Forecasting optimal treatments in relapsed/refractory mature T- and NK-cell lymphomas: a global PETAL Consortium study

Abstract

There is no standard-of-care in relapsed/refractory T-cell/NK-cell lymphomas. Patients often cycle through cytotoxic-chemotherapy (CC), epigenetic-modifiers (EM), or small-molecule-inhibitors (SMI) empirically. Ideal therapy at each line remains unknown. We conducted a retrospective, multiple-intervention, ‘target-trial’ using the PETAL global cohort. Patients received frontline CC, then second- and third-line (2L and 3L) with either CC again, EM, or SMI (12 possible treatment scenarios). Overall survival (OS; 2L- or 3L- to-death) was compared across treatment sequences using Cox, reinforcement learning, and synthetic intervention methods adjusting for age, histology, primary refractory disease, PIT score, response to 2L, and receipt of 2L transplant consolidation. Five-hundred-and-forty received 2L (EM=101, SMI=45, CC=394), and 290 received 3L (EM=65, SMI=44, CC=181). 2L SMI then 3L EM improved OS (aHR: 0.29, 95%CI: 0.11–0.74; p=0.010) versus 2L-3L CC-CC, and consistently across most other sequential strategies. In 2L stability analyses, benefit was notable with 2L SMI in AITL (verus CC: aHR: 0.23, 95%CI: 0.10–0.4; p<0.001; versus EM: aHR: 0.32, 95%CI: 0.12–0.82; p=0.020), and both SMI and EM in PIT-stratified high-risk groups (SMI: aHR: 0.40, 95%CI: 0.21–0.76; p=0.005; EM: aHR: 0.60, 95%CI: 0.39–0.92; p=0.020) versus 2L CC. Results were consistent across all other independent stability and causal inference analyses providing a treatment selection framework.

GRAPHICAL ABSTRACT

In relapsed/refractory T-cell/NK-cell lymphomas, patients often cycle through cytotoxic-chemotherapy (CC), epigenetic-modifiers (EM), or small-molecule-inhibitors (SMI) empirically. Ideal therapy at each line remains unknown. In this retrospective, multiple-intervention, ‘target-trial’ using the PETAL global cohort, these drug classes were compared sequentially across 2^nd followed by 3^rd line (12 possible treatment scenarios) using classical and machine learning causal inference methods. At second-line, benefit was especially evident in AITL and high-risk (PIT score ≥2) populations. Small molecule inhibitors followed by epigenetic modifiers improved survival compared to most other strategies.

INTRODUCTION

Several cohort-based studies have demonstrated that the majority of patients with relapsed/refractory (R/R) mature T and NK-cell Lymphomas (TNKL) die within 2 years.^1–6 While multiagent cytotoxic chemotherapy (CC) still remains the most commonly used second-line (2L) treatment, many rationally designed therapies have been used within the last decade, such as epigenetic modifiers (EM; e.g. histone deacetylase inhibitors [HDACi] or DNA methyltransferase inhibitors [DNMTi]) and small molecule inhibitors (SMI) that suppress oncogenic pathways (e.g PI3K, JAK/STAT, etc.).^7–13 Despite this evolution, there is still no universally accepted standard-of-care for these patients and regimens are chosen empirically in no specific order. This partly due to significant challenges with implementing traditional prospective controlled trials in R/R TNKL given their rarity, heterogeneity, and aggressive presentations at relapse. Most patients will cycle through therapies based on disease kinetics, drug availability, and physician/patient preference rather than a biological rationale.

Emerging data suggest that exposure to EM can sensitize tumors to subsequent therapies.¹⁴ Due to their pleiotropic properties, clinically relevant combinations of EM with immunotherapy, immunomodulatory drugs, and SMI have been investigated in TNKL.^11,15–17 Thus, there is a desperate need for data to inform optimal treatment at each line in TNKL, and in cancers broadly.

Given the sparse nature, diverse disease biology across subtypes of TNKL, and real-world challenges with clinical trial enrollment, cohort-based observational studies could be leveraged to elucidate treatment effects at each line to facilitate clinical decisions. In this study, we interrogated the overall survival (OS) impact of commonly used therapeutic regimens in second-line (2L) and third-line (3L). We also aimed to identify distinct subgroups with OS advantage to personalize treatment. We deployed three independent causal inference and machine learning models to prioritize treatment selection based on readily available and clinically relevant patient and disease characteristics. Thus, we provide a potential framework for maximizing efficacy of therapy in 2L and 3L in the largest global cohort of patients with R/R TNKL.

METHODS

Study Design

We conducted a retrospective, multi-intervention, ‘target-trial’^18,19 cohort study using the updated clinical information from the PETAL (PEripheral T-cell lymphoma) Consortium’s global cohort (Supplemental Table S1). Patients were included if they received first-line (1L) treatment with multiagent cytotoxic chemotherapy (CC; including brentuximab vedotin [BV]) followed by 2L treatment with either CC again, EM (e.g. HDACi or DNMTi), or SMI (broad or selective). Pralatrexate and BV were included in the CC arm to simplify sequential analyses based on mechanism of antineoplastic effects. Additional details on inclusion, exclusion, and cohort assignment are described in the Supplemental methods.

Study End Points

The primary outcome was OS (time from 2L start to death or lost-to-follow-up). Comparisons were made between all potential treatment arms in 2L and 3L which resulted in 12 possible treatment paths: EM, SMI, or CC at 2L without any subsequent 3L treatment, and 9 combinations of EM, SMI, or CC at 2L followed by EM, SMI, or CC 3L. We dissected the effects at 2L and 3L independently using stability and preplanned subgroup analyses further described in the supplemental methods.

The study was conducted in compliance with the Helsinki Declaration, and institutional review board approval was obtained from the coordinating center and at each participating center per institutional standards.

Statistical, Stability, and Machine Learning Causal Inference Analyses

We used Kaplan Meier and Cox Proportional Hazards (PH) methods^18–20 to assess survival adjusting for a priori covariates described previously: age, histology, primary refractory disease (defined as lack of complete response to frontline therapy), PIT score at diagnosis, and receipt of hematopoietic stem cell transplant (HSCT) consolidation.¹ For 2L-3L sequential treatment effects, and 3L only stability analyses, the response to 2L was used as an additional covariate (Supplemental Table S2).¹ Missing data were imputed using pooled multiple imputation procedures.²¹ Preplanned stability analyses assessing effects at 2L and 3L independently were performed to validate findings and dissect sequential treatment effects with BV-based therapy as a unique arm, across histologic, HSCT, and high-risk subgroups. We combined reinforcement learning (RL) with the Cox model to guide treatment selection over time, with survival risk serving as the reward function²², and survival synthetic intervention (SSI) to assess survival for all combinations of covariates while considering latent confounders which may determine intervention selection^23–25, adjusted for the same covariates used in the Cox PH model, using bootstrapping with 100 samples to estimate 95% confidence intervals. Additional details are provided in the Supplemental Methods, and Tables S2 and S3.

RESULTS

Baseline Demographics, Clinical Characteristics, and Covariates

Of the total 1240 patients screened for the updated PETAL global cohort, 1143 from 14 sites were eligible for inclusion (Figure 1 and Supplemental Table S1). A total 540 patients were included in 2L-3L combined treatment efficacy analysis and 2L-only stability analysis (EM n=101, SMI n=45, CC n=394; Table 1 and Supplemental Tables S4). Of those, 290 (53.7%) received 3L (EM n=65, SMI n=44, CC n=181; Table 1 and Supplemental Table S5) and 280 were included into the 3L stability analyses (Figure 1). Patients who received 2L EM or SMI more often had AITL (EM: 40.4%; SMI: 57.8%; CC: 25.1%) and more often received 1L autologous HSCT consolidation (EM: 33.3%; SMI: 31.1%; CC: 14.7%; Supplemental Table S4). Patients who received 2L SMI were less often 1L refractory (SMI: 37.5%; EM: 42.6%; CC: 49.2%) and more often received 2L allogeneic HSCT (SMI: 16.7%; EM: 10.9%; CC: 9.1; Supplemental Table S4). Demographics and clinical characteristics at 3L followed similar trends and are reported in Supplemental Table S5. Demographics and clinical characteristics did not significantly deviate from the overall population for patients who received 2L combination SMI and EM therapy (Supplemental Table S6), and for patients who were alive beyond 2 years from 2L (Supplemental Table S7). All causal inference models performed well based on concordance indices adjusting for covariates listed (Supplemental Table S9).

Figure 1: Study Flow Chart CONSORT diagram.

The total number of included patients for primary combined second- and third-line (i.e. sequential) analyses was n=687. Analyses were dissected into stability cohorts for stability analyses at treatment at second-line only (n=687), third-line only, irrespective of second line therapy received (n=351), and third-line only for patients who received SMI or EM at second line (i.e. excluded those treated with CC at second-line to assess for selection bias; n=83)

Abbreviations; PETAL, PEripheral T-cell Lymphoma; 2L, second-line; 3L, third-line; EM, epigenetic modifier; SMI, small molecule signaling inhibitor; CC, cytotoxic chemotherapy.

Table 1:

Therapeutic agents used within each group at second- and third-line treatment.

Second-line (N=540)
Epigenetic Modifiers (101); n (%)		Small Molecule Inhibitors (n=45); n (%)		Chemotherapy (n=394); n (%)
HDACi monotherapy	76 (75)	Combination with romidepsin; n=13 (29)		Ifosfamide-containing	100 (25)
HDACi combination therapy	12 (12)	Lenalidomide	8 (18)	Gemcitabine-containing	92 (23)
Azacitidine and Romidepsin combination	6 (5.9)	PI3K inhibitor	4 (8.9)	DHA + platinum	44 (11)
EZH1/2 inhibitor	5 (5.0)	Alisertib	1 (2.2)	Other cytotoxic chemotherapy	43 (10.9)
Azacitidine alone	2 (2.0)	Monotherapy; n=32 (71.1)		BV only	65 (17)
		Investigational pathway inhibitors	10 (22.2)	Pralatrexate only	25 (6.4)
		Alisertib	7 (15.5)	Anthracycline-containing	21 (5.3)
		Tipifarnib	2 (4.4)	BV-CHP	2 (0.5)
		Sulanemadlin	1 (2.2)	BV-Bendamustine	1 (0.3)
		PI3K inhibitor	6 (13.3)	Pralatrexate + Methotrexate	1 (0.3)
		JAK inhibitor	5 (11)
		Calcineurin inhibitor	4 (8.9)
		ALK inhibitor	3 (6.7)
		IMiD	2 (4.4)
		Retinoid inhibitor	1 (2.2)
		BCL2 inhibitor	1 (2.2)
Third-line (N=290)
Epigenetic Modifiers (n=65); n (%)		Small Molecule Inhibitors (n=44); n (%)		Chemotherapy (n=181); n (%)
HDACi monotherapy	45 (69)	Combination with HDACi; n=7 (15.9)		Gemcitabine-containing	47 (26)
EZH1/2i alone	9 (14)	Duvelisib	4 (9.1)	Ifosfamide-containing	37 (20)
HDACi combination therapy	2 (3.1)	Proteasome inhibitor	2 (4.5)	BV only	35 (19)
HDACi + DNMT1i	5 (7.7)	Lenalidomide	1 (2.3)	Other cytotoxic chemotherapy	30 (17)
Azacitidine alone	2 (3.1)	Duvelisib	9 (20.5)	Pralatrexate only	16 (8.8)
DNMT1i combination	2 (3.1)	Lenalidomide	6 (13.6)	DHA + platinum	9 (5.0)
		Investigational pathway inhibitor	5 (11.4)	BV + Bendamustine	6 (3.3)
		Tipifarnib	3 (6.8)	BV-CHP	1 (0.6)
		IAP antagonist	2 (4.5)
		Crizotinib	4 (9.1)
		Cyclosporine	4 (9.1)
		JAK inhibitor	3 (6.8)
		BCL2 inhibitor	2 (4.5)
		Duvelisib + Bortezomib	2 (4.5)
		Lenalidomide + Ibrutinib	1 (2.3)
		Bexarotene	1 (2.3)

Specific details on all drugs and regimens received can be found in Supplemental Table S6.

Abbreviations: HDACi, histone deacetylase inhibitor; DNMTi, DNA-methyltransferase inhibitor; EZH, enhancer of zeste homolog; IMiD, immunomodulatory imide drug; PI3K, phosphoinositide 3-kinase; JAK, Janus Kinase; ALK, anaplastic lymphoma kinase; BCL2, B-cell Lymphoma-2; BV, Brentuximab vedotin.

Outcomes based on Treatment Selections

The most common therapeutic agents used at 2L and 3L are shown in Table 1 and Supplemental Table S8. Notably, 13 (27.1%) patients in the SMI arm received a small molecule drug combined with romidepsin at 2L (Table 1 and Supplemental Table S6). Of the total 540 patients, 250 (46.3%) received 3L observation or best supportive care. The number of patients included into each of the 12 potential treatment scenarios can be seen in Supplemental Table S10.

The treatment scenario of SMI at 2L, followed by EM at 3L (SMI>EM) significantly improved OS compared to CC>CC (adjusted hazard ratio [aHR]: 0.29, 95% confidence interval [95%CI] 0.11–0.74; p=0.010; Figure 2). When comparing to all other treatment scenarios, SMI>EM consistently improved OS for most other treatment sequences (Supplemental Table S11). Global availability of therapeutic agents used can be found in Supplemental Table S12 for reference.

Figure 2: Overall survival using unadjusted Kaplan Meier and adjusted Cox Proportional Hazards methods based on sequence of therapy from second- third-line in all patients (n=540).

*For ease of viewing, these arms are not shown in the Kaplan Meier figure, however they are still included in the analyses.

A treatment sequence is defined as treatment in second-line followed by treatment in third-line, represented with the “>” symbol (e.g., CC>EM refers to the treatment sequence of receiving cytotoxic chemotherapy in the second-line followed by epigenetic modifiers in the third-line).

Abbreviations: CC; cytotoxic chemotherapy, EM; epigenetic modifiers, SMI; signaling inhibitors, HR; hazard ratio, CI; confidence interval.

Stability and Machine Learning Analyses

When comparing treatment at 2L only (irrespective of treatment at 3L) versus CC, SMI (aHR: 0.45, 95%CI: 0.27–0.75; p=0.002), EM (aHR: 0.72, 95%CI: 0.50–0.98; p=0.039), and BV-based (aHR: 0.55, 95%CI: 0.35–0.85; p=0.007) therapy significantly improved OS overall (Figure 3A). In the AITL subgroup, SMI significantly improved OS versus CC (aHR 0.23, 95%CI: 0.10–0.54; p<0.001) and EM (aHR: 0.32, 95%CI: 0.12–0.82; p=0.017; Table 2). In patients with high PIT score, OS benefit was seen with EM (aHR: 0.60, 95%CI: 0.39–0.92; p=0.020) and SMI (aHR: 0.40, 95%CI 0.21–0.76; p=0.005) versus CC (Table 2). Benefit was also seen in those that did not proceed to HSCT after 2L treatment (Table 2). When repeating analyses after reassigning HDACi+SMI treated patients from the SMI to the EM arm, OS benefit was retained with EM, SMI, and BV overall (Figure 3B), EM or SMI in high-PIT score, and SMI in AITL subgroups versus CC (Table 2). Across all subgroups, BV trended towards improved OS versus CC, however no estimates were statistically significant except for those that did not proceed to HSCT (Table 2). When examining 3L-only (irrespective of 2L), there were no significant differences seen across treatment arms using adjusted or crude comparisons (Figure 3C/D). However, among patients who received 2L SMI or EM, the median OS for 3L EM, SMI, and BV were numerically higher than CC (Figure 3D).

Figure 3: Stability analyses of overall survival using unadjusted Kaplan Meier and adjusted Cox Proportional Hazards methods based on each independent line of therapy.

Stability analyses of overall survival using Kaplan Meier and Cox Proportional Hazards methods based on second-line treatment only (n=540) with EM+SMI combinations in the SMI arm (A), and EM+SMI combinations in the EM arm (B) and third-line treatment only, irrespective of second-line treatment (C; n=290) and third-line treatment only among patients who were treated with epigenetic modifiers or small molecule inhibitors at second-line only (D n=74).

Abbreviations: CC; cytotoxic chemotherapy, EM; epigenetic modifiers, SMI; signaling inhibitors, HR; hazard ratio, CI; confidence interval.

Table 2:

Overall survival subgroup analyses using Cox Proportional Hazards comparing treatment at second-line only

	EM vs CC		SMI vs CC		SMI vs EM
	aHR (95%CI)	P-value	aHR (95%CI)	P-value	aHR (95%CI)	P-value
Patients treated with EM+SMI combinations included in the SMI group
PTCL-NOS	0.70 (0.45–1.09)	0.110	0.77 (0.40–1.50)	0.443	1.10 (0.53–2.33)	0.789
AITL	0.71 (0.39–1.28)	0.250	0.23 (0.10–0.54)	<0.001	0.32 (0.12–0.82)	0.017
ALCL	0.62 (0.12–3.13)	0.562	Unable to estimate*		Unable to estimate*
PIT Score 0–1	0.87 (0.50–1.50)	0.611	0.52 (0.22–1.24)	0.140	0.60 (0.24–1.53)	0.286
PIT Score 2 or higher	0.60 (0.39–0.92)	0.020	0.40 (0.21–0.76)	0.005	0.67 (0.33–1.37)	0.272
2L Allogeneic HSCT	0.59 (0.10–3.53)	0.562	0.60 (0.06–6.30)	0.672	Unable to estimate*
No 2L HSCT	0.65 (0.45–0.93)	0.018	0.44 (0.26–0.75)	0.002	0.68 (0.38–1.22)	0.194
Patients treated with EM+SMI combinations included in the EM group
PTCL-NOS	0.68 (0.45–1.02)	0.063	1.10 (0.44–2.79)	0.838	1.63 (0.61–4.37)	0.331
AITL	0.59 (0.34–1.03)	0.066	0.21 (0.07–0.60)	0.003	0.36 (0.12–1.04)	0.060
PIT Score 0–1	0.82 (0.49–1.63)	0.445	0.36 (0.08–1.51)	0.161	0.43 (0.10–1.89)	0.265
PIT Score 2 or higher	0.55 (0.36–0.83)	0.004	0.47 (0.22–0.97)	0.042	0.85 (0.39–1.88)	0.689
2L Allogeneic HSCT	0.68 (0.14–3.26)	0.626	Unable to estimate*		Unable to estimate*
No 2L HSCT	0.62 (0.44–0.87)	0.005	0.43 (0.22–0.83)	0.012	0.69 (0.35–1.39))	0.303
	EM vs BV		SMI vs BV		BV vs CC
	aHR (95%CI)	P-value	aHR (95%CI)	P-value	aHR (95%CI)	P-value
Patients treated with EM+SMI combinations included in the SMI group
PTCL-NOS	1.15 (0.56–2.37)	0.706	1.27 (0.53–3.05)	0.590	0.61 (0.32–1.15)	0.125
AITL	1.12 (0.38–3.37)	0.835	0.36 (0.10–1.29)	0.115	0.63 (0.22–1.78)	0.382
ALCL	1.22 (0.24–6.11)	0.812	Unable to estimate*		0.51 (0.24–1.06)	0.073
PIT Score 0–1	1.24 (0.60–2.57)	0.565	0.74 (0.27–2.04)	0.565	0.70 (0.38–1.28)	0.249
PIT Score 2 or higher	1.22 (0.53–2.78)	0.641	0.82 (0.32–2.11)	0.675	0.49 (0.24–1.02)	0.056
2L Allogeneic HSCT	Unable to estimate*		Unable to estimate*		Unable to estimate*
No 2L HSCT	1.42 (0.81–2.49)	0.222	0.96 (0.48–1.93)	0.914	0.46 (0.28–0.75)	0.002
Patients treated with EM+SMI combinations included in the EM group
PTCL-NOS	1.11 (0.55–2.25)	0.771	1.81 (0.61–5.37)	0.284	0.61 (0.32–1.15)	0.126
AITL	0.96 (0.32–2.84)	0.940	0.34 (0.08–1.37)	0.130	0.62 (0.30–1.74)	0.361
PIT Score 0–1	1.18 (0.58–2.40)	0.644	0.51 (0.11–2.37)	0.392	0.69 (0.38–1.27)	0.236
PIT Score 2 or higher	1.11 (0.49–2.51)	0.805	0.94 (0.34–2.61)	0.910	0.49 (0.24–1.02)	0.056
2L Allogeneic HSCT	Unable to estimate*		Unable to estimate*		Unable to estimate*
No 2L HSCT	1.36 (0.78–2.36)	0.280	0.94 (0.42–2.09)	0.882	0.45 (0.28–0.75)	0.002

^*Unable to estimate due to small sample size.

Abbreviations: EM, epigenetic modifiers; SI, small molecule signaling inhibitors; CC, cytotoxic chemotherapy; aHR, adjusted hazard ratio; CI, confidence interval; PTCL-NOS, peripheral T cell lymphoma – not otherwise specified; AITL, angioimmunoblastic T-cell lymphoma; ALCL, anaplastic large cell lymphoma; ENKTL, extranodal natural killer T-cell lymphoma; HSCT, hematopoietic stem cell transplantation; Auto-HSCT, autologous hematopoietic stem cell transplantation; Allo-HSCT, allogeneic hematopoietic stem cell transplantation; PIT, prognostic index for T-cell lymphoma.

Reinforcement Learning methods did not find any specific treatment sequence to significantly improve OS, however there was a trend towards improved survival with EM>EM (coefficient for risk of death [Cd]: −0.65, 95%CI: −1.56–0.27; p=0.2) and SMI>EM (Cd: −0.70, 95%CI: −1.64–0.23; p=0.1), consistent with Cox PH methods (Figure 4). Survival synthetic intervention estimated a total 213 unique combinations of patient characteristics and treatments at 2L and 3L, which similarly found that patients treated sequentially with SMI>EM or EM>EM had the best relative survival compared to all other treatment strategies in similar patient subpopulations (relative survival [95%CI]: 0.966 [0.917–1.000] and 0.990 [0.832–1.145]), respectively; Figure 5 and Supplemental Data File).

Figure 4: Survival estimates across second- and third-line therapy in RR TNKL patients using a novel reinforcement learning causal inference model.

Abbreviations: CC; cytotoxic chemotherapy, EM; epigenetic modifiers, SMI; signaling inhibitors, CI; confidence interval, PTCL-NOS; peripheral T cell lymphoma – not otherwise specified, AITL; angioimmunoblastic T-cell lymphoma, ALCL; anaplastic large cell lymphoma, HSCT; hematopoietic stem cell transplantation, 2L; second-line, PIT; prognostic index for T cell lymphoma

Figure 5: Relative overall survival estimates for three common real-world clinical scenarios based on 213 unique combinations of patient factors and therapy sequences using synthetic intervention methods for select prognostic scenarios.

Relative survival estimates of a select three unique clinical scenarios representing high, intermediate, and low relative survival from a total of 213 unique estimates using a survival synthetic intervention novel causal inference model.

Abbreviations: AITL; angioimmunoblastic T-cell lymphoma, PIT; prognostic index for T-cell lymphoma, CC ± BV; cytotoxic chemotherapy with or without brentuximab vedotin, EM; epigenetic modifiers, SMI; signaling inhibitors, CI; confidence interval, 2L; second-line, 3L; third-line.

DISCUSSION

This is currently the only study examining effects across multiple lines of therapy in R/R TNKL patients. Using the largest global cohort of R/R patients, classical and novel causal inference methods, and a rigorous study design, we assessed and described survival based on the therapeutic modalities used across multiple lines of therapy in this difficult to treat and heterogeneous group of malignancies. Previous studies in the R/R setting have been limited to single arm studies assessing only a single line of therapy, with limited correlation or adjustment for the prior or subsequent treatment used.^9–11,26 Our large sample size allowed for evaluation of treatment at 2L and 3L independently and sequentially for unique subgroups including PTCL-NOS and AITL, as well as high risk populations. We were also able to identify potentially synergistic combinations such as SMI and HDACi combination therapy using robust stability analyses which may explain the benefits seen with sequential 2L SMI then 3L EM treatment seen in the primary analyses.

There are very few studies in R/R TNKL comparing contemporary therapeutic options such as EM and SMI therapies, and various possible sequences following upfront chemotherapy. Given the consistently poor survival in the R/R setting, there is an unmet need for efficacy comparisons of therapies in conglomerate.^{1–5,8,27,28} The aggressive nature of the disease at relapse also creates a dynamic challenge that necessitates evidence-based strategic use of limited available options. This study was designed to address this unique complexity considering the unobserved effects of sequential intervention analyses and to dissect the effects through stability analyses. At 2L, we identified unique groups that confer benefit with certain therapies, primarily AITL with SMI, and high-risk patients with EM (with or without SMI combinations). This is in line with previous reports which observed distinct treatment effects on survival in AITL and other nodal T-follicular helper (nTFH) cell lymphomas.¹ The phase 2 PRIMO trial of duvelisib monotherapy (n=97) reported a complete response (CR) rate of 53% in AITL and 27% in PTCL-NOS.¹⁰ Additionally, results from the LEMON-T study of lenalidomide maintenance after salvage therapy has shown promising results with a median OS of 34 months.²⁹ The benefit of SMI in AITL, a nTFHL subtype, may be due to the unique tumor microenvironment of this subtype and the on- and off-target immunomodulation effected by this class of drugs.³⁰ Golidocitinib, a novel JAK1 inhibitor, has shown overall response rates (ORR) of 56.3% and 46% in AITL and PTCL-NOS, respectively, in a recent phase 2 study.³¹ Ruxolitinib and duvelisib combination have also shown exceptional activity in nTFHL (ORR 79% and CR 63%).³² Newer SMI agents under investigation are similarly finding encouraging early response rates, e.g. cemsidomide.³³ We also identified that the combination of SMI+HDACi at 2L, most commonly romidepsin with duvelisib or lenalidomide, was a likely driver of benefit in AITL and high-risk disease. This is consistent with a large analysis which observed that HDACi combinations (majority being combinations with SMI) had higher CR rates versus HDACi monotherapy in nTFHL (38.8% vs 25.4%).³⁴ A phase 2 single arm study also reported CR rates of 36% in AITL and 17% in PTCL-NOS with lenalidomide and romidepsin.¹⁶ Another recent phase 1b/2a trial of duvelisib with romidepsin in the R/R setting saw a CR rate of 60% in AITL/nTFHL and 30% in PTCL-NOS.¹¹ Studies have identified the potential benefits of EM therapy alone in nTFHL, with nTFHL shown to be a predictor of response to HDAC inhibitors.³⁴ The randomized phase 3 ORACLE study found that oral azacytidine improved PFS versus investigators choice (5.6 vs 2.8 months)³⁵ and multiple studies examining combined azacitidine and romidepsin have shown high response rates in both phase 1, phase 2, and real-world settings.^9,36,37 Similarly in the frontline, addition of azacytidine to CHOP resulted in robust CR rates of 75% and 88% for all patients and nTFHL, respectively.³⁸ A phase 1 study of valemetostat reported high ORR across subtypes with CR rates as high as 31% and 45% in PTCL-NOS and AITL, respectively.³⁹ In our study, HDACi monotherapy was the most common agent in the EM arm, with DNA-methyltransferase inhibitors and EZH inhibitors representing a minority of EM-treated patients which may explain the limited benefit seen in AITL subgroups. Regardless, we did find that EM improved OS at 2L in the overall population and that use of EM in 3L, particularly after 2L SMI, improved survival. Lastly, while estimates in the 2L allogenic HSCT-consolidation subgroup were nonsignificant, there were higher rates of 2L allogeneic HSCT-consolidation after SMI and HDACi+SMI therapies which may represent more successful bridging to HSCT with these therapies, although transplant-related-mortality may outweigh benefits.⁴⁰ We would like to point out that with increasing use of doublet combinations of SMI+HDACi such as duvelisib and romidepsin for instance, greater number of patients are proceeding to allogeneic HSCT, the only potential curative strategy, and this in turn is likely to favorably influence the OS. Given the very small percentage of patients that proceed to HSCT-consolidation in the R/R setting, our findings warrant consideration to adding the ability to undergo HSCT-consolidation as a study endpoint when investigating these combinations in randomized trials.

The understanding and molecular characterization of TNKL has outpaced therapeutic development in this disease and the need for personalized treatment decisions is becoming increasingly important. Due to the large sample size and global nature of this study, we were able to deploy rigorous methods to evaluate multiple treatment modalities in unique subgroups and modelled the study design to reflect clinical practice given the reasonable level of clinical equipoise with respect to treatment patterns and outcomes globally.¹ The target-trial framework limits biases that can be prevalent in observational studies and is conceptionally similar to a prospective interventional study.^18,19 In clinical practice, 3L decisions are not predestined and depends on response and tolerability of 2L. although the treatment choices at 3L are largely similar. There is a significant variation of therapeutic availability globally and the decision to receive a certain therapy, or no therapy at all, can be often based solely on drug availability in certain regions of the world (i.e. treatment location unrelated to clinical factors). We modelled the study and analyses to match these practices and included patients who did not receive 3L. The additional use of novel independent causal inference methods such as RL and SSI allow for validation of classical Cox PH methods and robust survival estimation. Results from RL and SSI methods were largely consistent with Cox PH methods. While dynamic treatment decisions have been successfully modeled in other clinical settings^41,42, sequential therapy has not been formally assessed in malignancies like T-cell lymphoma where there is an unmet need given limited options. Given the robust trial design and validation of results using multiple independent models, this study provides a potential scaffold for sequential therapy assessment, which can be a valuable strategy, especially in settings of high therapeutic complexity, agnostic of malignancy type.

This study has several limitations. Some highlighted previously include systematic errors based on the retrospective nature of these analyses specifically with varied documentation over the span of 10 years across multinational sites with lack of centralized pathology and response to treatment review among others. Varied treatment practices in the choice of therapy across academic centers worldwide is a major confounder. There may be potential biases, such as possible selection bias, evident by results of the 3L stability analyses. Effects from selection bias are likely minimized (but not mitigated) with covariate adjustment encompassing a majority of decision-making. There may also be unmeasured confounding, however these can be minimized with the robust trial design and covariate adjustment, clinical equipoise reflected across arms (since similar patients could receive different agents globally), and the inclusion of observation/best supportive care as a 3L option reflecting clinical practice. Cox PH methods are inherently limited in fully investigating the effects of sequential therapy due to the proportional hazards assumption, however the likelihood for biased survival estimation is minimal because the covariates used occur, or are measured, only once prior to 2L (i.e. the first study intervention), apart from response to 2L and 2L HSCT-consolidation (used in sequential and 3L stability analyses only) which are measured once after 2L but before 3L. Novel independent RL and SSI models supported findings to Cox PH methods, even with the inclusion of the dynamic covariate of 2L treatment response. Additionally, a time-dependent Cox PH model does not fully obviate the proportional hazards assumption or covariate time-dependency²⁰, so RL and SSI models would more accurately estimate survival. We also did not have toxicity data and were unable to differentiate whether a therapy switch was due to toxicity. While this is a potential confounder for treatment decisions, it is unlikely to have directly affected survival time because toxicity would be intricately linked to patient fitness (adjusted for in analyses), and the therapy administered (i.e. treatment-related-mortality is innate to the survival estimates for each therapy) reflected in broad clinical decision-making captured in our cohort. We also would like to note that the broad categorization of drugs with different mechanisms of action under the classes of SMI or EM is an over-simplification to the heterogeneity drugs lumped under it. However, this was only meant to make analyses feasible and hence to attempt to overcome this inherent limitation of the disease and its associated treatment practices, multiple machine learning methods were used independently to confirm similar results. Our goal is to not recommend one class of drug over other but to simply provide some tools to potentially consider certain drugs over conventional chemotherapy in R/R setting if available to the clinician. Finally, randomized controlled trials that are biomarker informed will be needed to validate these preliminary results. Lastly, pralatrexate and BV-based regimens were included in the CC arm for sequential analyses as their cell cytotoxicity is like chemotherapy. The overall utilization of these therapies was low so their effects in the CC arm are likely minimal. To explore these effects, BV-based therapies were a unique arm in 2L and 3L stability analyses. PETAL consortium investigators are now focused on integrating clinical and molecular information and harnessing machine learning for precision medicine on prospectively enrolled patients.

In conclusion, we used a rigorous trial design and multiple independent causal inference methods with the largest R/R TNKL global cohort to compare contemporary therapeutic strategies at 2L, 3L, and sequentially across these lines of therapy. The use of SMI or EM (both combined and monotherapy) at 2L followed EM at 3L demonstrated an OS benefit over other treatment sequences. At 2L, benefit was especially evident in AITL and high-risk (PIT score ≥2) populations, supporting their use over classical CC. Results were consistent across all stability analyses and independently applied methods despite differing model prediction metrics. This study highlights the urgent need for equitable drug access globally and supports the direction of novel therapeutic advancement in this difficult-to-treat population. Ultimately leveraging distinct molecular vulnerabilities through serial genomic profiling of PTCL subtypes will be critical to the rational use of drug-drug combinations in the right sequence.

Statement of Significance.

There is no universal therapeutic approach for relapsed/refractory T-cell/NK-cell lymphomas. Optimization of approved treatment sequences in second- or third-line could be critical to augmenting survival across cancers. Using the largest relapsed/refractory global patient cohort and causal inference methods, we propose a potential schema for assessing treatment effects at each line.

KEY POINTS.

• Targeted small molecule inhibitors with and without epigenetic modifiers offer a survival advantage in R/R AITL and high-risk groups.

• Therapy effect at each line can be assessed using high quality observational data with rigorously employed novel causal inference methods.

DATA SHARING STATEMENT:

For original data please contact salvia.jain@mgh.harvard.edu