Comparing Patient-Derived Xenograft and Computational Response Prediction for Targeted Therapy in Patients of Early-Stage Large Cell Lung Cancer

This article is featured in Highlights of This Issue, p. 2097

Non–small cell lung cancer (NSCLC) is the most common cause of cancer-related death in developed as well as in developing countries causing more than a million deaths each year on a global scale (1, 2). In recent years, molecularly targeted therapy (TT) improved treatment of NSCLC. Targets focus on inhibition or suppression of highly deregulated mitogenic signaling pathways in cancer cells, including EGFR, MAPK, MET, and MYC pathways, and others (3–5). Tyrosine kinase inhibitors (TKI) such as erlotinib, gefitinib, and cetuximab have been approved for clinical TT of NSCLC by the FDA, the European Medicines Agency (EMA), and the China Food and Drug Administration (CFDA). When applied appropriately, TT often demonstrates clinical efficacy associated with improved overall survival and the abatement of symptoms (6–8).

Despite these promising results, one major challenge is to develop effective strategies to select patients with considerable benefits from a specific TT. Studies have focused on potential driver mutations and their association with TT outcome. For instance, mutations in the EGFR gene might be a powerful biomarker for high sensitivity to targeted TKIs (9–11). Although it has been reported that mutated KRAS might contribute to the resistance of TKIs treatment in NSCLC, the prevalence of this mutations is low, about 5% to 10% in Asian and <20% in other countries (12). The fusion between the echinoderm microtubule-associated protein-like 4 (EML4) and the ALK gene, has been reported to be associated with positive clinical outcome of crizotinib treatment (13, 14). Despite these novel discoveries, the prognosis of NSCLC has remained poor, and the 5-year survival is below 20% in most countries (15).

Recently, several studies provide evidence that patient-derived xenograft (PDX) show biologic consistency with the tumor of origin and are phenotypically stable at the histologic and genomic level across multiple generation (16, 17). Moreover, drug-sensitivity pattern of PDXs can recapitulate response in the patients from which they derived (18). Unfortunately, the timeframe to develop these so-called Avatar models (PDXs) is usually long compared with the clinical decision timeframe for a patient. Furthermore, because of the technical limitations, normally in 30% of patients, the generation of a PDX will fail at the first time. This failure rate increases the expense of the PDX approach significantly. There is an ongoing discussion whether PDX is appropriate for clinical decision making (19–21).

Can the response prediction by cheap in silico strategies compete with that of the PDX? Therefore, we compare response prediction of a computational approach with that of the PDX in a small cohort of NSCLC patients. Our in silico strategy combines formal knowledge of cancer-related cellular processes (represented by a molecular signaling map, MSM) with individual patient data in a Flux-comparative-analysis (FCA) and focuses on the change of signal flux within the underlying molecular cellular system by facing external perturbation, including drug treatment. The change of signal flux in a simulated cellular system might reveal the treatment response of the corresponding individuals (22). Therefore, we reasoned that the FCA applied to our MSM might be able to mimic treatment response of the PDX. We use the next-generation RNA seq as specific information of an individual tumor sample to initialize our algorithm (23). The MSM contains a global cellular signaling network and quantifies different hallmarks of tumorigenesis. It also allows modeling the intervention of a TT and calculation of change of the signal flux between a treatment state and a control state, which is the basic strategy of FCA for enabling the individual response prediction to the TT. So far, the MSM has been applied to investigate the therapeutic responses of AML and colon cancer patients under TT. This is the first study where the algorithm is adapted to NSCLC.

This study consists of two arms (Fig. 1). The first arm involves the PDX, in which individual tumor samples are implanted. Two xenografts are produced for each patient, one to be treated with erlotinib, the other to be treated with cetuximab. The response measurement consists in the comparison of tumor volume at day 25 (V₂₅) with tumor volume at baseline (V₀). A tumor responds to a treatment if it's volume shrinkage below 40%: V₂₅/V₀ < 0.4 (<40%).

The second arm involves NGS-based RNA-seq and computational treatment simulation. Gene-expression profiles (GEP) are generated from tumor samples of each lung cancer patient, which is the input data to the MSM. Response is the relative change of a score quantifying cancer proliferation in the FCA (22). To this end, the proliferation score is calculated from the MSM without implementing the targeted treatment (control state, PRO_control) as well as from the MSM with implementing the targeted treatment (treatment state, PRO_treat). The quantified response prediction is S = log₂(PRO_treat/PRO_control). A negative value predicts an individual response to TT, zero or a positive value predicts a TT failure. Goal of the study is to compare responses as observed in the PDX with response as predicted by the MSM-based approach.

Surgical biopsies of 25 patients were obtained from primary diagnosed, early-stage large cell lung cancer (NSCLC) patients between January 2010 and March 2013 from the Yu Huang Hospital (Yu Huang, Zhe Jiang, PR China). Written informed consent was obtained from each patient. The PDX model was created to decide whether cetuximab or erlotinib would be the appropriate TT. This study was approved by the medical ethics committee of the Yu Huang hospital.

All tumor samples were put into medium under the controlled sterile condition immediately after surgical resection. For each sample, high-throughput dataset was generated by using an Illumina HiSeq (Illumina). The tool TopHat (version 2.0.6) was applied to generate reads from the raw RNAseq data with h19 genome. The quality control assessment was conducted by using the tool RSeQC (version 2.6.1). Gene annotations were obtained from the Illumina and gene counts were generated by applying the tool HTSeq (version 0.6.1). The reads per kilobase per million mapped reads (RPKM) was applied to measure the mRNA abundance. To reduce different possible potential bias factors (such as GC content, gene size, and others) of this type of mRNA measurement, the conditional quantile normalization with the cqn Bioconductor package (24) was performed.

Nine- to 10-week-old SCID and nude (nu/nu) mice (Vital River) were used for receiving surgical tumor samples. The gender of the mice was chosen according to the gender of donors. All animals were housed in pathogen-free facilities with abundant food and water under the guidance approved by the Institutional Animal Care and Use Committee (IACUC). PDXs were established by transplanting fresh tumor tissues on mice as described in the study (18). After stabilization of tumor-implanted mice, 25 mice were treated with cetuximab, which was administrated every 3 day for five total injections at the dose levels range from 1 mg per injection to 0.03 mg per injection. Another 25 mice were administrated with erlotinib every 3 day for five total injections at the dose levels ranging from 15 mg per injection to 5 mg per injection. Tumor volume was recorded twice per week and calculated with L × W² × π/6. Antitumor activity was defined on the basis of RECIST criteria (25). Tumor shrinkage (TS) ≤ 40% on the day 25 after start of treatment was considered as positive response: V₂₅/V₀ < 0.4.

In the previous study (23), we describe the construction of a large-scale MSM, which consists of 58 cancer-relevant signaling pathways, including different feedback-loops and miRNA regulations. To be able to reflect different aspects of human tumorigenesis, the MSM also incorporates 10 cancer hallmarks (26, 27). The basic modeling pattern in the MSM is that each gene participates in one or more transcription reactions to generate one or more mRNAs; each mRNA participates in one translation reaction to generate a corresponding protein. Each gene, mRNA and protein in the MSM are labeled with their corresponding unique ID, including Ensemble-ID for gene and mRNA; uniProt-ID for protein. Individual patient gene-expression data are used to initialize the MSM model: Each gene in the model is set to the values from gene-expression data and miRNA expression data as given by the associated IDs.

Cetuximab (Erbitux) is a chimeric human-murine monoclonal IgG1 antibody, which is designed to block EGFR-mediated signaling pathways through binding to the extracellular domain of the EGFR. Erlotinib (Tarceva) is a quinazolinamine, which binds in a reversible manner to the adenosine triphosphate (ATP) binding site of receptors, including EGFR, ERBB, and others. Both drugs have been approved by the FDA and are frequently applied drugs for TT for treating lung cancer (28, 29). From a molecular perspective, the drug effect of cetuximab is a single function inhibition at the extracellular domain 3 of EGFR, which is 10-fold greater affinity (0.087 nmol/L) than natural ligand, whereas the drug effect of erlotinib consists of inhibitions of multiple tyrosine kinases with different binding affinities that includes EGFR (0.67 nmol/L), ERBB4 (230 nmol/L), LYN (530 nmol/L), and SRC (700 nmol/L) (30). The mass action law is applied to model the inhibition effect of both drugs in the MSM:

where [Target:Drug] = [Target] × [Drug]/BA; [x]: concentration of component x; BA: binding affinity (Fig. 2A).

FCA is applied to investigate whether a drug could cause a significant flux change in a model system between two different states: control (C) and treatment (T). The MSM was reported in the previous study (23) and is the model system for the FCA. All MSM components are either reactions (including phosphorylation, transcription, and others) or bio-objects (including gene, protein, and others). Thus, the MSM is converted to a B-tree (31) with reaction-nodes (R) and bio-object-nodes (B; Fig. 2B). Each child node is a data-array, whose elements are reactions (R) and bio-objects (B) as defined in the MSM. After the B-tree conversion, the gene components in the MSM can be set to the values of transcriptome data as given by the Ensemble-ID. After initialization, the B tree is converted into a Petri Net (P) with the definition: P = (R, B), where the R contains all transitions (reactions) and the B contains all places (bio-objects; Fig. 2B). The simulation of P is performed by iterating transitions in R for firing evaluation X, until the P reaches a steady. The simulation for each treatment takes 50 minutes within a computer of duo CPU, each with 2.66 GHz, 4Gb memory and 400 MT/s.

Descriptive statistics are used to present differences between groups. Scores are compared by using boxplots. Differences are assessed by the non-parametric Mann-Whitney test. The discriminative power of scores between two groups is analyzed by ROC curves and the AUC summary measure (32). The correlation between two scores is studied by Spearman correlation.

Twenty-five surgical biopsies histopathologically diagnosed as early stage of NSCLC from Yu Huang hospital from January 2010 and November 2013 were collected for this study together with the patient's demographic and clinical information, including age, gender, stage, histology, and others. The mean age (SD) of patients was 62.08 ± 11.87. No patient received any anticancer treatment before diagnosis. The main smoking history of patients was non-smoking (20/25, 80%). The majority of NSCLC cases were in the stage I (19/25, 76%; See Table 1).

Because of high efficacy of TKIs in NSCLC (33, 34), cetuximab and erlotinib were chosen to investigate prediction of treatment response. The individual (MSM and PDX based) predicted responses scores under both treatments are listed in the Table 2 (two left columns). For the MSM-based prediction, negative values indicate a positive response.

Under cetuximab the minimal score (maximal score) was −2.400 (1.280) with mean (median, SD) of 0.004 (0.090 and 0.890). Under erlotinib the minimal (maximal) score was −2.900 (1.350) with mean (median, SD) of 0.144 (0.440 and 0.94). The paired results in both MSM models are presented in Fig. 3A. The reaction under cetuximab cannot be statistically distinguished from the reaction under erlotinib (Wilcoxon test P value, 0.3746), which is demonstrated by a tight scattering of the combined measurements around the diagonal (line of equality). This figure also shows the predicted response rate under cetuximab of 48% (12/25; 95% CI [27.8%, 68.7%]) and under erlotinib of 40% (10/25; 95% CI [21.13%, 61.34%]).

The FCA results show that within the responders of cetuximab treatment, signal flux of key components such as FLT1 (>30%), CTNNB1 (>31%), JAK2/3 (>43%), MYD88 (>42%), MYC (>45%), and PDGF (>30%) from diverse signaling pathways, including EGFR, VEGF, WNT, JAK-STAT, TLR, MAPK, and PI3K–AKT, PDGF signaling pathways are clearly downregulated, which indicates the downregulation of proliferation, growth, and angiogenesis in the responders (Supplementary Table S1). Interestingly, many components from the cell-cycle pathway are also downregulated. This reduces the replicative potential of cancers (27). Furthermore, signal flux of several components, including TSC1/2 (>23%), RPTOR (69%), AMPK (50%) from the AMPK pathway, are highly downregulated, which leads to restrict the high energy demand for cancer development. It is noteworthy that signal flux of several important transcription factors, including CREB (>24%), ELK4 (30%), HIF2 (47%), JNKs (33%), SRF (30%), are also downregulated (Supplementary Table S1).

For the responders of erlotinib treatment, the FCA results show that signal flux from many key components in pathways, including EGFR, PI3K/AKT, MAPK, WNT, VEGF, cell cycle, PDGF, and other signaling pathways, are downregulated (Supplementary Table S2). The downregulation of EGFR, PI3K/AKT, MAPK, JAK–STAT, cell cycle, AMPK signaling pathways is a common phenomenon for the predicted responders of both TKI treatments, which retains proliferation and growth of cancer cells. This conclusion is partially in an agreement with several independent studies that discovered crucial signaling pathways for tumorigenesis: The prosurvival arm of pathways comprising PI3K–AKT–mTOR and the proliferative arm consisting of the RAS–RAF–MEK–ERK pathways (35).

For each of these 25 patient two avatar models of his/her cancer are prepared and used for treatment response testing. The biologic response on the treatment was measured by a change score for the tumor volume (TS, tumor shrinkage: V₂₅/V₀): The percentage of volume between start and day 25 of treatment. Shrinkage below 40% was considered as response (Table 2, two right columns). Under cetuximab maximal (minimal) tumor shrinkage was 1% (88%) with mean (median, SD) of 44.68% (47.00%, 27.8%). Under erlotinib maximal (minimal) tumor shrinkage was 23% (98%) with mean (median, SD) of 62.64% (67.00%, 22.6%) (Table 2, two right columns). The paired results in the avatar model are presented in Fig. 3B. Tumor shrinkage under cetuximab is significantly more extend as under erlotinib (Wilcoxon test P value: 0.001562), which is demonstrated by a large number of combined measurement below the diagonal (line of equality). This figure shows the response rate under cetuximab of 40% (10/25; 95% CI [21.13%, 61.34%]) and a response rate under erlotinib of 20% (5/25; 95% CI [6.83%, 40.70%]).

Furthermore, we systematically compared the PDX responses with that in silico response. Figure 3C and D show the correspondence between the TS measurement and the in silico response score under cetuximab and erlotinib, respectively. The correlation (Spearman) between both measurements under cetuximab is as expected negative (rho = −0.66, P = 0.0003431). Higher tumor shrinkage implies lower MSM prediction scores. The correlation (Spearman) between both measurements under erlotinib is as well negative but not significantly different from 0 (rho = −0.35, P = 0.1115). Under cetuximab, both approaches result in positive response predictions for 9 patients and for 12 patients in simultaneous negative response prediction. Divergent results exist for 4 patients. The overall agreement between both responses under cetuximab is 84% (21/25; 95% CI [63.92%, 95.46%]). Under erlotinib, both approaches result in positive response predictions for 5 patients and for 15 patients in simultaneous negative response prediction. Divergent results exist for 5 patients. The overall agreement between both response under erlotinib is 80% (20/25; 95% CI [59.30%, 93.17%]). Under both treatments the overall agreement is significantly different from random agreement. Moreover, to investigate how well the MSM score predicts the biologic responses from PDX, the ROC analysis (36) has been performed, which shows that under cetuximab the AUC is 0.86 (95% CI [0.65, 0.96]); under erlotinib the AUC is 0.915 (95% CI [0.78, 0.99]) (Fig. 3E and F).

Among the 25 recruited lung patients, 11 left the hospital shortly after diagnosis and could not be followed. Five of the remaining 14 patients chose TT (cetuximab) as part of the adjuvant therapy according to the PDX results (Table 3). The responses of all 5 are in agreement with the individual in silico prediction result.

In the past several years, molecularly TT has been demonstrated its high potential for lung cancer treatment as monotherapy as well as combination therapy. How to identify patients who can derive most benefit from a specific TT is currently the major challenge, which is also one of the obstacles to make TT a standard routine in clinical practice.

To our knowledge, this study is the first study that compares PDX and a computational strategy with regard to the treatment response of TT for NSCLC patients. The use of PDX is one strategy under investigation to identify potential responders. Although the PDX is expensive, has a high labor-intensity and time-consumption, many recent studies have demonstrated its potential utility in the clinical application for different cancer types (37–39). Nevertheless, one major disadvantage is that the average timeframe for generating response data from PDX would not be less than 15 days, in this study, 25 days. Furthermore, recent studies provided compelling evidence about the strong impact of social distress on the tumor growth in PDXs (40, 41). Stress-related signaling cascades such as adenylyl cyclase/cAMP/PKA signaling (42), MAPK–ERK signaling (40), and Hedgehog signaling (41) may confound the treatment effects of interest. There is the potential bias of producing false response signals: The TT may be effective, but social distress may impair this effect.

As alternative, we propose to use a computational approach based on individual molecular patient data, an MSM and the FCA. The MSM-based response prediction suggests that TKI treatments (cetuximab and erlotinib) imply therapeutic effects, if the signal flux within several mitogenic signaling pathways, including EGFR, PI3K–AKT, and MAPK is downregulated. Activities of different mitogenic pathways are critical for the uncontrolled proliferation and invasion of lung cancer cells (3–5). This is seen in the individual simulation and may elicit the cancer-relevant deregulation at molecular level as well as individual treatment response prediction. The proposed computational approach has been explored in several studies (22, 23). The algorithm can be used to analyze data of an individual patient with respect to several targeted therapies. The computing effort using a standard computer is approximately an hour. By comparison, predicting response for several treatments using PDX is restricted by the amount of tumor tissue and is not affordable in most labs.

In this phase I prognostic study (proof of concept), the PDX- and MSM-based response prediction were compared in 25 NSCLC patients and for two targeted therapies (cetuximab, erlotinib). For both substances, the results demonstrated a potentially strong agreement between PDX- and MSM-based prediction [overall response agreement: 84% (cetuximab) and 80% (erlotinib)]. In terms of a, ROC analysis, it became obvious that MSM response scores may be good surrogates for a complicated PDX prediction [AUC: 0.86 (cetuximab) and 0.915 (erlotinib)]. It could not be clarified in this study how much the disagreement between MSM- and PDX-based response prediction (Fig. 3C and D) may be caused by social stress promoted growth of NSCLC in the xenografts. For both treatments, the MSM-based response prediction is always better compared with the PDX-based (cetuximab: MSM 48% vs. PDX 40%; erlotinib: MSM 40% vs. PDX 20%). The striking difference between MSM and PDX prediction for erlotinib may also be caused by modeling effects. So far, we did not check how the infrastructure of the MSM or the algorithms used for the FCA may be biased with respect to different backgrounds regarding cancer biology and implementation of treatment effects, because all mice were housed under the same standard facility condition we would not expect a noticeable stress effect. However, the stress-related signaling cascades such as adenylyl cyclase/cAMP/PKA signaling (42), MAPK/ERK signaling (40), and Hedgehog signaling (41) are well implemented in the MSM model. To settle this question, we started a project in which patient, PDX, and model-based response can be compared and quality differences between MSM-based and PDX-based response prediction can be quantified. These results will be available within next 2 years. In summary, these results of our study emphasize the potential of the computational molecular modeling for TT and an effective alternative for expensive, labor-intensive and time-consuming PDX studies.

This study is based on a homogeneous well-specified sample of patients. Regarding the functional principle of the computational approach, it is possible to extend the treatment simulation for predicting the therapeutic effect of multiple drugs, which might function as an important part for a drug discovery process. The limitations of our study are: the small sample size of 25 patients does not allow reaching substantial conclusions. It was possible to compare both strategies, but we could not validate the prediction with final clinical result of every patient. Only 5 patients were treated with TT of cetuximab. Moreover, the current limitation of the MSM-based computational approach is that MSM is a large-scale signaling network map, and can therefore only predict the responses from treatment of signaling agents such as cetuximab, imatinib, lapatinib, temsirolimus, and others. Although the FCA, which is based on Petri net, is not so much dependent on kinetic parameters as the standard ordinary differential equation system (ODE) does, the more detailed kinetic parameters are applied, the more accurate could be the result of FCA. Therefore, information from kinetic databases such as Brenda (43) will be used to improve the kinetic parameters of MSM in future studies.

In conclusion, we see the MSM-based computational approach as a potentially effective prediction strategy to determine individual TT for NSCLC patients. Our proof-of-concept study (phase I prognostic study) demonstrates the potential substantial agreement between a PDX- and an MSM-based response prediction.

U.R. Mansmann reports receiving speakers bureau honoraria from Baxter Austria and Pfizer Germany. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J. Li, C. Ye, U.R. Mansmann

Development of methodology: J. Li, U.R. Mansmann

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C. Ye

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J. Li, U.R. Mansmann

Writing, review, and/or revision of the manuscript: J. Li, U.R. Mansmann

Study supervision: J. Li, U.R. Mansmann

This work is funded by the German Cancer Consortium (DKTK) and German Cancer Research Center (DKFZ).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.