Combination Olaparib and Temozolomide in Relapsed Small-Cell Lung Cancer

Small-cell lung cancer (SCLC), which accounts for approximately 15% of all lung cancers, is a high-grade neuroendocrine carcinoma with high metastatic potential and poor clinical outcomes. The most common genetic alterations in SCLC are inactivation of TP53 and RB1, and no clear targetable alterations have been described (1–3). Untreated SCLCs tend to be highly sensitive to cytotoxic chemotherapy; response rates to first-line etoposide/platinum (EP) are 50% to 70% (4, 5). The duration of time from completion of first-line chemotherapy (generally 4–6 cycles are given) to disease progression, often referred to as the degree of “platinum sensitivity,” predicts likelihood of response to second-line cytotoxics (6). However, response rates in the second-line setting are significantly lower, ranging, approximately, from 5% to 30% (6–10).

Recently, targeting DNA-damage repair by inhibition of PARP has emerged as a potential therapeutic strategy in SCLC (11). Although SCLCs are not characterized by homologous recombination (HR) deficiency or mutations in BRCA1/2, PARP inhibitors have shown some activity in SCLC preclinical models and early-phase trials (12–15). However, the single-agent activity of PARP inhibitors in SCLC appears to be minimal. Notably, the UK STOMP trial failed to show an improvement in progression-free survival (PFS) for patients treated with olaparib in the maintenance setting (16). PARP inhibitors may synergize with agents that increase the prevalence of single-stranded (ss) DNA breaks (17). Mechanistically, this combination strategy is supported even in HR- and BRCA-intact tumors by the observation that trapping of PARP complexes to sites of ssDNA breaks leads to failure of repair and induction of double-stranded (ds) breaks (15, 18). Addition of the PARP inhibitor veliparib to temozolomide in a randomized phase II trial led to improved response rate (39% in the veliparib/temozolomide arm vs. 14% in the placebo/temozolomide arm), but no significant improvement in the 4-month PFS or median PFS (mPFS; ref. 13). The ECOG-ACRIN 2511 study demonstrated a modest improvement in mPFS with the addition of veliparib to cisplatin and etoposide (mPFS, 6.1 vs. 5.3 months; HR, 0.75, one-sided P = 0.01), but no significant improvement in overall survival (OS; ref. 14). Olaparib, which has stronger PARP-trapping activity than veliparib (18), may offer significant antitumor efficacy while maintaining a manageable therapeutic window in combination with temozolomide.

As a field, SCLC research suffers from a paucity of available tissue for preclinical and coclinical investigations. Initial diagnoses are often made from scant aspirate tissue, and repeat biopsies after initial treatment are typically outside of standard of care. We have previously reported generation of a panel of patient-derived xenograft (PDX) models of SCLC that recapitulate patient tumor genomic features and sensitivity to EP (19). Here, we develop and test the clinical activity of combination olaparib and temozolomide (OT) in patients with SCLC, and then apply a PDX panel to discover molecular signatures predictive of response to this treatment.

The trial consisted of a phase I dose-escalation portion and a phase II multistage portion. The phase I portion (Fig. 1, left) was a conventional 3+3 dose-escalation design with a primary objective of determining the recommended phase II dose (RP2D) of combination olaparib tablets and temozolomide, both dosed days 1 to 7 of each 21-day cycle. Dose-limiting toxicities (DLT) were monitored during cycle 1. One patient enrolled at dose level 3 was not DLT-evaluable and was therefore replaced, per protocol. The phase II portion of the study (Fig. 1, right) was a dose expansion at the RP2D, dose level 3, with the primary objective to assess efficacy as measured by overall response rate (ORR). This was a multistage optimal design to allow for early termination due to lack of efficacy.

Between October 2015 and April 2018, we enrolled 50 patients with previously treated SCLC to the clinical trial. The baseline demographics of patients enrolled study-wide are shown in Table 1. The majority of patients (86%) had an Eastern Cooperative Oncology Group Performance Status (ECOG PS) of 1, and the number of prior lines of cancer therapy ranged from 1 to 7 (median 2). Seventy-two percent were platinum-sensitive, defined here as having a chemotherapy-free interval from completion of first-line therapy to initiation of second-line therapy of at least 90 days. The study required baseline brain imaging, and asymptomatic untreated brain metastases measuring <1 cm were allowed for eligibility. Twenty patients (40%) had metastases at baseline, 12 of which were untreated at study entry. Demographics of the subset of patients treated at dose level 3/RP2D are shown in Supplementary Table S1.

At the four dose levels tested in the phase I portion, there were no DLTs, serious adverse events (SAE), or grade 4 or 5 treatment-related toxicities. We observed increasing neutropenia, anemia, and thrombocytopenia in dose levels 3 and 4 compared with dose levels 1 and 2. In addition, one patient treated at dose level 4 experienced grade 3 vomiting. Although none of these adverse events met DLT criteria, dose level 3 was selected as the RP2D based on the phase I experience.

Summaries of treatment-related adverse events (TRAE) study-wide and among patients treated at dose level 3/RP2D are shown in Supplementary Table S2. Study-wide, the most common TRAEs were thrombocytopenia, anemia, and neutropenia (occurring in 68%, 68%, and 54% of patients, respectively), although the majority were grades 1 to 2. Rates of fatigue, nausea, and vomiting were 50%, 42%, and 22%, respectively, and again the majority were grade 1 or 2. Notable AEs included two grade 5 events that occurred during the phase II portion and that were deemed possibly related to study drugs. One was due to pneumonia and occurred during cycle 1. The other was due to neutropenic sepsis and occurred during cycle 3. There was one case of pneumonitis, which was grade 3 and thought to be related to study drugs. This event occurred after two cycles of treatment and resolved with steroids.

Dose reductions were performed per protocol and at the investigator's discretion. Among 41 patients treated at the RP2D, dose reductions occurred in 44% overall and 64% of those who received at least 3 cycles (Supplementary Table S3).

Study-wide, 48 patients were evaluable for response assessment. The confirmed ORR was 41.7% (20/48; Fig. 2A and B). Four additional patients had unconfirmed partial response (PR); two had PR on the first scan and then progressive disease (PD) on the subsequent scan, and two had a PR on the first scan and then expired prior to a confirmatory scan. Among patients treated at dose level 3/RP2D, confirmed ORR was 41% (16/39). Responses were seen at all dose levels (Fig. 2B). Among the responders and nonresponders, the median number of prior lines of therapy was 2 and 1 (mean 2.1 and 2), respectively. The median duration of response (mDOR) was 4.3 months (Fig. 2C). After a median follow-up of 7.1 months, the mPFS was 4.2 months (95% CI, 2.8–5.7; Fig. 2D) and the median overall survival (mOS) was 8.5 months (95% CI, 5.1–11.3; Fig. 2E). Among patients treated at dose level 3/RP2D, the mDOR was 5.3 months (95% CI, 2.7–5.8), mPFS was 4.2 months (95% CI, 2.5–5.7), and mOS was 6.7 months (95% CI, 4.6–12.6). Nine patients continued on treatment post-progression because of ongoing clinical benefit in the opinion of the treating investigator, with the duration of treatment post-progression ranging from 3 weeks to 12 months.

In an exploratory analysis, we further assessed efficacy based on platinum sensitivity (Fig. 3). The confirmed ORR was 47.1% among the 34 patients with platinum-sensitive disease, and 28.6% among the 14 patients with platinum-resistant disease (Fig. 3A and B). The mDOR among patients with at least one time point demonstrating PR was 4.2 months for platinum-sensitive (20 patients) and 4.5 months for platinum-resistant (4 patients; Fig. 3C). The mPFS was 4.5 months and 2.9 months, and mOS was 9.4 months and 7.4 months, among platinum-sensitive and platinum-resistant patients, respectively (Fig. 3D and E). There was no statistically significant difference observed in PFS (HR = 0.76, P = 0.400) or OS (HR = 1.05, P = 0.898) when comparing outcomes among platinum-sensitive and platinum-resistant patients, although the small size of these cohorts limits the power to detect a significant difference in this study.

To model clinical OT sensitivity and resistance in the laboratory setting, we generated a series of 6 PDX models from 4 patients treated during the phase I portion of the trial. All 4 patients, MGH1518, MGH1528, MGH1543, and MGH1514, had clinical benefit on OT, with RECIST 1.1 best responses of −61%, −50%, −41%, and −28%, and time to progression of 6.8, 5.5, 3.9, and 3.1 months, respectively (Fig. 4A–D). These patients had received 1, 7, 1, and 4 prior lines of therapy, respectively. PDX models were developed from these patients prior to receiving OT (MGH1518-1B, MGH1528-1, and MGH1514-5) and at the time of disease progression (MGH1518-3, MGH1528-2, and MGH1543-1). These 6 PDX models were treated with a single 5-day cycle of twice-daily olaparib plus daily temozolomide. Following treatment, PDX tumor responses were compared with those of their donor patients (Fig. 4E–H). To quantify tumor responses, we measured maximum tumor regression after treatment (best response) and days to 200% initial tumor volume [time to progression (TTP)]. By both metrics, PDX sensitivity mirrored patient tumor sensitivity to OT at the time of model generation. Specifically, PDXs from OT-naïve patients who went on to have durable PRs, MGH1518-1B and MGH1528-1, regressed completely in all mice (100% response) and did not relapse for at least 80 days (Fig. 4E and F, blue curves). Patient MGH1514 had a modest tumor response (−28%) and a brief PFS (3.1 months); the corresponding PDX MGH1514-5 derived at the time this patient enrolled on the OT study showed transient partial regression (−40% ± 4.4% days SEM) followed by progression (42.5 ± 3.1 days SEM; Fig. 4H, yellow curves). In contrast, in PDXs derived from 3 patients at the time of progression on OT, tumors showed minimal responses and progressed at nearly the same paces as vehicle-treated controls (Fig. 4E–G, red curves). Collectively, these data gave us confidence that the PDX tumors accurately recapitulated the sensitivity and resistance of their donor patient tumors.

We extended the coclinical trial to a total of 32 PDX models derived from 22 patients (Fig. 4I; Supplementary Fig. S1A). The additional 26 models in this panel were not derived from patients who had been treated with OT, but were an unselected cohort of consecutively derived models, with the goal of representing the biological diversity seen among patients with SCLC. This panel included 13 models from chemotherapy-naïve patients and 19 models generated from patients after at least one prior line of therapy. Two to six replicate xenografts per model were assessed for OT sensitivity using the 5-day single-cycle treatment protocol, with best response and TTP measured to assess efficacy (Fig. 4J; Supplementary Fig. S2A). We observed a range of responses across the models (Fig. 4J), consistent with the range of responses seen in our patient cohort (Fig. 2A). Replicate xenografts demonstrated highly concordant tumor volume curves following OT treatment (average SEM for model response = ±8.4% and TTP = ±3.1 days), and across the 32-model panel the best response and TTP metrics were tightly inversely correlated (Pearson r = −0.87, Supplementary Fig. S2B). OT-sensitive models were substantially less sensitive to single-agent olaparib or temozolomide (Supplementary Fig. S2C), indicating that PDX responses are dependent upon the combined activity.

The clinical course of SCLC is generally characterized by initial sensitivity to DNA-damaging therapies such as EP, followed by relapse and broad cross-resistance to second-line agents (6). Acquisition of cross-resistance remains an ongoing clinical challenge, and yet it is difficult to study because direct comparison of different regimens is not possible in the same patient. The SCLC PDX panel captures intertumoral heterogeneity and provides a unique opportunity to study cross-resistance. We have previously demonstrated that PDX models faithfully recapitulate clinical sensitivity to EP (19). A direct comparison of the patient platinum sensitivity and corresponding PDX model EP sensitivity (Supplementary Fig. S3A–S3D) further supports this conclusion. For OT, we also observe a high degree of clinical fidelity in models derived from trial patients (Fig. 4A–H). We therefore applied the models to investigate patterns of cross-resistance between EP and OT.

We compared in vivo drug response to each regimen with both PDX clinical history and gene-expression levels across 32 models (Supplementary Fig. S1A). PDX TTP on EP was superior for those models derived from treatment-naïve patients compared with those derived from previously treated patients, and a similar trend was observed for OT in the same models (Fig. 5A). We then directly compared in vivo sensitivity for each regimen in each model and observed a moderate correlation (Pearson r = 0.56; Fig. 5B). Both results suggest partial cross-resistance to these regimens. To interrogate the molecular underpinnings of these observations, paired-end transcriptome sequencing was performed in untreated replicate xenografts for each model, and replicates were averaged for transcript abundance (Supplementary Table S4). The transcriptional profiles associated with EP sensitivity correlated well with the profiles associated with OT sensitivity (Pearson r = 0.68, Fig. 5C). Furthermore, there was a notable absence of genes whose expression could stratify the regimens by marking sensitivity to one but not the other. These results argue that clinical resistance to EP may also predict resistance to OT, and no clear markers emerge from this data set to distinguish tumors sensitive to one regimen but not the other.

To identify candidate molecular pathways associated with EP and OT sensitivity and resistance, gene set enrichment analysis (GSEA) was performed across PDX models using the transcription data. When we compared enrichment scores for the regimens, we found that significantly enriched gene sets for both EP and OT were closely related (Fig. 5D). Interestingly, inflammatory-response gene sets (IFNα and IFNγ, inflammation, TNFα and TGFβ) enriched for sensitivity to both EP and OT across regimens (Fig. 5D). In contrast, MYC-regulated transcripts correlated with resistance to the regimens (Fig. 5D). Leading-edge analysis revealed a high degree of overlapping genes between the members of each set responsible for the enrichment scores. These were combined into expression signatures for MYC targets (65 genes) and inflammatory response (82 genes), and when mapped to the PDX panel these signatures labeled models that were most resistant to both EP and OT (MYC target-high, inflammatory-low; Fig. 5E and F). We further assessed whether the recently described subclassification of SCLC tumors by expression of the transcriptional regulators ASCL1, NEUROD1, YAP1, and POU2F3 could identify EP/OT-sensitive or resistant models (Supplementary Fig. S4A–S4C). All four categories were represented by at least one model, but the number of models in each category was limited and did not significantly correlate with cross-resistance (20, 21). This analysis reveals the potential of clinically representative PDX models to interrogate molecular signatures associated with cross-resistance to therapies.

Although our data suggest that EP resistance may predict OT resistance, this can be a difficult metric to apply clinically, and there may still be a valuable role for a molecular biomarker in further clinical development of OT. We therefore analyzed the relationships between OT sensitivity and gene expression to identify candidates predictive of response to this regimen. The six models derived from patients treated with OT on the clinical trial (Fig. 4A–H) were used as calibration points to divide the 32-model discovery set into OT-sensitive and OT-resistant cohorts. MGH1514-5, which was derived pretreatment from a patient with RECIST stable disease (−28% best response), and which had a deeper best response and longer TTP than the post-relapse models MGH1528-2 and MGH1543-1, was selected to mark the boundary delineating sensitive and resistant cohorts in the PDX panel (Figs. 4J and 6A; Supplementary Fig. S1A). Each transcript was scored for differential expression between the sensitive and resistant cohorts (Fig. 6A; Supplementary Table S5). Genes with significant differential expression (>2-fold, FDR < 10%) were further evaluated for (i) performance as a classifier (ROC curve AUC > 0.8), and (ii) ease of distinguishing expression states (e.g., on/off, Bimodality Index > 1.1; ref. 22). In total, 216 genes satisfied these criteria for marking OT sensitivity, and 79 genes for OT resistance (Supplementary Table S5).

We treated an additional 11 SCLC PDX models for which RNA sequencing had not been performed with OT in vivo to serve as a validation set for candidate biomarkers (Fig. 6B; Supplementary Fig. S1B). The validation set was separated into sensitive and resistant cohorts by best response (−40% threshold from MGH1514-5), which correlates tightly with TTP (Supplementary Fig. S2B). We hypothesized that genes that were members of underlying pathways upregulated in sensitive or resistant tumors would be more likely to validate in unknown tumors. Therefore, we used the EP/OT cross-resistance signatures (Fig. 5E and F) to guide candidate selection (Fig. 6A, blue and red). Twenty-four of the 82 members of the inflammatory-response signature satisfied all statistical criteria for enrichment in the discovery set (Supplementary Table S5). We tested four of these candidates by quantitative RT-PCR in the validation set: CEACAM1 (CD66a), TNFSF10 (TRAIL), TGIF1 (TGFβ-induced factor I), and OAS1. Although our analysis represents basal gene expression across the PDX panel, IFN-induced expression has been well characterized for three candidates (CEACAM1, TNFSF10, OAS1) and TGFβ-induced expression for TGIF1 (23–29). CEACAM1, TNFSF10, and TGIF1 were expressed at significantly higher levels in the 5 OT-sensitive validation models, and OAS1 performs as a complementary biomarker that enhances the combined specificity for OT sensitivity (Fig. 6C; Supplementary Fig. S5A). Aggregate expression of these four candidates strongly distinguished sensitive from resistant models in both the discovery and validation sets (Fig. 6C; Supplementary Fig. S5A).

Although the MYC-target signature marked OT resistance, the dynamic range for its component transcripts was low (Fig. 6A; Supplementary Table S4), with only EIF4A1 demonstrating significant differential expression in resistant models. Even for EIF4A1, basal expression across the panel was high, and distributions in resistant versus sensitive cohorts showed small differences and tight variation consistently in the discovery and validation sets (Supplementary Fig. S5B). The MYC-target signature represents an aggregate of moderate differences in gene expression, rather than the standout individual transcripts in the inflammatory-response set. Interestingly, SNAI2 (SLUG), encoding a regulator of the epithelial-to-mesenchymal transition (EMT) and neural crest development (30–33), was among the best transcriptional classifiers of resistance outside the MYC signature (Fig. 6A), and a role for EMT has been proposed for chemoresistance in both SCLC and other solid tumors (12, 31, 34). High SNAI2 expression also marked OT resistance in the validation set, and classifier performance was maintained at the protein level (Supplementary Fig. S5C and S5D).

Expression biomarker candidates that emerged from the RNA-sequencing screen were compared with genes with strong preclinical rationales: PARP1, MGMT, and SLFN11 (11–13, 15, 35). In the discovery set, SLFN11 expression performed as a classifier for OT (AUC = 0.71, 95% CI = 0.53–0.89), with optimal performance observed at thresholds that would distinguish the most resistant PDX models (Fig. 6D; Supplementary Fig. S5D and S5E). Neither PARP1 nor MGMT significantly distinguished OT-sensitive and OT-resistant cohorts, as shown in ROC curve AUC values (Fig. 6D). In contrast, the inflammatory-response genes (CEACAM1, TNFSF10, OAS1, TGIF1) significantly outperformed the hypothesis-driven candidates as markers of OT sensitivity, and SNAI2 as a marker of resistance, in both the discovery and validation sets regardless of threshold (Fig. 6E and F; Supplementary Fig. S5A). These candidate biomarkers emerged from an analysis of cross-resistance and are unlikely to differentiate EP from OT, but instead may identify SCLC tumors with broad vulnerability to DNA-damaging regimens.

SCLC remains one of the most aggressive and lethal malignancies, with a minimal role for curative surgery, no targetable driver mutations, and an mOS for metastatic disease generally less than 1 year. Nonetheless, SCLC is initially highly sensitive to DNA damage, and combinations that pair DNA-damaging agents with inhibitors of DNA-damage checkpoints and DNA repair, such as inhibitors of PARP, WEE1, CHK1, and ATR, have recently emerged as promising new strategies to treat relapsed SCLC (13–15, 18, 36–39). Two challenges unify these approaches: (i) development of biomarkers to identify sensitive cases, and (ii) determination of the common features of tumors that are cross-resistant to EP and subsequent DDR inhibitors. Recent progress has been made on both fronts, with the identification of SLFN11 as a marker of sensitivity for both PARP inhibition and EP (12, 13, 40), although unbiased screens for clinically relevant biomarkers have not been achieved. Here we describe a single-arm phase I/II study of OT in relapsed SCLC, with an integrated coclinical trial in PDX models to discover candidate biomarkers and dissect cross-resistance to chemotherapy.

Among 50 patients with relapsed SCLC, combination OT shows a strong signal of clinical efficacy. Although cross-study comparisons are difficult to interpret, the confirmed ORR of 41.7% and mPFS of 4.2 months are numerically superior to those attained in several recent second- and third-line SCLC studies (9, 41, 42). The addition of veliparib to temozolomide also showed improved responses in a placebo-controlled randomized phase II study, further supporting this strategy (13). In contrast, monotherapy with the PARP inhibitor talazoparib led to only two partial responses among 23 patients with SCLC treated in a phase I expansion study (43), and monotherapy olaparib did not confer a PFS benefit compared with placebo when used as a single agent in the maintenance setting after first-line EP (16). The superiority of the combination over monotherapy was also evident in our PDX models, where OT was significantly more active than either drug alone (Supplementary Fig. S2C). Collectively, these data support the further development of strategies that combine induction of DNA damage with PARP inhibition. Such an approach in the SCLC second-line setting is timely, given that standard of care is shifting to incorporate immune checkpoint blockade in combination with EP in the first-line setting (5).

A recent preclinical study also demonstrated striking synergy when DNA damage–repair inhibitors (either olaparib or the CHK1 inhibitor prexasertib) were combined with PD-L1 inhibition (44). Unfortunately, combination durvalumab and olaparib showed minimal clinical activity (ORR 10.5% among 19 evaluable patients) in an early-phase clinical trial (45). Although the low response rate to durvalumab/olaparib is disappointing, it should not stop further investigation of similar combinations. Mechanistically, if combination PARP inhibitor + temozolomide more potently induces DNA damage than single-agent PARP inhibitor, as appears to be the case in clinical studies, OT may also be more robustly potentiated than olaparib alone when combined with an immune checkpoint inhibitor. A follow-on study combining OT with a PD-1/PD-L1 inhibitor may be warranted. We would argue that this strategy could be beneficial even in SCLCs that have progressed after first-line chemotherapy/immuno-oncology (5), if, for example, cGAS and STING pathways were more robustly activated by OT than by carboplatin/etoposide.

Cross-resistance to DNA-damaging regimens remains a major challenge in the management of relapsed SCLC. We established an experimental system tailored to study cross-resistance: a large collection of PDX models with diverse clinical histories in which the models derived from patients resistant to either EP or OT recapitulated this behavior in vivo. We compared EP and OT efficacy with gene expression across the panel and found that the transcriptional profiles of drug resistance were highly similar, permitting an initial description of cross-resistant SCLC: low basal expression of mediators of innate immunity and increased expression of MYC-regulated transcripts. The high MYC-target signature echoes our previous findings for EP alone across a smaller panel of PDX models (19), but the low inflammatory signature was both robust and novel. Although the SCLC PDXs accurately recapitulate clinical sensitivity to EP and OT, they are tested in NSG mice, with near-total impairment of adaptive immunity. A similar signature of 38 IFN-stimulated genes (ISG) was recently observed in a pan-cancer analysis of The Cancer Genome Atlas, although not in the context of drug resistance (46). This signature could not be attributed to immune or stromal infiltrates, but instead to chronic IFN production from the cancer cells themselves. Furthermore, the same ISG signature was observed in a large collection of 379 PDX models of diverse tumors grown in immunocompromised nude mice, and in the complete absence of an immune system in cancer cell lines grown in vitro (46). The SCLC PDX inflammatory signature in drug-sensitive models may represent a marker of increased endogenous DNA damage, as through replication stress, and this may decrease cellular tolerance of exogenous DNA-damaging agents and inhibitors of repair. In this case, transcription of inflammatory-response genes could result from low levels of cytosolic dsDNA detected via the cGAS–STING pathway (47–50), and it has recently been demonstrated that treatment with olaparib can activate the pathway in genetically engineered mouse models of SCLC (44). Alternatively, components of the inflammatory response may play an active role in mediating the cytotoxicity of DNA damage, independent of CD8⁺ T-cell or natural killer–cell function. The cross-resistant PDX models identified here will provide valuable tools for investigating these possibilities and uncovering the mechanistic underpinnings of a clinical phenotype that has a profound impact on survival in SCLC.

We were able to generate six PDX models from four patients on the OT trial. Although additional PDX models from patients on trial may have further enhanced our translational studies, these six models nonetheless facilitated a search for dichotomous biomarkers for OT by forging a direct connection between the clinical and coclinical trials (Fig. 4A–H). These patient:PDX anchor points were used to calibrate in vivo model responses according to clinical history, allowing a rational division of the 32-model discovery set into sensitive/resistant cohorts. From the comparison of EP and OT transcriptional profiles, we recognized that the strongest connection between drug response and gene expression was the low inflammatory signature for cross-resistance. Four components of this signature that have been well characterized as genes upregulated in response to either IFN signaling or TGFβ were tested in a validation set of 11 additional PDX models. These genes significantly outperformed current hypothesis-driven biomarker candidates for PARP inhibitor combinations in SCLC that have emerged from recent studies (12, 15, 40): SLFN11, MGMT, PARP1, ATM, and EMT signature. The differences across these studies may be due to differences in model systems, specific PARP inhibitor drugs and dosing strategies, study designs, or other factors. Ultimately, the optimal strategy for clinical application may be to combine candidates, as aggregate expression of CEACAM1 (CD66b), OAS1, TNFSF10 (TRAIL), and TGIF1 strongly distinguished OT-sensitive from OT-resistant models. The performance of these IFN response genes as classifiers in the validation set further supports the inflammatory-response signature as a marker of sensitivity to OT, and likely other DNA-damaging regimens, in SCLC. One limitation of our study is that the clinical trial did not require archival or fresh tissue for eligibility, and we had insufficient archival tissue available from patients on the study to be able to perform further validation of proposed biomarkers. This experience highlights the potential added value of requiring tissue samples from patients enrolling onto studies in the future.

Although SCLC has historically been treated as a homogenous disease, this paradigm is poised to shift with identification of functionally distinct subgroups (20) that may benefit from different therapies. Our PDX coclinical trial strategy can be applied to early-phase SCLC trials to accelerate this progress. The anchor-point trial design leverages the size of the PDX panel without sacrificing clinical context, presenting advantages over other laboratory models such as established cell lines as well as scarce archival patient samples. Each model within this diverse and well-annotated panel can be assessed for sensitivity to multiple therapies, permitting direct comparisons to segregate SCLC into functional as well as molecular subgroups. Predictive biomarker discovery is not the only application of the coclinical trial strategy; well-validated PDX models from trial patients can be used to optimize dosing schedules, and serial models derived before and after therapy represent a powerful tool for uncovering mechanisms of acquired resistance. We anticipate that PDX coclinical trials will provide critical contributions to breaking the monolithic view of SCLC and establishing personalized medicine for this recalcitrant disease. This approach may be further applied more broadly to refine patient selection and drug development strategies for therapies with a signal of clinical activity in early-phase trials.

This was a single-arm, open-label, single-institution phase I/II study (NCT 02446704). The study was reviewed and approved by the Dana-Farber/Harvard Cancer Center Institutional Review Board (IRB). Written informed consent was provided by all participants, and the study was performed in accordance with the Declaration of Helsinki. Patients who met the key inclusion criteria were those ≥18 years old with histologically or cytologically confirmed SCLC who were not candidates for curative therapy, and who had received prior first-line EP with subsequent radiographic progression. Any number of interval prior therapies were allowed. Eligibility criteria were the same for the phase I and phase II portions of the study, as detailed in the clinical protocol (Supplementary Data).

In the phase I portion of the study, patients were treated using a 3+3 dose-escalation study design. The starting dose level was olaparib tablets 100 mg orally twice daily and temozolomide capsules 50 mg/m² orally every evening, both administered days 1 to 7 of a 21-day cycle. Subsequent dose levels escalated one drug at a time, with the same days 1 to 7 administration schedule (Fig. 1). Treatment was continued until disease progression, concurrent illness preventing further administration of treatment, or DLT not resolving within 7 days, or at the discretion of the participant or investigator. Continuation of treatment post-progression was allowed in cases of ongoing clinical benefit. Cycle length could be extended to 28 days at the investigator's discretion in cases where additional time was needed to recover from treatment-related adverse events. Growth-factor support was not allowed.

The phase II portion of the study was a dose expansion at dose level 3, which was selected as the RP2D (Fig. 1), with the primary objective to determine ORR. During the initial 20-patient expansion, patients requiring dose reductions could be reduced by >1 dose level at a time (i.e., dose level 3 to 1) at the treating physician's discretion. A second 20-patient expansion was added in a protocol amendment. In the second expansion group, if dose reductions were required below the RP2D, patients were reduced first to a new dose level 2A (olaparib 200 mg orally twice daily and temozolomide 50 mg/m² orally every evening) for at least 1 cycle, and then to dose level 1 if further reduction was required.

Clinical and laboratory evaluations were required weekly during treatment cycle 1 and on day 1 of each subsequent cycle. DLTs were monitored during cycle 1. Toxicity assessments were performed using the NCI Common Terminology for Adverse Events (CTCAE v4.0). Response assessment by CT of the chest, abdomen, and pelvis was performed every 6 weeks. Brain imaging by MRI or CT was required at baseline and subsequently at the discretion of the investigator. Best objective response was determined using RECIST 1.1 criteria (51).

The primary endpoint of the phase I portion was presence of DLT, with the objective of determining the MTD and RP2D of combination olaparib and temozolomide. The primary endpoint of the phase II portion was ORR. For the phase II portion, a multistage optimal design was used to allow early termination due to lack of efficacy. After the first 9 evaluable patients were treated at the RP2D, accrual was continued only if at least one partial or complete best overall response had been observed, in which case the protocol then allowed enrollment of an additional 11 patients for a total accrual of 20 at the RP2D. If at least 4 responses were observed among these 20 patients, the protocol proceeded to enroll an additional 20 patients at the RP2D. The multistage design provided 87% power to determine that the RP2D of combination olaparib and temozolomide was associated with a 30% ORR if at least 6 patients achieved a response in total. The study design is associated with type I error of 9% if the underlying rate of overall response were truly 10% suggesting a lack of efficacy. Best objective response was determined using RECIST 1.1 criteria (51). All patients who had at least one response assessment performed were included in the response evaluation. All enrolled patients were included in the PFS and OS analyses. Duration of response, PFS, and OS rates were estimated by the Kaplan–Meier method, and platinum-sensitive and resistant groups were compared using the score test of the proportional hazards model. Data analysis was performed using SAS 9.4 (SAS Inst Inc), and P values were based on a two-sided hypothesis. A data cutoff of May 31, 2018, was used for toxicity analysis, and November 6, 2018, was used for efficacy analysis.

All tissue and blood samples from patients were collected per IRB-approved protocols with written informed consent from the patients and in accordance with the Declaration of Helsinki. All mouse studies were conducted through Institutional Animal Care and Use Committee–approved animal protocols in accordance with Massachusetts General Hospital institutional guidelines. The source of tumor material for model generation was either core-needle biopsy or circulating tumor cells. PDX model generation and treatment with EP and OT were as described previously (19). Notably, for models previously reported (19), additional replicate xenografts have been added. Furthermore, new models not treated with either regimen are included in this dataset. Briefly, treatment studies were initiated at xenograft volumes = 400–600 mm³ for 2–6 mice per model per treatment arm, and tumors were measured 2–3 times weekly. OT: olaparib 50 mg/kg oral gavage (OG) d1–5 + temozolomide 25 mg/kg OG d1–5. EP: cisplatin 7 mg/kg intraperitoneal (i.p.) d1, 8 + etoposide 10 mg/kg i.p. d1–3, 8–10. Trial tumor metrics: TTP = days from start of treatment to 2× initial tumor volume (ITV), response = change in tumor volume between ITV and d7–28 minimum, endpoints: tumor volume > 2× ITV or 100 days after start of treatment.

Paired-end transcriptome sequencing was performed in biologic duplicate for each model. Two untreated mice per model were selected with xenograft volumes of 800–1,200 mm³, and tumor fragments were flash-frozen in liquid nitrogen within 5 minutes of euthanasia. Frozen tissue fragments were lysed and homogenized with the TissueLyser (Qiagen), and RNA extractions were performed with the Qiagen RNeasy Mini Kit. RNA quality was assessed with a 2100 Bioanalyzer DNA Chip 7500 (Agilent Technologies) and samples with an RNA integrity number of more than 8 were selected for library construction. cDNA libraries were prepared with the Illumina TruSeq Stranded mRNA Sample Kit and sequenced with a paired-end 2 × 75 bp protocol on an Illumina HiSeq 4000 instrument at the Novartis Institute for Biomedical Research. Sequence alignments were performed against a chimeric transcriptome index consisting of human and mouse reference genomes (hg38 and mm10) to sort PDX- and mouse-specific reads. Transcript abundance values, transcripts per million (TPM), were estimated by Salmon (0.8.2; ref. 52) using the following options (-l ISR, –seqBias, –gcBias, –useVBOpt). Gene-level expression was generated using the R tximport package (53). As expression values approximated a log-normal distribution, expression values (TPM) were log-transformed [log₂(1+TPM)]. Model-level expression values were calculated by averaging gene expression across biological replicates. Low-expression genes [max log₂(1+TPM) <1] and expressed genes with more than 25% average variation between replicates (mean replicate CV > 0.25) were removed.

Untreated PDX transcript abundance [log₂(1+TPM)] was compared with mean PDX TTP following treatment with EP versus OT for each of the 32 models in the discovery set by Pearson correlation, and the threshold for significant correlation was assigned at an FDR-adjusted P value of 10% (Benjamini–Hochberg method). GSEA (54) was performed for transcript abundance versus EP and OT TTP (treated as continuous variables, Pearson method), using the Hallmark gene set collection (MSigDB v6.2, 5,000 permutations). Normalized enrichment scores were compared and significant enrichment was assigned for FDR <25%. Leading-edge analysis was performed for each significantly enriched gene set for EP or OT, and genes present in the leading edge of at least one gene set (union) for both regimens (intersect) were assigned to one of two expression signatures: inflammatory response (positively correlated with EP/OT TTP, 82 genes) or MYC target (negatively correlated with EP/OT TTP, 65 genes). Gene set signature scores were calculated as the average transcript abundance z-scores of the signature genes for each model versus the PDX panel.

Transcript expression levels were compared in OT-sensitive versus resistant cohorts, with model MGH1514-5 demarcating the two cohorts. Genes with at least 2-fold change in expression between the cohorts and an FDR-adjusted Welch t test P value of less than 10% (Benjamini–Hochberg method) were selected for further analysis. Each differentially expressed transcript was assessed for biomarker suitability using two criteria: performance as a two-population classifier and bimodal distribution of expression levels. Bimodal expression was estimated using the bimodality index (ref. 22; BimodaIndex R package), a measure of the degree to which gene expression across the PDX panel fits a two-component mixture model, and classifier performance was estimated by the ROC AUC (pROC R package; ref. 55).

Total cell RNA was extracted from PDX snap-frozen tissue using the Qiagen RNeasy Mini Kit (Qiagen, catalog no. 74104). cDNA was synthesized from 1 μg total RNA, using oligo-dT priming and the TaqMan Reverse Transcription Kit (Applied Biosystems, catalog no. N8080234) following the manufacturer's instructions. Relative expression of transcripts was quantified by real-time PCR, using the FastStart Universal SYBR Green Master Mix (Roche, catalog no. 4913914001) and the LightCycler 480 System (Roche). mRNA levels were normalized to ACTB. Primer sets are listed in Supplementary Methods.

A.F. Farago reports receiving commercial research grants from AbbVie, AstraZeneca, PharmaMar, Merck, Bristol-Myers Squibb, Loxo, Bayer, Ignyta, Genentech/Roche, and Amgen and is a consultant/advisory board member for PharmaMar, AbbVie, AstraZeneca, Loxo, Bayer, Genentech, Boehringer Ingelheim, Roche, Bristol-Myers Squibb, and Millennium Pharmaceuticals. R.S. Heist is a consultant/advisory member for Tarveda, Apollomics, and Boehringer Ingelheim. D. Rangachari is a consultant/advisory board member for DynaMed and Advance Medical. D.A. Barbie reports receiving a commercial research grant from Novartis, has received other commercial research support from Bristol-Myers Squibb, has ownership interest (including stock, patents, etc.) in Xsphera Biosciences, and is a consultant/advisory board member for Tango Therapeutics, N of One, and Madalon Consulting. L.V. Sequist is a consultant at AstraZeneca, Janssen, Merrimack, Genentech, Blueprint, and Pfizer and reports receiving commercial research grants from AstraZeneca, Boehringer Ingelheim, Merrimack, Blueprint, LOXO, Genentech, and Novartis. A.N. Hata reports receiving commercial research support from Novartis, Pfizer, Amgen, Relay Therapeutics, and Roche/Genentech. S. Maheswarana has ownership interest (including stock, patents, etc.) in CTC-iChip. A.T. Shaw is a member of the board of directors of Syros and is a consultant/advisory board member for Pfizer, Novartis, Chugai, Ariad/Takeda, Ignyta, Achilles, TP Therapeutics, Taiho, Daiichi-Sankyo, Foundation Medicine, Guardant, Genentech/Roche, Blueprint Medicines, KSQ Therapeutics, Servier, EMD Serono, Bayer, Natera, and Loxo. M. Mino-Kenudson is a consultant/advisory board member for Merrimack Pharmaceuticals and H3 Biomedicine. N.J. Dyson has received commercial research support from Novartis. B.J. Drapkin reports receiving commercial research grants from Novartis, AbbVie, Merck, and AstraZeneca. No potential conflicts of interest were disclosed by the other authors.

Conception and design: A.F. Farago, B.Y. Yeap, A.T. Shaw, N.J. Dyson, B.J. Drapkin

Development of methodology: A.F. Farago, S. Maheswaran, D.A. Haber, B.J. Drapkin

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): A.F. Farago, M. Stanzione, Y.P. Hung, R.S. Heist, J.P. Marcoux, D. Rangachari, D.A. Barbie, S. Phat, D.T. Myers, T.D. Dubash, L.V. Sequist, A.N. Hata, D.A. Haber, M. Mino-Kenudson, B.J. Drapkin

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A.F. Farago, B.Y. Yeap, M. Stanzione, R.S. Heist, J. Zhong, R. Morris, L.V. Sequist, M.S. Lawrence, A.T. Shaw, M. Mino-Kenudson, B.J. Drapkin

Writing, review, and/or revision of the manuscript: A.F. Farago, B.Y. Yeap, Y.P. Hung, R.S. Heist, J.P. Marcoux, D. Rangachari, S.R. Digumarthy, L.V. Sequist, A.N. Hata, A.T. Shaw, M. Mino-Kenudson, N.J. Dyson, B.J. Drapkin

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A.F. Farago, B.Y. Yeap, M. Kem, E.A. Kennedy, S. Maheswaran, B.J. Drapkin

Study supervision: A.F. Farago, E.A. Kennedy, N.J. Dyson, B.J. Drapkin

Other (image interpretation and response evaluation): S.R. Digumarthy

We are grateful to the patients and families who participated in these research studies. We thank the members of the MGH thoracic oncology group and other MGH Cancer Center staff for assistance with recruitment of patients and collection of samples, and Edwin Choy for his guidance around initial protocol design. We thank I. Sanidas, A. Guarner-Peralta, B. Krishnan, P. Rumde, V. Kamesan, J. Grinnell, and current and former members of the Dyson and Farago research groups; and L. Zou, R. Corcoran, and C. Benes for critical discussions and scientific input. We thank Jeffrey Engelman for his vision and guidance in launching this project. This work has been supported by AstraZeneca (ISS22810111, to A.F. Farago), NCI grant U01CA220323-A1 (to N.J. Dyson and A.F. Farago) and U24CA213274 (to A.F. Farago), the V Foundation translational grant no. T2016-003 (to N.J. Dyson), NIH career development award K12CA087723 (to A.F. Farago), career development awards from Uniting Against Lung Cancer (to A.F. Farago) and the Lung Cancer Research Foundation (to B.J. Drapkin), the ASCO Young Investigator Award (to B.J. Drapkin), the research fellowship grant from the Deutsche Forschungsgemeinschaft (to M. Stanzione), NIH RO1 CA1299933 (to D.A. Haber), the Howard Hughes Medical Institute and National Foundation for Cancer Research (to D.A. Haber), and a research agreement through the Novartis-MGH Alliance.