Apoptosis Modelling for Treatment Decisions in Colorectal Cancer
ROYAL COLLEGE OF SURGEONS IN IRELAND
Saint Stephen'S Green 123
Higher or Secondary Education Establishments
€ 932 803,80
Jochen Prehn (Prof.)
Sort by EU Contribution
€ 508 000
THE QUEEN'S UNIVERSITY OF BELFAST
€ 353 442
UNIVERSITE PARIS DESCARTES
€ 362 400
€ 461 160
€ 64 800
KLINIKUM DER JOHANN WOLFGANG VON GOETHE UNIVERSITAET
€ 316 876,20
Grant agreement ID: 306021
1 November 2012
31 October 2014
€ 3 896 336
€ 2 999 482
ROYAL COLLEGE OF SURGEONS IN IRELAND
This project is featured in...
New biomarkers for colorectal cancer management.
Grant agreement ID: 306021
1 November 2012
31 October 2014
€ 3 896 336
€ 2 999 482
ROYAL COLLEGE OF SURGEONS IN IRELAND
This project is featured in...
Final Report Summary - APO-DECIDE (Apoptosis Modelling for Treatment Decisions in Colorectal Cancer)
APO-DECIDE is an SME-focused project that explored the potential of a systems-based analysis of apoptosis signalling pathways to deliver new prognostic and predictive biomarkers for the management of advanced colorectal cancer (CRC), and new tools for patient stratification for therapies targeting apoptosis deficiencies in cancer (IAP antagonists). APO-DECIDE focused in particular on caspase signalling pathways which are activated in response to genotoxic chemotherapy and that are targeted by IAP antagonists.
APO-DECIDE delivered a valuable biobank of CRC and matched normal tissue from three CRC patient groups (untreated, genotoxic chemotherapy-treated [5-Fluorouracil (5-FU)-based] and anti-EGFR therapy-treated) collected from sites in Ireland, Northern Ireland and France. As data integration and exchange is key in systems biology-based research, APO-DECIDE has also developed an integrated database, complete with genetic tumour characterisation, clinical data, and patient follow-up collected prior to and during the project. This database was instrumental for the project and is an important resource for future European systems medicine research projects.
APO-DECIDE utilized and developed deterministic models of caspase activation pathways based on ordinary differential equations (ODEs) previously characterised by the partners (APOPTO-CELL; Rehm et al., EMBO Journal, 2006). As protein quantities are key for a systems-based evaluation of cancer cell apoptosis in response to chemotherapy, APO-DECIDE also developed and validated two quantitative protein profiling platforms for use in a clinical setting, based on formalin-fixed, paraffin-embedded (FFPE) tissue: reverse phase protein (lysate) array (RPPA) technology and digital immunohistopathology. The RPPA platform was validated and proved to be a reliable protein profiling platform. Digital immunohistopathology was also used to identify protein biomarkers.
Systems modelling of caspase signalling using RPPA data and the APOPTO-CELL modelling platform identified a high variability in the predicted ability of individual patients to activate executioner caspases. Importantly, it delivered a novel tool that separated stage III CRC patients who will respond to 5-FU-based chemotherapy from those who will not, and may require additional targeted therapies instead.
APO-DECIDE also investigated the ability of the APOPTO-CELL systems model to predict the impact of IAP antagonists (SMAC mimetics) on induction of apoptosis in CRC. The project validated the potential of IAP antagonists as anticancer therapeutics in combination with 5-FU-based chemotherapy in CRC, using data from in vitro, in vivo and in silico experiments. Initial data acquired from patient-derived xenograft mouse models also indicated that the use of IAP antagonists led to apoptosis induction and reduced tumour cell viability. Single cell imaging and in vitro experiments also indicated that IAP antagonists have dual activity in inducing caspase activation. The APOPTO-CELL-SMAC model extension was developed. With further validation, APOPTO-CELL-SMAC model could select patients for IAP antagonist/Smac mimetic therapy.
Working with the academic partners, SME OPTIMATA integrated in vitro and in vivo protein expression data relevant for APOPTO-CELL parameterisation, as well as PK/PD data for 5-FU, Oxaliplatin and IAP antagonists within their proprietary virtual patient OVP modelling environment. OVP-based simulations allowed the quantitative replication of tumour growth curves in xeno-transplanted mice. Using an in silico human model, we conducted a virtual trial that demonstrated that CRC patient populations treated with 5-FU-based chemotherapy should see prolonged and increased survival from co-administration of IAP antagonists.
In summary, APO-DECIDE has successfully developed a clinical research infrastructure and an associated, SME-focused quantitative protein profiling and advanced systems modelling platform. The project demonstrated the potential of systems modelling of caspase activation pathways as patient stratification tools for the clinical management of stage 2 and 3 CRC patients. APO-DECIDE has also delivered a platform for a new generation of pharmacokinetic/phamacodynamic stratification tools for 5-FU-based genotoxic therapies and IAP antagonists that require testing in future clinical trials.
Finally, APO-DECIDE delivered proof-of-concept for the incorporation of virtual clinical trials into clinical trial design.
Project Context and Objectives:
Colorectal cancer (CRC) has among the highest incidence and mortality rates of any cancer with over 1.2 million cases diagnosed and over 608,700 deaths from this disease worldwide. The current standard of care for CRC patients is primarily dictated by its disease stage. When diagnosed at an early, localized disease stage (stage 1), patients undergo curative surgical resection. Stage 2 patients, who exhibit localized spread with no lymph node involvement, are treated surgically and those with high risk clinical-pathological features (T4, lymphovascular invasion, poor differentiation, perforation, obstruction) will receive adjuvant 5-fluorouracil (5-FU)-based chemotherapy. The 5-year survival of stage 2 patients undergoing surgical resection alone is approximately 75- 80%, and adjuvant 5-FU-based chemotherapy has been shown to result in improvement in survival of only 3-4%. In stage 3 disease, where there is nodal involvement (N1: 1-3 nodes involved and N2: > 3 nodes involved) the current treatment paradigm is surgical resection followed by adjuvant 5-FU-based therapy with the addition of the platinum agent oxaliplatin (FOLFOX regimen) or the topoisomerase I inhibitor, irinotecan (FOLFIRI regimen). Adjuvant 5-FU/oxaliplatin treatment in node-positive stage 3 disease benefits approx. 15%-20% of patients. However, a majority of these patients will relapse or develop distant metastases within 5 years following their surgery.
Due to its asymptomatic nature, colorectal cancer (CRC) is frequently diagnosed in its later stages in which there is local, lymphatic or metastatic spreading of the disease, requiring adjuvant or palliative chemotherapy. Overall, it is estimated that 50% of CRC patients eventually develop metastases for which palliative systemic treatment is administered. In metastatic stage 4 disease, mortality from CRC is high. The response rate to palliative FOLFOX or FOLFIRI treatments are around 40%-50%, but median overall survival remains low at around 16-19 months. Recently, the treatment paradigm for metastatic stage 4 CRC has evolved in complexity to include newly developed, ‘targeted’ therapeutics.
Research into targeted molecular therapies has led to the approval of anti-EGFR therapy for CRC. Addition of the EGFR-targeted monoclonal antibodies cetuximab or panitumumab to standard chemotherapy treatment has significantly improved progression-free survival (PFS) and overall survival (OS) of patients with metastatic stage 4 CRC. Anti-EGFR therapy inhibits cell survival signalling in cancer cells when the small GTPase KRAS is not mutated. KRAS is a small GTPase that is essential for the activation of downstream cell survival signalling pathways, including the Ras/Raf/ERK and PI3K/Akt pathways. The existence of activating KRAS mutations therefore abolishes the ability of cetuximab to inhibit cell survival signalling in cancer cells.
Since its approval, cetuximab has become a $2.5 billion per annum drug, and more than 50% of patients with advanced CRC are currently treated with anti-EGFR therapy. Nevertheless, half of these patients will not benefit from the addition of anti-EGFR agents despite the fact that their tumour is KRAS wild-type. This situation clearly demonstrates the need for predictive markers that can be used in the context of ‘targeted’ treatment approaches.
DNA damaging agents, such as 5-FU and inhibitors of cell survival signalling such as cetuximab, seek to induce tumour regression through induction of apoptosis or sensitization to apoptosis. Dysfunctional apoptosis is well recognized as a key contributing factor in malignant progression and development of chemotherapy resistance in cancer. The APO-DECIDE consortium has developed an integrated platform that delivers combinatorial predictive biomarkers and associated protein profiling tools for a systems-based analysis and prediction of therapy responses in CRC patients in order to deliver novel systems medicine tools for clinical decision making and personalised oncology approaches that will lao enable ‘smart’ clinical trials designs in the future. APO-DECIDE has developed novel comprehensive and holistic predictive models of apoptosis sensitivity that integrate EGFR signalling, pharmacokinetics/-dynamics (PK/PD) and whole body physiology, and has experimentally and clinically validate model predictions. In addition, APO-DECIDE has delivered novel quantitative protein profiling techniques based on high-throughput protein arrays and digital histopathology technologies that enable the application of systems medicine approaches in routine clinical laboratory settings or large clinical trials.
APO-DECIDE responds to the need to define a holistic and more global strategy for the treatment of colorectal cancer patients. In order to achieve this, the project has addressed three key areas:
1. Clinical validation of novel systems medicine approaches including the creation of a clinical research infrastructure
2. Implementation and development of novel quantitative protein profiling techniques enabling high-throughput protein profiling in clinical trials and in the routine hospital laboratory setting, and
3. Development and experimental and clinical validation of more holistic systems models.
APO-DECIDE has made significant progress in each of these areas. The project has delivered a platform that enables a successful and co-ordinated implementation of these approaches at a European level by incorporating leading experts in the field of apoptosis signalling, apoptosis deficiency in cancer, mathematical modelling, as well as clinical oncology and targeted therapies for CRC. APO-DECIDE has targeted important bottlenecks in systems medicine, and delivered research and commercial outputs that will give rise to novel and significant scientific and commercial opportunities.
The main Scientific Objectives of APO-DECIDE have been as follows:
1. Delivery of predictive computational models and validation of combinatorial biomarker sets that predict chemotherapy responses in CRC patients based on a system assessment of apoptosis sensitivity. Based on successfully completed, clinical proof-of-concept work that modelling of apoptosis execution pathways is capable of predicting treatment responses to adjuvant 5-FU-based chemotherapy in CRC patients, the APO-DECIDE consortium has delivered new tools that predict chemotherapy responses in CRC patients based on a system assessment of apoptosis sensitivity. For the future implementation of these tools into clinical trials and clinical decision-making processes, the APO-DECIDE consortium has validated the predictive and prognostic power of these tools in two independent clinical studies. At the completion of this project, the APO-DECIDE consortium has delivered clinically validated sets of combinatorial biomarkers and the associated computational modelling platform that enable the clinical implementation of systems medicine tools for the treatment of CRC, and potentially other cancers.
2. Development of tools that enable a case-specific a priori assessment of treatment responses to FDA- and EMA- approved therapies for the treatment of CRC with a view towards personalised oncology. By integrating EGFR and apoptosis signalling models, the APO-DECIDE consortium has developed new systems medicine tools that enable a case-specific assessment of treatment responses of metastatic CRC (mCRC) patients to anti-EGFR therapy based on the analysis of both EGFR and apoptosis signalling pathways, and has clinically validated the application of these models through a dedicated, multi-centre clinical study. The aim of the APO-DECIDE consortium will be to deliver new systems medicine tools that aid in the identification of mCRC patients with KRAS wild-type status (or ‘quadruple’ wild-type status) who will benefit from anti-EGFR therapy as second line, and potentially first line treatment.
3. Provision of tools for patient stratification in clinical trials that test new treatment regimes In response to the recognised deficiency of tumour cells to undergo apoptosis in response to currently used chemotherapeutics. Most global pharmaceutical R&D companies have developed specific agents that target apoptosis deficiency in tumour cells, most notably Smac mimetics. The APO-DECIDE consortium has recognised these important developments, and has provided novel systems medicine tools that enable a prediction of patient responses to Smac mimetics. APO-DECIDE has performed key pre-clinical experimental validation of such approaches by employing translational in vivo tumour models based on xenografts and CRC patient explants, and by using Smac mimetics that are currently in Phase I clinical trials. Our ultimate goal has been to deliver novel patient stratification tools that enable a prediction of individual patient responses to Smac mimetics and that can be implemented in future Phase 2 or 3 clinical trials, enabling a ‘smart’ clinical trial design.
4. Delivery of new holistic predictive models of apoptosis sensitivity that integrate pharmacokinetics/-dynamics (PK/PD) and whole body physiology, and the validation of these model predictions. One of the key limitations of current systems biology approaches and one of the key challenges of systems medicine is the lack of integration of whole-body physiology and PK/PD data, as well as additional patient-specific information into existing modelling strategies. APO-DECIDE has recognised this important deficiency and addressed it through an integrative multi-scale systems medicine approach that is capable of merging intracellular signalling networks with whole body physiology and patient clinical data within a single modelling framework. APO-DECIDE has performed pre-clinical and clinical validation of this approach in the context of anti-EGFR therapies and Smac mimetics. By doing so, APO-DECIDE has significantly extended the capacity of its systems medicine expertise to conduct knowledge- and data-based virtual trials in order to optimize treatment regimes (dosages/timings) in silico. This allows for the optimisation of future trial design and clinical practice, with the ultimate aim to improve patient care and to provide novel commercial opportunities.
5. Validation of novel quantitative protein profiling techniques that enable the application of systems medicine approaches in routine hospital settings. Current protein profiling approaches in systems biology studies frequently employ quantitative studies performed on freshly frozen tissue material. However, freshly frozen tissue material is not routinely available in the clinical practice, and often of limited quality or quantity. Histological sections based on formalin-fixed, paraffin-embedded (FFPE) tumour material, on the contrary, are routinely constructed within pathology departments, and can be retrospectively investigated. While FFPE sections are increasingly used for genetic testing of tumour material, their implementation as a source for individual protein profiling is as yet underdeveloped. APO-DECIDE therefore views the application of quantitative protein profiling techniques based on FFPE material as one of its core activities. APO-DECIDE has integrated its modelling platforms with reverse phase protein-array (RPPA) and quantitative digital histopathology approaches to provide new systems medicine work flows that can be integrated into clinical trials and future routine hospital settings.
In addition to these Scientific Objectives, APO-DECIDE identified additional Strategic Platform Objectives:
1. Provision of a sample biorepository. APO-DECIDE has created a sample biorepository of a) prospectively collected, primary tumour tissue samples (FFPE) and associated clinical follow-up data from the Northern Ireland NI240 trial of stage 2 and stage 3 colorectal cancer patients treated with 5-FU-based chemotherapy or observation only; b) a retrospective, multi-centre collection of primary tumour tissue samples (FFPE tissue) and associated clinical follow-up data from stage 3 CRC patients treated with 5-FU-based chemotherapy (FOLFOX); and c) a prospective sample biorepository (fresh frozen (FF), FFPE tissue, serum) of stage 4 metastatic CRC KRAS wild-type patients across three centres (PDUM, QUB, RCSI), who had received combination anti-EGFR therapy with tissue/response data collated.
2. Comprehensive molecular and pathological characterisation of primary tumours, including sequencing to identify common gene alterations in CRC, as well as chromosomal alterations.
3. Creation of a fully integrated genome, protein profiling, and clinical database portal that was used to exchange, analyse and integrate emerging data sets.
4. Provision of a systems modelling platform that allows integration of diverse source data and model components into a coherent standardized framework, compliant with the latest emerging international data and modelling standards.
5. Creation of an SME-led commercial development strategy to facilitate timely protection of emerging IPR of primary relevance to participating companies. The SMEs involved in this project have ensured that all commercially viable biomarkers and/or algorithm developments arising from APO-DECIDE have been protected.
6. Optimisation and early development of robust, reliable and standardized tissue combinatorial biomarkers for eventual translation post-project completion into validated companion in vitro diagnostic (IVD) tests, leveraging proven SME-driven strategies.
7. Additional (secondary) IPR in-licensing opportunities provided to SME partners with a view to expanding company product portfolios and diversifying R&D efforts. The foreground of this project has offered each SME partner ample IPR opportunities including biomarkers, systems models and algorithms. Ongoing opportunities to foster advantageous complementarities between projects and partners have been facilitated and encouraged by project networking events and meetings.
WP1 – Biorepository and Sample Distribution
Establishing the APO-DECIDE biorepository
APO-DECIDE undertook the development of a collaborative infrastructure that allowed for the successful transfer and processing of clinical samples for future quantitative experimental analysis and the secure exchange and compilation of detailed clinical data documentation. The efficient and reliable collection and exchange of patient tissue and corresponding clinical data was a crucial enabler for the work and success of APO-DECIDE. In particular, the project’s work utilising tissue microarrays (TMAs) and quantitative analysis of proteins by RPPA in WPs 2, 3 and 4, as well as associated systems modelling work based on these analyses, depended on the accomplishment of the objectives of WP1.
In addition to serving the immediate needs of the project, the provision of this sample biorepository was a key strategic platform objective for APO-DECIDE. The biorepository will be maintained beyond the lifetime of the project, and will be an important element of research infrastructure for ongoing work on the personalised treatment of CRC.
The APO-DECIDE biorepository comprises samples and data from three different cohorts, assembled from three different institutions. Clinical samples (FFPE tumour sections) and data were gathered from FOLFOX cohorts hosted in RCSI, PDUM and QUB and from the Cetuximab cohorts predominantly hosted in PDUM with contributions from RCSI and QUB.
Once the biorepository had been established and successfully populated with samples and data, it was enhanced by the addition of data from other work packages. In particular, data from the TMAs and quantitative analysis of proteins by RPPA carried out in WPs 2, 3 and 4 were added to the database.
Distribution of samples
In addition to gathering clinical data and samples, WP1 was responsible for the distribution of samples throughout the consortium. This involved the careful orchestration of a complex sequence of tissue movements. FOLFOX FFPE patient tissue blocks were successfully distributed from PDUM (n=55) and RCSI (n=28) to QUB for tissue microarray construction and DNA extraction. FOLFOX FFPE patient blocks were then sent to RCSI for protein extraction and the quantitative expression of proteins of interest were analysed using reverse phase protein arrays (WP2). Following TMA construction of the FOLFOX tissue by QUB, the TMAs were sent to ONCO for immunohistochemical staining of proteins of interest and advance digital pathology was carried out (WP3). FOLFOX patient blocks from PDUM were sent back to the Department of Pathology in PDUM with all blocks in excellent condition and accounted for. The return of the FOLFOX patient blocks to PDUM demonstrates the successful implementation of the workflow structures for the distribution and processing of patient tissue.
This workflow was also implemented for the FFPE primary tumour tissue blocks from the PDUM Cetuximab cohort (n=93). These samples were successfully distributed to RCSI for protein extraction and reverse phase protein arrays analysis (WP4). Protein was extracted from all tissue blocks. In addition, protein lysates from primary tumours of 45 patients who received Cetuximab were also sent to RCSI from PDUM. QUB sent 17 FFPE blocks to the Cetuximab collection and RCSI identified 4 patients who were suitable for inclusion in the Cetuximab study. Protein extraction for these FFPE blocks was also performed.
WP1 successfully developed and implemented the necessary workflows for the identification of suitable samples, the transfer and movement of samples and data between partners, and the creation of an appropriate and effective biorepository. These achievements were fundamental to APO-DECIDE, and provided the essential groundwork for the work of the other work packages in the project.
WP2 – Combination of high throughput proteomics (RPPA) with APOPTO-CELL modelling to predict 5-FU responsiveness
WP2 focused on the clinical validation of the APOPTO-CELL systems modelling approach and the development of a calibrated RPPA platform for large-scale protein quantification. RPPAs represent an emerging approach for quantitative profiling of the levels of multiple proteins in tumour samples and have the potential to map protein levels and function in intracellular pathways in a convenient and sensitive manner. RPPAs provide objective quantification using small sample amounts, while being a high-throughput and low-cost technology.
WP2 involved the extraction of protein from biobanked FFPE tissue samples from the NI240 and FOLFOX patient cohorts and case-specific absolute quantification of APOPTO-CELL proteins using RPPA technology. Quantification results were stored in the relational database management system and were annotated with anonymised patient medical information. Through the database, data could be imported into the established systems modelling environments of RCSI and OPTI to provide automated processing, analysis and storage of results. Modelling outputs could then be correlated with clinical outcome. The standard APOPTO-CELL model requires five protein values as inputs, namely the molar amounts of Apaf-1, caspase-9, Smac, XIAP and caspase-3. Due to the costs associated with standardised, calibrated and highly reproducible protein quantification tests in laboratory and clinical environments, RCSI investigated whether it would be possible to reduce the number of individual protein inputs in the differential equations-based APOPTO-CELL model.
Reverse phase protein array [RPPA] preparation and data acquisition
Large-scale protein profiling through RPPA technology is becoming a key technological development to facilitate systems medicine research, enabling a quantitative interrogation of large numbers of individual proteins and phosphoproteins. Expansion of RPPA technology to FFPE-based material is a key development process, and was one of the main undertakings of the APO-DECIDE consortium. RPPA was successfully implemented for analysis of APOPTO-CELL proteins of interest in the NI240 and FOLFOX cohorts and data was subsequently correlated with clinical outcomes. Technologies such as RPPA will be amenable to large-scale clinical trials as well as clinical care centres of all sizes, so APO-DECIDE’s work in this area may have a significant impact on future systems medicine approaches. Although time consuming, macro dissection for RPPA is a relatively simple and beneficial technique to maximise tumour content. A high percentage of tumour is important to ensure accurate and representative analysis of the tumour is performed. RPPA was carried out on protein from macrodissected tumour material and biostatistically analysed. All deliverables of this work package were achieved.
RPPA analysis of tissue samples from the APO-DECIDE biorepository was carried out at RCSI. The first step was the macro dissection of the FFPE blocks. Subsequent to this, protein quantification was carried out, to identify viable samples for RPPA. In the case of tissue samples with a high volume but low concentration, spin columns were used to concentrate the sample to a lower volume with a higher concentration. Concentration of samples using the spin columns was a technique that allowed us to have a lower margin of sample elimination, which was essential given the value of each FFPE block. Once protein had successfully been extracted from all possible FFPE blocks, RPPA analysis was performed. Raw data obtained was distributed to RCSI bioinformatics team for analysis.
The NI240 cohort contained 254 stage 2, 3 and 4 colorectal cancer patients, and the expression of APOPTO-CELL related proteins was successfully determined in 220 patient samples by RPPA, 211 cases of which had sufficient clinical data to compare model output with clinical data. Of these 211 cases, 133 patients had stage 2 colorectal cancer with 64 of these patients having received 5FU-based chemotherapy. A further 78 patients had stage 3 colorectal cancer and 39 of these patients received 5-FU-based chemotherapy. The FOLFOX cohort was composed of 182 stage 3 colorectal cancer patients. The expression of the five proteins of interest to be inputted in APOPTO-CELL could be assessed by RPPA in 135 patient samples (~72 %). Of these 135 samples, 131 belonged to patients who underwent chemotherapy and for whom sufficient clinical data were available to be correlated with the model output.
Correlation of results from RPPA-based APOPTO-CELL modelling with clinical outcome and patient-specific clinical data
The APOPTO-CELL model output was compared with the clinical data on patient outcomes and the model’s ability to predict patient outcome was determined. Apoptosis responsiveness was defined by comparing the activity of apoptosis executing caspases with therapy responsiveness (time to recurrence, time to death, PVS, OS). Correlation was scored to determine the predictive capacity of the approach. Model performance was compared to classical statistical approaches.
Data from the NI240 cohort demonstrates the effectiveness of APOPTO-CELL in accurately identifying stage 3 colorectal cancer patients who will respond to chemotherapy and those that may require additional targeted therapies to achieve meaningful treatment responses (p = 0.059). Substrate cleavage computed by APOPTO-CELL on the basis of the RPPA protein measurements for FOLFOX cohort patients who underwent chemotherapy showed separation between good/bad outcome supporting NI240 results. However these results – which are still preliminary – were below the level of statistical significance (p = 0.1232). Further work, including longer follow up of patients, will be required to definitively determine the importance of this finding.
Modelling the system behaviour of protein interactions is extremely complex. Approximately 0.25 million simulations were executed in APOPTO-CELL to describe the system behaviour for the range of protein concentrations that can reasonably be expected in biological systems. In order to simplify and streamline the process of modelling, RCSI undertook a number of experiments to investigate the possibility of describing a minimum model that would minimise the complexity of the simulations required, while maintaining the usefulness of the results produced.
By examining the comprehensive results obtained from the 250,000 simulations, RCSI investigated the kinetics of apoptosis execution signalling and the associated amounts of cleavage of effector caspase substrates (i.e. the capacity to execute apoptosis). Importantly, these results prove that simulations for time points >300 min are not required. Termination of simulations at 300 min allowed a computationally less expensive approach towards the implementation and further validation of APOPTO-CELL.
The comprehensive analysis of the APOPTO-CELL phase space has allowed RCSI to conduct and evaluate a number of model simplification strategies. As part of this work, we identified that the model can assist in determining that the proteins XIAP, caspase-3 and Smac alone or together should be investigated as a reduced set of combinatorial biomarkers that may be particularly valuable for subgroups of patients in the APO-DECIDE cohorts. Importantly, the systems modelling approach has provided specific concentration ranges for these proteins at which they will have dramatic effects on the capacity to execute apoptosis, irrespective of the concentration of the other proteins involved in the signalling network. This may provide additional and attractive routes for exploitation and potential commercialisation.
Development of a clinical systems medicine workflow solution
WP2 also developed a workflow which integrates APOPTO-CELL into a clinical environment, where it can improve treatment of patients as a means predicting the treatment outcome. Integrating the APOPTO-CELL software into a clinical setting required the provision of a simple graphical user interface (GUI), which would make the functionality of the modelling system easily accessible for clinicians who are not trained in using complex systems modelling software. RCSI defined the requirements for this GUI in consultation with clinicians and pathologists at QUB and RCSI Beaumont Hospital. Based on these requirements, a prototype implementation of the GUI was produced. This was further refined following extensive testing.
This GUI and the workflow developed by WP2 demonstrate the possibility of integrating a sophisticated systems modelling package into a clinical workflow, making it a routine element in treatment choice. This promises improved outcomes for patients with CRC, and offers opportunities for commercialisation of APOPTO-CELL.
WP3 – Combination of APOPTO-CELL with advanced digital pathology for implementation into clinical settings
WP3 focused on another important aspect of integrating systems modelling into clinical settings. Previous studies have confirmed the usefulness of systems modelling of apoptosis as a predictor of patient response, but they were performed using quantitative Western blotting, which is not easily transferrable into a clinical setting. In order to maximise the efficient and effectiveness of APOPTO-CELL as a routine part of a clinical decision-making workflow, APO-DECIDE implemented a number of important new techniques in the area of pathology. The objective of these changes was to make input data (and the model system itself) more amenable to the hospital pathology laboratory, by including both standard (i.e. immunohistochemistry) and advanced histology techniques (i.e. tissue microarrays and digital image analysis). Over the last few years, there have been several advances in histological techniques, including tissue TMA technology, which allows for the simultaneous analysis of several hundred clinical specimens, while retaining morphological features. For this project, QUB created TMAs for the CRC tumour samples from the NI240 cohort and the FOLFOX cohort, which were used by ONCO to examine the APOPTO-CELL relevant proteins using conventional immunohistochemistry (IHC).
The second key development in the histology field is digital pathology, whereby glass slides are digitised using a state-of the art scanning system to generate high-resolution images (digital slides). Once a digital slide is created, this can be stored, viewed via the web, annotated, and crucially examined with image analysis software. This software has several advantages, including providing quantitative data for each biomarker of interest. As part of WP3, ONCO and RCSI made use of this quantitative data in conjunction with APOPTO-CELL to predict patient response and to provide the basis for the future development of a companion diagnostic test for use in personalisation of cancer therapy.
Construction of TMA from tissue samples
Two sets of TMAs were prepared using the NI-240 and FOLFOX cohorts. In the case of the NI240 cohort, 402 blocks from 239 cases of CRC were supplied from the biorepository (WP1) to QUB. Fresh hematoxylin and eosin (H&E) stains were performed for cases using new sections from these blocks. All H&E stained slides were reviewed by pathologists at QUB. A new aspect was introduced into the design of the new TMAs, in the inclusion of cell line controls for each antibody to be examined. The same batch of cell lines that was used to construct the TMAs were also used for the antibody validation and protein quantification studies. As a result, these cell lines serve as a cross-referencing system for target protein concentration across experiments and between technology platforms.
Protein expression analysis via immunohistochemistry
Western blot (WB) analysis of cell line samples was carried out by ONCO. For this purpose, Hela-D98 cells were used as a positive control. For negative controls, we used cell lines where the target antigen of interest was selectively knocked down or was naturally deficient (e.g. MCF-7 cells are deficient in Caspase-3). Frozen vials of these cell lines were collected from our collaborators, and the cells were cultured in ONCOs tissue culture facility. Target antibodies from different vendors were selected based on the available information provided in the data sheet by the company. These antibodies were subjected to SDS-PAGE and WB analysis in order to confirm their specificity.
Antibody validation by IHC
As an antibody recognises a denatured linearized protein in the context of WB analysis, it is known that the same antibody might not recognise the protein in its native conformation in IHC. Hence, a second line of validation is to test whether the antibody can recognise the native protein by performing IHC. ONCO constructed FFPE cell pellet arrays from the cell lines used in the WB experiment (notably using an equivalent collection of cells cultured alongside) and performed IHC on these materials. All five antibodies we selected performed well in IHC using different antigen retrieval buffers at varying pH.
Using image analysis, we calculated the H-scores to clearly demonstrate the differences in staining intensity of positive and negative control cell lines.
Automated image analysis of TMA data
To complement the potential for high throughput analysis offered by TMA, ONCO adapted its image analysis software (IHC-MARK) specifically for the APODECIDE project in order to create a high throughput image analysis platform to minimise time/inter-observer variability. The process relied on the high-resolution digital capture of TMA slides. The resulting images were analysed to obtain quantitative IHC data for each protein. The images were categorised into four different categories (Negative, Mild, Moderate and Strong) based on the staining intensity.
The image analysis software (IHC-MARK) was modified to quantitatively analyse the positive (protein) staining. The sensitivity and specificity of this software was compared to another commercial image analysis software (and manual scoring, where available) to validate the performance of the algorithm. This analysis provided us with numerical data on eight different categories for each core within individual TMAs. The most relevant outputs of the image analysis are: a) average positive intensity of staining, b) percentage total positive staining, and c) total stained area, and d) H-score. For quality control, manual assessment of representative slides (manual histological scoring) was carried by the consortium pathologist, Prof. E. Kay, at RCSI. The quantitative image analysis data produced was used in the APOPTO-CELL systems model, and all image data were deposited to the APO-DECIDE database.
‘Average positive intensity’ and ‘percent positivity’ values were obtained from the image analysis data and were used for the statistical data analysis. In case of CD8 and CD45, the total number of positive cells per core was used as a final output of image analysis.
WP4 – Integration of EGFR/apoptosis signalling pathways and molecular tumour data to predict responsiveness
WP4 focused on the identification of biomarkers in patients treated by anti-EGFR therapy associated with different chemotherapy regimen in advanced colorectal cancer. The APO-DECIDE partners prospectively collected primary tumour samples from metastatic colorectal cancer patients with wild-type KRAS status treated with anti-EGFR therapy (Cetuximab) and 5-FU-based regimen (FOLFIRI or FOLFOX). The search for new biomarkers has been done at different molecular levels, on DNA, RNA and protein. WP4 relates to the characterization of tumours by sequencing and by search for copy number variation using comparative genomic hybridization. Expression of mir31-3p, which has been implicated in response to anti-EGFR therapy, was also determined. Protein profiling of apoptosis signalling and EGFR signalling proteins by RPPA have been performed. Correlations between protein and microRNA profiling, molecular characterizations and response to treatment has been implemented in the database and subsequently integrated into the higher-level apoptosis/EGFR signalling model.
Molecular characterization of the primary tumour
PDUM carried out a detailed molecular characterization of all primary tumours collected in WP1, through sequencing of common genes altered in CRC, and characterising chromosomal alterations. Specific genomic regions of the stable tumours identified were then re-sequenced. This sequencing focusing on the 46 genes contained in the AmpliseqTM panel (Ion Torrent). Among the panel the following genes were be analyzed: KRAS, BRAF, TP53, NRAS, EGFR, PIK3CA, SRC, CDKN2A, SMAD4, GNAS, PTEN, STK11, MET, RB1, AKT1, FBXW7, CTNNB1, APC, MLH1 FGFR1, FGFR2, FGFR3.
The sequence analysis was completed by profiling chromosomal aberrations, such as amplifications, deletions, copy number changes, and copy-neutral loss of heterozygosity (LOH) events, by genotyping tumour samples. This was carried out using DNA extracted from frozen samples. The characterization of copy number variation was performed with Genome Alteration Print (GAP) software, which allows the visualization of complex cancer genomic profiles obtained by SNP arrays. The presence of activating or inactivating mutations and aberrations for the members of EGFR and/or apoptosis related signalling was reported and implemented in the relational database system.
According to criteria previously defined, we found one mutation in 53 tumours, two in 38 tumours and three in 25 tumours. In five cases, we identified four or more mutations. In 29 samples, we did not identify any alteration. The most frequent mutated gene was TP53, followed by APC and SMAD4. Interestingly, three patients from the PDUM cohort showed a mutation of KRAS gene, which have been missed by our initial screening. Similarly, three patients from QUB showed KRAS mutations, as well as one patient from RCSI that was previously detected in RCSI screening and that we could confirm. We identified a NRAS mutation in four patients, and a BRAF mutation in 10 samples. KRAS- and NRAS-mutant tumours receive little or no benefit from anti-EGFR therapies as single agents or combined with chemotherapy and the European Medicine Agency (EMEA) and the Food and Drug Administration has restricted the use of cetuximab and panitumumab to patients with wild-type KRAS and NRAS tumours. For this reason, the 11 patients those tumours carry either KRAS or NRAS mutations were analysed separately from the rest of the series.
Quantification of miR31-3p expression has been performed on 99 samples from PDUM, 19 from QUB and five from RCSI. Analysis showed an inverse correlation between miR31-3p expression and overall survival and progression free survival.
WP5 – SMAC mimetics
SMAC mimetics are a promising class of drugs that directly interfere with the apoptosis execution pathway by antagonising inhibitor of apoptosis proteins (IAPs) and are currently being explored in several clinical trials. Based on a prototype systems biology model for SMAC mimetics, the approach adopted in WP5 was to extend the core APOPTO-CELL systems model to reflect the consequences of TL32711 on apoptosis signal transduction in order to derive a new predictive algorithm (APOPTO-CELL SMAC) for SMAC mimetics in colorectal cancer. In vitro and in vivo experiments were carried out by QUB, RCSI and UHF. The integration of this data into higher-level models was performed by OPTI.
In vitro validation of APOPTO-CELL SMAC
QUB performed a number of cell viability and cell death in two model colorectal cancer cell lines: HCT116 and LoVo. In addition, Western blot analysis of expression of core APOPTO-CELL proteins (cyt-c, SMAC, Apaf-1, caspase-9, caspase-3, XIAP) and additional proteins which may be pathway relevant for APOPTO-CELL SMAC extensions (cIAP1, cIAP2, FLIP, caspase-8) was carried out as well as flow cytometric assessment of cell surface expression of cell surface receptor tumour necrosis factor receptor 1 (TNFR1). A number of HCT116 daughter cell lines were used to assess the impact of clinically relevant genetic contexts on response to SMAC mimetics, namely: p53, mismatch repair (hMLH1) and KRas.
QUB also generated cell lines over-expressing both main splice forms of FLIP (FLIP(L) and FLIP(S)), which were assessed for sensitivity to SMAC-based therapy alongside cells in which caspase-8 was depleted using siRNA approaches. In addition, QUB developed an immunoprecipitation protocol to determine the formation of apoptosis-inducing cytoplasmic complexes between caspase-8 and RIPK1 following SMAC mimetic treatment; formation of these complexes are negatively regulated by IAPs. The role of TNF-α was assessed in a number of ways:
• Addition of exogenous TNF-α to cell line cultures;
• Assessment of secreted levels of TNF-α using ELISAs;
• Use of neutralising anti-TNF-α antibodies to disrupt ligand binding to TNFR1.
RCSI used their published methodology to monitor caspase activation in single cells. HCT116 and LoVo cells were transfected to express a FRET-based reporter for caspase activation, and caspase substrate cleavage was measured by flow cytometry. Key data for the single cell understanding of signal transduction kinetics in HCT116 cells following 5FU/oxaliplatin treatment, alone or in combination with SMAC mimetic TL32711 were already described. The data acquisition was continued and expanded into the LoVo cell system setting.
In these experiments, we unexpectedly identified that CRC cells frequently display caspase activation prior to the initiation of the canonical apoptosis execution phase. These events, which undoubtedly are central to understanding the establishment of cell death responses upon SMAC mimetic TL32711 (co-) treatment, are not captured by the core APOPTO-CELL SMAC systems model. Therefore, we had to expand the imaging studies to collect data from a sufficient number of cells to understand the frequency with which these events occur and the kinetics by which these events are propagated to lead into the classical execution phase.
Studies on cell death and apoptosis
To assess the impact of chemotherapy on response of colorectal cancer cells to TL32711, QUB determined the effects of 5-FU and oxaliplatin (alone and in combination to mimic the clinically relevant FOLFOX regimen) on apoptosis induced by TL32711. It was found that co-treatment with TL32711 increased the levels of apoptosis induced by 5-FU alone, oxaliplatin alone or 5-FU+oxaliplatin. Similar results were obtained with SN38, the active metabolite of irinotecan, another chemotherapeutic used to treat colorectal cancer. Notably, treatment with a TNFα neutralising antibody failed to prevent the enhanced apoptosis observed in TL32711/chemotherapy co-treated cells indicating that TNFR1 ligation is not required for this effect. Moreover, chemotherapy treatment (such as TL32711 treatment) failed to significantly induce secretion of TNFα from either colorectal cancer model. However, similar to co-treatment with TNFα chemotherapy treatment enhanced formation of the apoptosis-inducing complex between RIPK1 and caspase-8, suggesting that formation of this complex (termed the ripoptosome) is responsible for the synergy between chemotherapy and TL32711.
QUB performed in vivo experiments using xenografts derived from HCT116 and LoVo cells inoculated subcutaneously into the flanks of Balb/c nude mice. As well as monitoring tumour growth, a subset of animals was sacrificed during treatment to assess pharmacodynamic biomarkers of drug response (cIAP1 expression). To complement these experiments, UHF carried out experiments using the Chorion allantois membrane (CAM) mode. In these experiments, LoVo cells were implanted on fertilized chicken eggs on day eight of incubation and treated with TL32711 and/or 5-FU for three days. Thereafter, tumours were sampled with the surrounding CAM. The tumour area was analyzed by immunohistochemistry and images were digitally recorded. Tumours turned out to be completely resistant to treatment with 5-FU alone. Exposure to 5-FU even slightly increased tumour growth compared to control tumours. Notably, the addition of TL32711 prevented this 5-FU-stimulated increase in tumour area. By comparison, no evidence of cooperative antitumour activity of TL32711 and 5-FU was found at the concentrations used.
Protein expression for model parameterisation
To enable parameterisation of the APOPTO-CELL-SMAC model, baseline protein expression of core APOPTO-CELL proteins (cytochrome c, SMAC, XIAP, caspase-3, caspase-9 and APAF-1) was assessed by Western blot in samples obtained from untreated LoVo and HCT116 xenografts by QUB and RCSI. These samples were assessed alongside protein lysates processed from contemporary tissue cultures of each cell line, with absolute protein quantification carried out using a reference sample from HeLa cells in which levels of each APOPTO-CELL protein are known. The in vitro and in vivo protein concentrations (mean values and standard deviations) were used to parameterise the APOPTO-CELL model.
APOPTO-CELL SMAC model adaptation and expansion.
The experimental evidence described above was used to adapt the APOPTO-CELL SMAC model. First, we employed the original model and compared model results for protein concentrations for HCT116 and LoVo cell conditions that were determined in vitro and in vivo. We took into account the experimental variance of the measured protein expressions to obtain simulation results that reflect heterogeneity within the cell populations. This was implemented by randomly sampling 1000 combinations of protein concentrations within the experimentally determined range to simulate an ensemble of 1000 individual cells. These 1000 virtual cells will reflect an in silico tumour.
APOPTO-CELL SMAC was then expanded to include the newly identified processes upstream of the canonical apoptosis execution phase into the model. Our results suggest that ripoptosome formation results in pre-MOMP caspase activation, leading to ramps of substrate cleavage prior to initiation of APOPTO-CELL SMAC. The model extension took into account that the resulting ‘ramps’ of substrate cleavage can occur with specifiable frequencies and that these ramps deplete substrate until a critical threshold is reached that initiates the execution phase. The execution phase then proceeds as in the previous APOPTO-CELL SMAC version. In addition, the administration of TL-32711 was modelled to increase the frequency of ramp occurrence. Initial simulation results for LoVo cell conditions demonstrated that death kinetics is strongly affected by TL-32711. The simulations are currently being adapted to take into account in vivo HCT116 and LoVo protein concentrations, experimental kinetic data obtained by time-lapse imaging as well as experimentally determined frequencies of ramp occurrence. Together this will allow us to identify how many of the 1000 in silico cells are capable of executing apoptosis in presence or absence of TL-32711. The number of dying and surviving in silico cells will then be compared to tumour growth rates measured in vivo. This will validate the predictive capacity of the adapted APOPTO-CELL SMAC model.
To enable a full efficacy study, two PDX tumours are currently being grown in a small cohort of NOD/SCID mice to provide tumour material to implant two full experimental cohort of mice in the coming weeks. Subsequently a full efficacy experiment including 18F-FDG-PET imaging and ex vivo analysis of the effect of treatment of TL-32711 alone and in combination with 5-FU + oxaliplatin will be completed.
The results of WP5 identify an alternative therapeutic approach by which colorectal cancer patients predicted by the basic APOPTO-CELL model to be resistant to standard 5-FU-based chemotherapy could be treated with a therapeutic combination regimen containing a SMAC mimetic. Patients would be selected for such a combination using a predictive algorithm derived from the prototype APOPTO-CELL-SMAC model, which is still being refined.
The information generated will be used beyond the lifespan of APO-DECIDE to continue the development of personalised medicine approaches for the clinical application of SMAC mimetics in metastatic and high-risk early stage colorectal cancer.
WP6 – Higher level integration of APOPTO-CELL systems models
APO-DECIDE’s research was based on a range of pre-existing data, and generated a considerable amount of additional data. The aim of WP6 was to integrate the data collected in other work packages and outputs of the extended APOPTO-CELL model within the high-level virtual patient platform developed by OPTI. The cross-scale integration of modelling platforms enabled the most efficient processing of the data collected and generated by the partners in the different WPs. Together with the other parts of the APODECIDE project (such as establishment and integration of the database, WP7), WP6 provided a framework and an environment for applying the range of computational models used and developed by the project to the data collected from the clinical reports and experiments. Developing a comprehensive, end-to-end workflow starting from the experimental and clinical data records, through a specifically developed database, to the relevant components of the mathematical and computational models, and eventually to output reports that present the model-generated predictions in response to the clinician’s requests was crucial for the optimal utilization of the project results, and should facilitate the application and integration of APO-DECIDE’s findings in clinical settings.
In the course of the APO-DECIDE project, the partners have collected and generated clinical and experimental data in four cohorts (see WP 1). All these datasets were analysed in WP6, in order to develop predictive models that use the different types of pre-treatment data collected for the specific cohort as an input. In addition to this pre-treatment data, the models used the outputs of the APOPTO-CELL model, and predict the various response endpoints. The predictive models were trained using the collected data, and, whenever applicable, were retrospectively validated.
Higher level integration and validation – APOPTO-CELL RPPA/IHC
WP6 undertook the development and application of the integrated computational modelling platform for predicting the efficacy of adjuvant treatments, such as to the one applied in the NI240 and FOLFOX patient cohorts (i.e. 5FU-based chemotherapy). This platform integrated the laboratory and clinical data and the predictions of the APOPTO-CELL model into a single environment, connected with predictive models based on Optimata technology. This integrated system provided a base to which further expansions were added subsequently added.
To design this platform, OPTI employed a data-driven approach, by which the modelling technology is chosen based on the specific type and extent of the data available for model training and testing. To achieve this, OPTI worked with the other partners to carefully define and describe the available data for each clinical cohort. Based on this, appropriate measures for response evaluation were defines for each cohort. In order to assure comparability across datasets, it was necessary to develop and apply a special methodology for predicting survival-related end points, relevant to the adjuvant treatment. This work required the expansion of existing modelling platform to include more advanced multivariate survival models and machine learning algorithms.
Generally, the datasets of the NI240 and FOLFOX cohorts include for each patient the following information: (i) individual recurrence date/survival time/time of the last examination, (ii) measures of individual covariates (i.e. standard clinical evaluations and genetic and protein data from normal and tumour tissue) taken pre-treatment and (iii) the treatment protocol. For each patient, the APOPTO-CELL model output is also available, providing prediction of cell-level response of the apoptotic pathway to treatment on the cellular level. OPTI has used those data to develop and train tools for predicting patient-specific recurrence and survival from her/his individual pre-treatment covariates.
Once the available data for each cohort had been determined, a workflow for development of such predictive tools was outlined. This involved:
(I) Data preparation and preliminary analysis. During this stage, the lists of potential predictive covariates and of response variables to be predicted were defined, including the range of possible values per each variable. Statistical analysis was applied in order to determine possible dependences and influences within the covariates and response data (e.g. correlations, collinearity). This stage concluded with final choice of data to be used for model training.
(II) Univariate analysis. Each response variable was analysed vs. each candidate covariate, using appropriate models (e.g. ANOVA, Kaplan-Meier (KM) survival curves, linear regression). This stage resulted in ranking of most potentially significant covariates, per each response variable.
(III) Multivariate analysis and model training. For each response variable, a subset of covariates used for its prediction and the best-performing prediction algorithm was selected. This was a multi-step process, which included application of various methods of feature transformation and selection (e.g. Principle Component Analysis (PCA), stepwise selection algorithms) and of several statistical or machine learning methods. For the categorical response prediction (‘yes/no’ type), those could be logistic regression, various classification methods such as neural nets (NN), nearest neighbours (kNN), support vector machines (SVM), etc. For numerical response (such as time to progression) those could be regression or survival models, e.g. Cox, Weibull, log-logistic, which are used to generate most likely estimation of time.
(IV) Validation. When required and feasible, predictive tools should be prospectively validated by new independent data. If no new data are available, it is possible to divide the dataset at hand into training and validation parts, in an unbiased way, select the model using the training set and validate it on the other set. This approach requires the whole available dataset to be large enough, which is not the case presently. Therefore, we used cross-validation (CV), as follows. The whole dataset was randomly divided into k parts (folds), and the CV proceeded iteratively in k stages. At each stage, one fold of data was omitted and all the rest were used to train the model, which was then tested on the omitted part. In this way, the training-validation process was repeated and all the data were used for both training and independent retrospective validation. This method is denoted ‘k-fold CV’. A special case, when k equals the number of points in the dataset (patients, in our case), and one patient is omitted at each iteration, is termed ‘Leave-one-out’ (LOO) CV.
This methodology was initially applied to the NI240 cohort. In this context, our analysis consistently determined several factors that are predictive of Recurrence-Free Survival (RFS), both when clustering the patients, and when considering the exact value of RFS. Patients that are untreated, or treated by 5-FU can be classified into responders and non-responders (RFS<26.6months or RFS>26.6 months) with high accuracy, and both the clustering and the numerical value of RFS can be predicted with high accuracy. We found that most of the collected data are required to generate accurate predictions, and proper pre-processing (such as Principle Component Analysis) helps to avoid numerical problems.
Following the successful analysis of the NI240 cohort, the same methodology was applied to the FOLFOX cohort. This cohort contains data of 203 CRC patients followed up for 5-8 years after resection of primary tumour, treated by FOLFOX or XELOX. Survival data were recorded for most patients, Time To Recurrence (TTR) was given for 195 patients, with 143 patients (73%) censored at the time of death or last follow up; Overall Survival (OS) is given for 195 patients, with 170 patients (87%) censored at the last follow up; RFS times were almost similar to TTR, with exception of 4 patients that died before recurrence. Pre-treatment covariates included standard clinical evaluations, blood tests, several key mutations and levels of apoptosis-related proteins measured by RCSI (RPPA-based data) and ONCO (TMA-based data).
We chose to focus on predicting the TTR and OS, due to redundancy of RFS data. For both variables, clustering analysis revealed statistically significant division into two clusters (‘short’ vs. ‘long’ time), with thresholds of 47 months for TTR and 74 months for OS. Applying correlation analysis, we found that there are many moderately strong correlations between the covariates - notably, protein levels are correlated. Further, we have found that several covariates can be significant for predicting TTR or OS, but no strong associations were found. We applied the multi-step procedures for dimension reduction and model selection in order to train models that predict TTR and OS cluster (0 or 1). We found that using the whole dataset, the logistic regression models can be trained to separate perfectly the two clusters in both cases.
The perfect separation achieved when training both cluster predictors suggests that there can be a case of model over-fitting, especially taking into account the large number of potential covariates and the large number of multivariate models tested during the steps of model selection. In order to check the reliability of the claim that the evaluated covariates can predict chances of recurrence or death, one would need to test those on the new independent set of data, i.e. perform prospective validation. As explained above, since such data are not yet available, we performed cross-validation (CV), using the same dataset. By applying LOO CV we have found that the model for predicting TTR cluster has specificity of 0.85 and sensitivity of 0.81 and the accuracy rate was 83%; the model for predicting OS cluster has sensitivity of 0.96 and specificity of 0.79 and the accuracy rate was 93%. On the other hand, the models for predicting actual numerical value of TTR or OS were less accurate, and probably require more non-censored data to develop reliable predictions.
Higher level integration and validation – APOPTO-CELL EGFR
Having developed the basic modelling platform, we continued to develop it by incorporating data for the Cetuximab cohort, featuring patients treated by anti-EGFR therapy against advanced metastatic CRC. The methodology outlined above was applied to this cohort. Statistical/machine-learning and survival models were trained to predict response and time to progression or survival for individual patients using their pre-treatment data. The general workflow in this task was similar to that described above.
The dataset of the Cetuximab cohort includes for each patient: (i) the individual records of response, namely, Best Response in terms of Response Evaluation Criteria In Solid Tumours (RECIST), Progression-Free Survival (PFS) and OS; (ii) individual clinical data collected pre-treatment; (iii) phosphoprotein array measurements for 9 proteins; (iv) RPPA measurements of 32 apoptosis and EGFR pathway-related proteins; (v) evaluation of mutation status of 22 genes and level of microRNA (miRNA), miR-31-3p.
Best Response, PFS and OS were chosen as potential as response variables. Further, Best Response was represented by two clinically relevant binary variables, Objective Response (OR; is 0 when Best Response=CR or PR, and is 1 when Best Response=SD or PD) and Disease Control (DC; is 0 when Best Response=CR, PR or SD, and is 1 when Best Response= PD). The potential input for the predictive models was defined as all the data collected during the project, i.e. (ii), (iv) and (v) above, namely, the standard clinical data, RPPA protein measurements, gene mutations and miRNA levels. In addition, we have added the ‘Substrate’ output of the APOPTO-CELL EGFR model as an input. These data together include 59 factors (covariates), of which we selected a subset to be used as an input for each of the four models developed for individual prediction of the four response variables described above (OR, DC, PFS and OS).
Of the 160 patients contained in Cetuximab cohort, 92 had enough records of covariates data for model training and validation. OR and DC were available for 88 of these 92 patients; PFS was available for 88 patients, with 58 patients censored at the last follow-up; OS was available for 89 patients, with 55 patients censored at the last follow-up. Preliminary analysis has indicated that the dataset has high multi-collinearity among the covariates (especially the protein levels) and dimension reduction might be required. Several covariates had high correlations with the response variables, stressing the need for multivariate analysis.
Results were validated using two methods: LOO CV and 10-fold CV, the latter being a more challenging test to model accuracy and robustness. For LOO CV, we found that the sensitivity in predicting DC was 0.71 and specificity was 0.88; in total, in 85% cases, the patients were predicted correctly. For OR, the sensitivity was 0.77 and specificity was 0.75 with 76% of patients classified correctly. When applying 10-fold validation, we have obtained essentially similar results, suggesting that the estimation of model performance is robust: for the prediction of OR we obtained both sensitivity and specificity of 0.73 with accuracy rate of 72.7%; for the prediction of DC, sensitivity was 0.71 and specificity was 0.89 with 86.3% accuracy.
The results achieved by WP6 suggest that the collected data contains important information that may be used before treatment, in order to predict the chances of success, both in adjuvant and advanced CRC. These findings pave the way for development of practical tools that will help physicians select better treatment options, by comparing individual patients’ likelihood of response. Retrospective cross-validation shows that the binary-type (‘yes/no’) predictions are quite reliable. The quantitative predictions, being of much higher resolution can be less reliable, and additional patient data are needed to substantiate the predictive models.
Higher level integration and validation – APOPTO-CELL SMAC
The modelling platform was then further developed to provide an integrated higher-level model for the effect of Smac mimetics, for improvement of standard chemotherapy treatment of CRC. WP6 used the in vitro and in vivo experimental results generated by the partners in WP5 to develop the pharmacokinetic (PK)/ pharmacodynamics (PD)/disease model of effect of co-treatment by novel Smac mimetic TL32711 (also known as Birinapant) and FOLFOX in CRC. In order to achieve this, we applied the OVP modelling platform, which allows simulation and parameter estimation of complex models of drug PK/PD and disease progression and response to treatment, both on population and individual levels.
The experimental data generated by WP5 included the effect of Birinapant combined with 5FU and Oxaliplatin, in vitro and in vivo, in two colorectal carcinoma cell lines: LoVo and HCT-116. The in vitro data included cell viability measurements under various concentrations of the three drugs (alone and in combination). The in vivo experiments were carried out in mice xenografts of the two cell lines, where tumour sizes were measured over time.
We have trained PD models for all the three drugs using the results of in vivo experiments. Using the in vivo experimental results, in conjunction with published data regarding PK of the three drugs in mice, we trained the full PK/PD/disease model to reproduce the dynamics of tumour size over time, for both cell lines in mice xenografts.
By modelling the experimental data, we obtained four sets of PD model parameters, each one specific to experimental setup: in vitro/in vivo, HCT-116/LoVo cells. For each of the four setups, six protein concentrations were measured by RPPA technique, in WP5. We investigated the relationships between the protein levels and corresponding parameter values and found that for each of the six variable PD parameters, at least one of the proteins had correlation coefficient above the significance level. We constructed regression models, using the most correlated protein as a predictor factor (covariate), separately for each PD parameter.
This PK/PD/disease model was then upscaled to humans and used to predict the clinical efficacy of Birinapant, combined with FOLFOX in advanced CRC patients. This involved carrying out a clinical trial simulation and predicting the distribution of PFS times in a population of virtual patients. This approach demonstrates that pre-clinical data may be used for developing models for predicting the expected clinical benefit of the new treatment, even when no clinical data are available. In addition, the response to treatment can also be predicted on the individual level, when concentrations of the relevant proteins (or other equivalent biomarkers) can be evaluated for the specific patient. The methodologies developed as part of WP6 have clear potential to lead to improvement in personalized medicine and drug development.
WP7 – Database & Exchange Portal
APO-DECIDE relied heavily on data exchange between the academic and non-academic partners. Therefore, a major objective in APO-DECIDE was to establish a data framework for efficient data transfer between the partners and to provide a data alignment/standardisation concept that allowed pooling, comparing, and structuring data, to facilitate data queries, transfer, and joint analyses in the context of systems-level research and modelling.
To provide a reliable system that allowed the exchange, storage and access of the data for all partners, WP7 developed a centralized data exchange platform. This included the construction and programming of a database system that stores clinical data as well as structured quantitative data from all partners to allow the data entry and access of the annotated data. The performance of the database environment was monitored, taking into account the aspects of data collection, entry, annotation, and distribution.
Project co-ordinator RCSI developed and hosted the central project database. This database was fed with clinical, experimental, and modelling data from all partners involved. An associated data integration strategy, which allows data sets from multiple partners to be merged, was developed. In parallel, partners RCSI and OPTI ensured that their modelling environments (based on MatLab or OVP) could be linked to the database portal in order to semi-automatically retrieve data sets for modelling tasks and to feedback and store modelling outputs in the database.
Once the database structure had been developed, the database was populated with clinical data and experimental data from the partners involved. In this context, the project database environment was further developed to establish a dedicated server infrastructure for the project and to generate an associated internal data portal. The performance of the database environment and the data flow was continuously and retrospectively assessed, including the performance of the data collection, annotation, and distribution procedures. In this context, RCSI also evaluated the associated workflows and highlighted aspects. This work should ensure smoother data collection and implementation procedures for subsequent studies and additional collaborations that go beyond the duration of APO-DECIDE.
The aim of the APODECIDE consortium was to develop systems medicine tools that predict treatment responses in colorectal cancer (CRC) patients to 5-FU based chemotherapy and anti-EGFR therapy, based on a systems analysis of apoptosis and EGFR signalling pathways. The consortium’s goal was to develop new clinical decision making tools that enable personalised medicine approaches and ‘smart’ clinical trials design in the future.
Colorectal cancer is the third most common cancer worldwide. In 2014, it is estimated that 136,830 patients (71,830 men and 65,000 women) will be diagnosed with, and 50,310 patients will die from, cancer of the colon or rectum in the United States . Colorectal cancer was the second most common cancer in Europe in 2012. In addition, it was estimated that 447,000 patients (242,000 men and 205,000 women) were diagnosed with and 215,000 patients died of colorectal cancer in Europe during 2012 .
The primary treatment of early colorectal cancer is surgical resection of the primary tumour and regional lymph nodes. Surgery is curative in early stage tumours but in more advanced stages, adjuvant therapy is recommended to prevent recurrence and improve survival. Stage 4 disease will require systemic chemotherapy. 5-FU is the cornerstone of treatment for patients with early stage (stage 2/3) and advanced (stage 4) CRC. More than half of early stage CRC patients will develop local recurrence or distant metastases, despite adjuvant 5-FU based treatment. In stage 4 disease, half of patients do not derive any objective benefit from 5-FU based chemotherapy. Hence, identification of patients that do not benefit from 5-FU based therapies may lead to reduction in the cost of treating these patients and may spare unresponsive patients unnecessary treatment toxicities. The goal of adjuvant 5-FU based chemotherapy treatment is cure (or to increase chance for cure); there are no other chemotherapy treatment strategies available in this disease setting.
Improved Patient Care
The APO-DECIDE project uses new predictive tools that will provide guidance in deciding whether or not a colorectal cancer patient will benefit from combination chemotherapy. Key to APO-DECIDE is the development and validation of systems-based biomarkers for apoptosis susceptibility, combining high-throughput proteomics (reverse-phase protein arrays (RPPA)) with APOPTO-CELL novel systems modelling to predict responsiveness to DNA damage, and combining APOPTO-CELL with advanced digital pathology for use in clinical settings. Other technologies within the project, such as tissue microarray (TMA) construction, protein expression analysis via immunohistochemistry (IHC), automated image analysis of TMA data, the correlation of results from IHC-based APOPTO-CELL modelling, and the integration of IHC quantification and systems modelling, can be used to develop a full clinical workflow.
Our novel combinatorial biomarkers and associated systems modelling framework will allow stratification of patients based on systems analysis of apoptosis and EGFR signalling. In particular, it will allow the identification of those CRC patients who will (or will not) benefit from 5-FU-based chemotherapy and anti-EGFR therapy. Only a very small portion of node-negative Stage 2 CRC patients will benefit from chemotherapy, however there are no known effective markers to readily identify this subset of patients. Adjuvant 5-FU/oxaliplatin treatment in node-positive stage III disease benefits only ∼15%-20% of patients, and 30% of these will relapse within 5 years following surgery . Approximately 50% of metastatic, stage IV CRC patients do not benefit from the addition of anti-EGFR despite the fact that their tumour is KRAS wild-type. The identification of other genetic biomarkers that determine anti-EGFR therapy resistance have so far been largely unsuccessful in terms of their clinical implementation, and even patients with a ‘quadruple-wild-type’ tumour status show only maximal response rates between 40-50% .
The system modelling results we obtained could allow clinicians to classify patients based on systems analysis of apoptosis and EGFR signalling, coupled with their interplay with other biomarkers for drug responsiveness, such as somatic gene mutations. This will allow clinical oncologists to improve their clinical decision making, and to provide a framework in which specific treatment regimens are implemented as second-line, or in fact first-line treatments, which will ultimately benefit the patient.
Furthermore, possible solutions for identifying if and how highly promising novel targeted therapeutics, such as Smac mimetics (currently in several Phase 0/I trials), will be for patients was developed. Such patients may already display resistance to classical genotoxic therapy, or to novel ‘personalized’ therapies including anti-EGFR therapy or anti-VEGF therapy, but may be sensitive to Smac mimetics, either as second-line, or indeed as first-line treatment. Overall, the approach taken by APO-DECIDE will provide novel personalised medicine solutions, based on new predictive knowledge of how patients will respond to specific combinations of therapies that have the potential to greatly improve patient care.
Importantly, as apoptosis deficiency is a major contributor to cancer progression and cancer resistance, our achievements can easily be translated and applied to other major types of cancer as well. While the APO-DECIDE project focused on CRC, our technologies and approach will be applicable to a broad spectrum of cancers with associated increased health economic impact. These include the ‘big four’ cancers, but also cancers that are characterized by a high, intrinsic resistance to current chemotherapy, such as pancreatic cancer and glioblastoma. Our development and validation of systems-based biomarkers for apoptosis susceptibility will have significant impact on systems medicine research by identifying patients who will respond to standard chemotherapy and also by identifying if and how novel targeted therapeutics may be used at maximal benefit for patients.
Improved Clinical Trial Efficiency
The impact of the modelling environment is not limited to acute decision making for personalised medicine. The model may have further impact through the capacity to evaluate novel drugs and treatment schedules a priori within in silico trials. Such tools are urgently needed, since most novel anti-cancer drugs tested in classical trials fail late (at phase II or III), with associated enormous costs (typically >$500 million). Failure is often due to poor in vivo efficacy, possibly from suboptimal treatment schedules or dosing. The overall attrition of late-stage drug development is unsustainably high [6, 7]. Our in silico tools and virtual trials could assist in moving decision points earlier in the drug development process. This will reduce costs in drug development. In silico systems medical modelling furthermore will allow identification of boundary conditions for patient cohorts that qualify for effective treatment. This will increase impact through re-design of clinical trials through ‘smart’ patient recruitment, thereby addressing a major translational bottleneck.
The APO-DECIDE consortium has excellent contacts with companies who developed and are developing Smac mimetics (such as Astex Pharmaceuticals and Genentech, see Scientific Advisory Board), as well as companies that introduced anti-EGFR monoclonal antibody therapy (such as Merck). These interactions will accelerate the implementation of the APO-DECIDE technologies into clinical trials, maximizing the impact of the APO-DECIDE outputs.
Future Commercial Outputs
Currently, there are no validated, commercially available assays predicting likelihood of CRC patient response to 5-FU based chemotherapy that have been adopted into standard clinical practice. In APO-DECIDE, we have developed the following prospects of possible developments to fill this gap in the market:
The APOPTO-CELL model as a whole has the potential to validate apoptosis execution proteins as predictive systems biomarkers, and may allow the identification of stage 3 CRC patients who will benefit from chemotherapy. The comprehensive analysis of the APOPTO-CELL phase space has allowed us to conduct and evaluate a number of model simplification strategies. As part of this work, we identified that the model can assist in determining that the proteins XIAP, caspase-3 and Smac alone or together should be investigated as a reduced set of combinatorial biomarkers that may be particularly valuable for subgroups of patients in the APO-DECIDE cohorts. Importantly, the systems modelling approach has provided specific concentration ranges for these proteins at which they will have dramatic effects on the capacity to execute apoptosis, irrespective of the concentration of the other proteins involved in the signalling network. This may provide additional and attractive routes for exploitation and potential commercialisation.
Caspase-3 protein level measured by RPPA was identified as prognostic marker in both the NI240 and the FOLFOX cohorts. Smac protein expression is a prognostic marker for both stage 2 and stage 3 NI240 patients who did not receive 5-FU-based chemotherapy. The analysis of the FOLFOX cohort data revealed that caspase-9 protein concentration constitutes a prognostic marker for CRC stage III patients who received 5-FU-based chemotherapy. The NI240 data demonstrates the effectiveness of APOPTO-CELL in accurately identifying stage III colorectal cancer patients who will respond to chemotherapy and those that may require additional targeted therapies to achieve meaningful treatment responses. Substrate cleavage computed by APOPTO-CELL on the basis of the RPPA protein measurements for FOLFOX cohort patients who underwent chemotherapy showed separation between good/bad outcome supporting NI240 results, however did not reach statistical significance. These results are preliminary as they are computed on the basis of only 36 months follow-up. Follow-up at 48-60 months are being gathered and it will be crucial to repeat the analysis as soon as these data will become available.
Our project also addressed the need for the development and validation of absolute quantitative protein measurements directly in tissue samples. Large-scale protein profiling through reverse phase protein array (RPPA) technology is a key development in systems medicine research, enabling a quantitative interrogation of large numbers of individual proteins and phosphoproteins. Expansion of RPPA technology to FFPE-based material was a key development process, and a main undertaking of the APO-DECIDE consortium. Furthermore, APO-DECIDE indicated the value of digital histopathology in providing quantitative protein profiles for single protein biomarker and combinatorial biomarker studies.
Advancing Systems Medicine
Further, the project have provided a significant impact on the field of systems medicine by delivering new mathematical tools that incorporate PK and PD data, as well as genotype and clinical data. The importance of cell-cell heterogeneity in the prediction of treatment responses, the modelling of drug responses on the tissue/organ level, and the modelling of patient data influencing PK and PD of drug interventions and overall therapy responsiveness represent significant advances in systems biology and systems medicine research. The project delivered a validated and versatile holistic modelling environment with future clinical applicability that will be able to predict not only treatment responses, but also will provide guidance regarding dosages and combinatorial therapies, ultimately benefiting the patient. Modelling of drug responses at the tissue/organ and whole-body level (‘holistic’ modelling approaches) is therefore also a major project impact.
Our approach will impact in evaluating novel targeted drugs in clinical use (cetuximab) and in clinical development (in particular Smac mimetics), as well as the optimal interventions/dosages for individual patients. Our activities therefore represent important new developments in the translation of systems biology approaches into the clinic. Apoptosis sensitizers are increasingly investigated in the pre-clinical and clinical setting as these drugs target the intrinsic apoptosis deficiency of cancer cells, which frequently leads to cross-resistance to chemotherapy. Targeting of such novel approaches through systems modelling, pre-clinical validation and clinical implementation addresses a key aim of the project call, i.e. research into the application of systems medicine approaches to patient stratification, and to the design of ‘smart’ clinical trials for novel, ‘targeted’ treatment approaches.
Protein measurement through quantitative digital pathology has high potential to become a routine diagnostic procedure , enabling the translation of systems biology approaches into the clinic in the future. We performed important development steps regarding the integration of digital histopathology data into existing modelling frameworks. Such technologies will be amenable to large-scale clinical trials as well as clinical care centres of all sizes. The results obtained in the APO-DECIDE project may therefore have a significant impact on future systems medicine approaches.
We are confident that there is novelty in this combination of the apoptotic computational model and biomarker panel, and that there is space in the colorectal cancer market for an assay that can guide clinicians towards the right treatment decisions. Due to the low success rate of many of the approved therapies for colorectal cancer, there is a need for diagnostics that can predict the likelihood of response, thus enabling clinicians to find alternative treatments for the ‘non-responders’. The probability that this would reduce the cost of healthcare (by reducing the unnecessary prescription of drugs and by reducing the risk of severe side effects that require hospitalisation) for each patient indicates that we can expect the relevant regulatory authorities to approve the use of such a diagnostic, and the associated reduction in adverse events for the patients also increases the likelihood of adoption into routine practice by clinicians.
Advancing Cancer Research
APO-DECIDE has also had major impact on cancer research. This is demonstrated by the promising clinical proof-of-concept data and results that demonstrated the viability of our approach and by the inclusion of EGFR modelling, and somatic mutations, inflammatory markers and other clinical relevant data as co-founding factors of our systems model predictions.
Impact on Health Economics
The APO-DECIDE activities address areas of highest clinical and economic relevance. Colorectal cancer (CRC) has among the highest incidence and mortality rates of any cancer with over 1.2 million cases diagnosed and over 0.5 million deaths from this disease worldwide p.a. despite recent advances in the therapy of CRC. For current therapies, the weekly cost per patient can vary between €2,000-5,000, but benefits are limited. The global anti-cancer drugs market exceeded $50 billion in 2009, and the global incidence is on the rise due to aging populations and an increased prevalence of CRC in the developing world . Where patients do not respond, this investment is wasted (as well as placing a needless burden of toxic side-effects on the patient). Furthermore, devising new therapies but not knowing which patients will ultimately benefit from these new therapies is unsustainably costly in the current economic climate. Currently, expenses of 1.5 billion USD are estimated to accumulate on average for each new drug to reach the market. The majority of these expenses are associated with phase II and III trials, of which >80% and >50% failed in recent years, respectively [6,7]. These failure rates and costs obviously are unsustainably high, and constitute unprecedented economic challenges to R&D active pharmaceutical and biotech companies. New clinical trials for CRC chemotherapeutics also have to be designed in the context of the already complex treatment regimens for CRC. Classical clinical trials designs are therefore becoming increasingly inappropriate, especially for testing novel and targeted drugs and combination treatments and the recent development of adaptive clinical trial designs is starting to address this significant problem .
During the project, partners presented at scientific events and have published and submitted publications on our work. For example, the project coordinator participated in the SYSCOL networking symposium, Barcelona, where he introduced the APO-DECIDE project. Further, the project coordinator and the SME ONCO were invited to introduce the APO-DECIDE project at the FP7 Cancer Workshop, Dublin held for FP7 coordinators on networking for collaboration and future joint programmes. Further, partly on the basis of the experience from the APO-DECIDE project, the project Coordinator was given a personal invite from Molecular Therapeutics for Cancer Ireland to join a workshop entitled, ‘National Cancer Research Centre of Ireland’ (NCRCI), with the purpose of creating strategy for developing a national umbrella organisation for large scale cooperation between clinical, laboratory, and translational scientists and investigators from all disciplines of cancer research. The impact from these dissemination activities has been an overall interest in the project both for the academic research and commercial SME partners. As an indicator for the interest the project has created in the research environment, the project has been featured on the ‘Retell’ website (a contractor to the European Commission’s DG Research and Innovation that helps EU-funded projects publicize their achievements and work) as a ‘DG Research Success Story’. The coordinator, through the Centre for Systems Medicine, has also contributed to the ‘Roadmap to Systems Medicine’, now published https://www.casym.eu/index.php?index=90.
1) Cancer facts & figures, 2014. American Cancer Society.
2) Ferlay, J. et. al. 2012. Cancer incidence and mortality patterns in Europe: Estimates for 40 countries in 2012. European Journal of Cancer. 49; 1374-1403
3) de Gramont A, et. al., 2000. Leucovorin and fluorouracil with or without oxaliplatin as first-line treatment in advanced colorectal cancer. J Clin Oncol 18(16): 2938-2947
4) De Roock W, et. al., 2011. KRAS, BRAF, PIK3CA, and PTEN mutations: implications for targeted therapies in metastatic colorectal cancer. Lancet Oncol 12(6): 594-603
5) Faratian D, et. al., 2009. Systems pathology--taking molecular pathology into a new dimension. Nat Rev Clin Oncol 6(8): 455-464Arrowsmith J. 2011. Trial watch: Phase II failures: 2008-2010. Nat Rev Drug Discov 10(5): 328-329
6) Arrowsmith J. 2011. Trial watch: phase III and submission failures: 2007-2010. Nat Rev Drug Discov 10(2): 87
7) Arrowsmith J. 2011. Trial watch: Phase II failures: 2008-2010. Nat Rev Drug Discov 10(5): 328-329
8) Ledford H. 2011. Translational research: 4 ways to fix the clinical trial. Nature 477: 526-528
9) Center MM, et. al., 2009. Worldwide variations in colorectal cancer. CA Cancer J Clin 59(6): 366-378
10) Van Schaeybroeck S, et. al., 2011. Implementing prognostic and predictive biomarkers in CRC clinical trials. Nat Rev Clin Oncol 8(4): 222-232
List of Websites:
Grant agreement ID: 306021
1 November 2012
31 October 2014
€ 3 896 336
€ 2 999 482
ROYAL COLLEGE OF SURGEONS IN IRELAND
This project is featured in...
Deliverables not available
Grant agreement ID: 306021
1 November 2012
31 October 2014
€ 3 896 336
€ 2 999 482
ROYAL COLLEGE OF SURGEONS IN IRELAND
This project is featured in...
Grant agreement ID: 306021
1 November 2012
31 October 2014
€ 3 896 336
€ 2 999 482
ROYAL COLLEGE OF SURGEONS IN IRELAND