirect comparison rankings should be interpreted as exploratory descriptors of relative performance within the existing evidence structure. Furthermore, our results represent the landmark mortality and progression probability ratios, not the results of separate Cox modeling for each time point. This approach provides a more dynamic picture of the change in treatment effect over follow-up, such as the delayed clinical benefit characteristic of immunotherapy, which single-point Cox modeling may obscure.

A major strength of this study is the use of Bayesian modelling and RMST to quantify temporal effects where proportional hazards assumptions fail. However, limitations must be acknowledged. The small number of studies and the lack of head-to-head comparisons necessitate reliance on indirect comparisons. Heterogeneity in PD-L1 assays and chemotherapy partners, and steroid use, introduces noise that limits causal inference. While these factors suggest caution, the consistency of the trends across multiple datasets and IPD reconstruction strengthens the overall interpretability of the findings.

Conclusion

The integration of ICIs in advanced TNBC treatment provides clear clinical benefits, particularly for PD-L1-positive patients. P + CT demonstrated more consistent performance across ITT and long-term landmark progression probability ratio across time-point observation, while A + CT showed a pronounced but more time-sensitive signal in PD-L1-positive subgroups. RMST analysis confirmed a longer PFS for pembrolizumab, whereas the OS advantage for atezolizumab appeared more inconsistent. These findings underscore that ICI efficacy is highly dependent on the trial context, chemotherapy backbones, and biomarker selection. Future research should prioritize standardized direct comparisons and more precise biomarkers to optimize treatment selection in advanced TNBC.

Authors’ contributions

I.G.W.W.W., M.S.T., S.K., and N.H. managed conceptualization, supervision, and project administration. C.T.M. managed formal analysis. I.G.W.W.W., I.G.K.A.S., N.P.K.M., P.M.W.S.P., K.M.K.T., A.B.N.K., K.I.P.A., and I.K.R.P.P. contributed to data curation, investigation, methodology, and writing - original draft. M.S.T., S.K., and N.H. aided in conceptualization and contributed to writing - original draft. I.G.W.W.W., M.S.T., S.K., and N.H. participated in methodology and writing - review & editing. I.G.W.W.W., M.S.T., and C.T.M. were involved in data curation, investigation, and visualization. N.H. and M.S.T. provided critical writing – review & editing, supervision, and secured funding acquisition.

Funding

This project was funded by the National Research Council of Thailand (NRCT): High Potential Research Team Grant Program (N42A680423) and Ratchadapisek Somphot Matching Fund, Faculty of Medicine, Chulalongkorn University (RA-MF-01/69). The academic endeavors of the Thailand Hub of Talents in Cancer Immunotherapy (TTCI) receive support from the National Research Council of Thailand.

Data availability

All data analysed during this study are included in this published article and its supplementary information files.

Declarations

Ethics approval and consent to participate

Ethical approval was not required for this study as it is a meta-analysis of previously published, publicly available data and does not involve the collection of primary data from human participants. However, several amendments to the PROSPERO protocol registration have been made, following the necessary changes and adjustments to the protocol, such as the addition of RMST analysis, changes to the title, the addition of GRADE assessment analysis and other adjustments.

Competing interests

The authors declare no competing interests.