A variety of factors influence a patient’s clinical outcome, including intrinsic characteristics of the patient, disease, or medical condition, and the effects of any treatments that the patient receives. This project’s primary purpose is to develop dedicated predictive markers for specific clinical treatment regimens using convolutional neural networks (deep learning) combined with our prognostic biomarkers.
A prognostic biomarker provides information about the patient’s overall cancer outcome, regardless of therapy, whilst a predictive biomarker gives information about a therapeutic intervention’s effect. It follows from this definition that a biomarker can be both prognostic and predictive and that the distinction can be difficult. As most cancer patients will receive some form of treatment and since no treatment has a documented effect on all patients in a cohort, one cannot know if the biomarker applied provides information about the patient’s overall outcome or response to the treatment. For example, DNA ploidy is regarded as a generic marker linked to genomic instability and referred to as a prognostic marker. In several studies, we have demonstrated that non-diploid tumours indicate a worse prognosis for patients with prostate cancer. The test has most often been applied to the tumour after removal (radical prostatectomy), and we have referred to the statistically significant correlation between disease-free survival and ploidy as a prognostic finding. We could have referred to the same finding as a predictive marker for radical prostatectomy, as radical prostatectomy was the only therapy given before the event (time to recurrence). The generally accepted definition of prognostic vs predictive marker is perhaps a bit confusing, and the language we use certainly is, as a prognostic marker predicts the outcome for a patient. The keyword here is regardless, meaning that when labelling a biomarker “prognostic”, we ignore the potential effect of therapy. To label a biomarker “predictive”, the design must be such that the effect measured is due to a specific therapy. A predictive biomarker can, therefore, also be a target for therapy.
In this project, we define predictive studies as studies of response, and predictive markers as markers that can predict a given response to a given treatment.
Good materials to develop and test a predictive marker are hard to come by, as they must be both prospective and large. Prospective studies normally include several institutions, and the biobank is rarely centralised as the biopsies and surgical tissue most often will remain at the institution treating the patient. Successful prospective studies with a distinct result (required for the development of a marker) are often stopped before they become large enough. Once it becomes evident that one arm of the study is beneficial to the patients, it would be unethical to withhold the best treatment from the other patients in the study. As a result, ICGI is constantly searching for materials, and we have some good leads with big pharma.
The Short Course Oncology Therapy (SCOT) Study
Six months of oxaliplatin-containing chemotherapy is usually given as adjuvant treatment for stage III colorectal cancer. The Short Course Oncology Therapy (SCOT) study is an international, randomised, phase 3, non-inferiority trial conducted at 244 centres. Patients aged 18 years or older with high-risk stage II and stage III colorectal cancer underwent central randomisation with minimisation for the centre, choice of regimen, sex, disease site, N stage, T stage, and the starting dose of capecitabine. Patients were assigned (1:1) to receive three months or six months of adjuvant oxaliplatin-containing chemotherapy. The chemotherapy regimens could consist of CAPOX (capecitabine and oxaliplatin) or FOLFOX (bolus and infused fluorouracil with oxaliplatin). The regimen was selected before randomisation in accordance with choices of the patient and treating physician. The primary study endpoint was disease-free survival, and the non-inferiority margin was a hazard ratio of 1.13. The primary analysis was done in the intention-to-treat population, and safety was assessed in patients who started study treatment. The trial recruited patients across six countries (UK, Denmark, Spain, Sweden, Australia, and New Zealand), followed up for a minimum of three years, up to eight years post-randomisation. The SCOT study was designed as a non-inferiority trial, aiming to exclude a maximum 2.5% fall in 3-year disease-free survival by halving the adjuvant treatment duration. Secondary endpoints were overall survival, toxicity, quality of life and cost-effectiveness.
A total of 6088 patients underwent randomisation between March 27, 2008, and November 29, 2013. The intended treatment was FOLFOX in 1981 patients and CAPOX in 4107 patients. There were 3044 patients assigned to the 3-month group, and 3044 assigned to the 6-month group. Nine patients in the 3-month group and 14 patients in the 6-month group did not consent for their data to be used, leaving 3035 patients in the 3-month group and 3030 patients in the 6-month group for the intention-to-treat analyses. At the cut-off date for analysis, there had been 1482 disease-free survival events, with 740 in the 3-month group and 742 in the 6-month group. The 3-year disease-free survival was 76.7% (95% CI 75.1–78.2) for the 3-month group and 77.1% (75.6–78.6) for the 6-month group, giving a hazard ratio of 1·006 (0.909–1.114, test for non-inferiority p=0.012), significantly below the non-inferiority margin. Peripheral neuropathy of grade 2 or worse was more common in the 6-month group (237 [58%] of 409 patients for the subset with safety data) than in the 3-month group (103 [25%] of 420) and was long-lasting and associated with worse quality of life. There were 1098 serious adverse events reported (492 reports in the 3-month group and 606 reports in the 6-month group), and 32 treatment-related deaths occurred (16 in each group).
In the whole study population, three months of oxaliplatin-containing adjuvant chemotherapy was non-inferior to six months of the same therapy for patients with high-risk stage II and stage III colorectal cancer and was associated with reduced toxicity and improved quality of life. Although the study was underpowered, these data suggest that a shorter duration leads to similar survival outcomes with a better quality of life and thus might represent a new standard of care.
Predictive marker development
One would think that when two different treatment durations give the same result, it is either because the treatment has no effect, or all effect sets in already after the shortest treatment duration. Should there be patients that do not have sufficient effect unless treated over the longer period, one would expect a lower number of events in that group unless the follow-up time is too short or events are death from toxicity of the prolonged treatment. The SCOT study demonstrated non-inferiority of the 3-month regimen for all patients regardless of treatment, including CAPOX treatment, but not FOLFOX treatment.
Although the disease-free survival was remarkably similar for the two arms, indicating that three months of therapy is non-inferior to six months, there might be individual variations to treatment response, e.g., that there is a group of patients that need six months of therapy and another group where six months is excessive, causing neuropathy, other side-effects and even death (which actually would be recorded as an event in this study). ICGI has recieved access to tissue sections from 2500 UK patients.
One hypothesis should be that patients classified as poor prognosis with one or more of our prognostic markers will benefit from six months of treatment, whereas those classified as good prognosis will show no difference in disease-free survival after three vs six months of treatment. The first marker of choice would be Histotyping using the DoMore-v1 marker. We will run an inference with the algorithm published in The Lancet as soon as all cases have been received, at the end of Q2 2021. The result to confirm this hypothesis would be a significantly higher number of poor prognosis disease-free patients in the 6-month arm compared to the 3-month arm. For the good prognosis group, the difference should only be on frequency and grade of neuropathy and other safety parameters, not disease-free survival. As we expect 1891 disease-free patients and 609 patients experiencing an event in our cohort of 2500, the power should be sufficient for this study to evaluate the DoMore-v1 marker as a predictive marker for optimal treatment duration (if the difference is clinically significant).
Another hypothesis could be that the patient’s sensitivity to the toxicity of the treatment could be detectable in the tumour tissue. Safety was only assessed in 868 patients or 14.3%. If we receive a representative group with the 2500 patients, safety has been assessed in 356 patients and 50% in each arm. Sensory neuropathy of grade 2 or worse was seen in 58% in the 6-month arm and 25% in the 3-month arm, indicating that we will have a positive sensitivity group of 148 and a negative group of 208 patients. This might be enough to train a CNN but might require another dataset for validation. An alternative might be to validate on the patient-recorded outcome FACT/GOG-Ntx4 questionnaire, where peripheral neuropathy data are available from 47% of the patients and assumably for 1179 of our 2500 cases. Otherwise, we could use this group for training and validate on the group assessed by the investigators. This could potentially be a predictive marker for neuropathy.
This text was last modified: 18.08.2021