Maddalena Centanni1, Afroditi Nanou2, Leon W.M.M. Terstappen2,3, Frank A.W. Coumans2,3, Mats O. Karlsson1, Lena E. Friberg1
1. Department of Pharmacy, Uppsala University, Uppsala, Sweden 2. Department of Medical Cell BioPhysics, Faculty of Science and Technology, University of Twente, The Netherlands 3. Decisive Science, Amsterdam, The Netherlands
Objectives: Liquid biopsies represent a promising advancement in cancer diagnostics, offering a minimally invasive approach to sampling that has potential to provide clinicians with real-time insights into the disease status and progression (1). Challenges persist however in the identification and enumeration of circulating tumor cells (CTCs) and tumor-derived extracellular vesicles (tdEVs) in patient samples, as they are time-consuming, costly and prone to error, even under the high level of standardization provided by the CellSearch system (2). To optimize the automated enumeration of CTCs and tdEVs, deep learning (DL) approaches were developed over several iterations (2). The objectives of this study were to compare CTC and tdEV enumeration methods in terms of their connection to the tumor size time course and their association to overall survival (OS) in patients with metastatic colorectal cancer (mCRC).
Methods: Blood samples from 446 mCRC patients were available from the CAIRO2 trial (3). CTCs were counted manually. In addition, three DL counting methods were employed (i) a k-nearest neighbors’ (kNN) classifier for CTC and tdEVs, (ii) a kNN classifier with higher precision and lower recall (kNNs) for CTCs, and (iii) a decision tree classifier (Tree) for CTCs (2). Consequently, four unique CTC and one tdEV dependent variable were available from each liquid biopsy sample. First, count models of the Poisson family were applied on the data to describe the mean count over time (λ(t)) for each individual. The λ(t) was linked to a previously developed modeling framework (4), including a tumor size model describing the time course of the sum of the longest diameters (SLD(t)) and a logistic regression function for dropout. In the second step, a log-logistic time-to-event model evaluated the relations between the five separate λ(t) of each individual and OS.
Results: For each CTC and tdEV variable, 4961 observations were available. A negative binomial Poisson model, accounting for overdispersion, described λ(t) well for all counting methods. The fit of the models for the manually counted CTCs and the model for kNN tdEVs improved when absolute SLD(t) was related to λ(t) through an exponential function. For CTCs quantified by the kNN, kNNs and Tree algorithms, λ(t) was affected by the ratio of SLD(t) and baseline SLD (SLD(t)/SLDbaseline) in a power function. The model fit further improved when the effect by absolute SLD(t) was delayed through an effect compartment, i.e. (SLDdelayed/SLDbaseline). The λ(t) for the tdEVs and manually counted CTCs exhibited half-life from baseline of 2 weeks, unlike the λ(t) for kNN, kNNs and Tree counted CTCs, which showed a half-life from baseline of 15 weeks. All five λ(t) were significantly associated with OS, as demonstrated by their relationship to the hazard function through an Emax function within the hazard function’s exponential component. With respect to objective function, the λ(t) of the manually quantified CTCs was the most significantly related to OS, followed by the λ(t) of the tdEVs. However, when assessing the hazard ratios (HRs) based on the mean λ(t) after six months of treatment, the λ(t) for kNN, kNNs and Tree counted CTCs (HR = 3.44, 3.73 and 3.61, respectively) was considerably higher than those of the λ(t) for manually counted CTCs and kNN counted tdEVs. This indicated that kNN, kNNs, and Tree could better identify subgroups of patients with significantly higher risk of death.
Conclusions: Counts of CTCs and tdEVs were significantly related to OS in patients with mCRC, potentially offering opportunities for tailoring treatment on a personalized basis. The developed modeling framework may be further adapted to explore how treatment modifications based on dtEV and CTC counts may impact OS, and enable exploration of biomarker thresholds for identifying patients at elevated risk of mortality.
References:
(1) Di Meo A, Bartlett J, Cheng Y, Pasic MD, Yousef GM. Liquid biopsy: a step forward towards precision medicine in urologic malignancies. Mol Cancer [Internet]. 2017;16(1):80. Available from: https://doi.org/10.1186/s12943-017-0644-5
(2) Nanou A, Stoecklein NH, Doerr D, Driemel C, Terstappen LWMM, Coumans FAW. Training an automated circulating tumor cell classifier when the true classification is uncertain. PNAS Nexus [Internet]. 2024 Feb 1;3(2):pgae048. Available from: https://doi.org/10.1093/pnasnexus/pgae048
(3) Tol J, Koopman M, Miller MC, Tibbe A, Cats A, Creemers GJM, et al. Circulating tumour cells early predict progression-free and overall survival in advanced colorectal cancer patients treated with chemotherapy and targeted agents. Ann Oncol Off J Eur Soc Med Oncol. 2010 May;21(5):1006–12.
(4) Netterberg I, Karlsson MO, Terstappen LWMM, Koopman M, Punt CJA, Friberg LE. Comparing Circulating Tumor Cell Counts with Dynamic Tumor Size Changes as Predictor of Overall Survival: A Quantitative Modeling Framework. Clin cancer Res an Off J Am Assoc Cancer Res. 2020 Sep;26(18):4892–900.
Reference: PAGE 32 (2024) Abstr 11222 [www.page-meeting.org/?abstract=11222]
Poster: Drug/Disease Modelling - Oncology