Analysis of individual target lesions for tumor size models of drug resistance: a new methodology encompassing signal processing and machine learning
N. Terranova (1), P. Girard (1), U. Klinkhardt (2, 3), A. Munafo (1)
(1) Merck Institute for Pharmacometrics, Merck Serono S.A., Lausanne, Switzerland, (2) Merck KGaA, Darmstadt, Germany, (3) Current affiliation: CureVac GmbH, CureVac GmbH, Paul-Ehrlich-Str. 15, Tübingen 72076, Germany
Objectives: Exploratory data analysis is a fundamental step in the model development effort. It becomes even more relevant when, for developing semi-mechanistic tumor size models of drug resistance, we consider individual target lesions rather than their sum [1]. In this work, we propose a new methodology, inspired by signal processing and machine learning, for clustering classified individual tumor lesions according to the degree of similarity in their dynamics. Resulting information can drive the modeler in rationally selecting the most appropriate modeling strategy for describing the tumor resistance profile of cancer patients.
Methods: A new classification of tumor individual target lesions, based on functional and location criteria, was defined, validated by a clinical expert and applied to clinical studies in metastatic colorectal cancer (mCRC). This classification was implemented in SAS® software through keywords recognition on the lesion description recorded by physicians. When lesions were similarly classified for one patient, the sum of tumor measures was computed. Cross correlation (CC) was used to measure the similarity among classified lesion dynamics by also considering potential delays. Resulting correlations were clustered with K-means to obtain a straightforward and overall interpretation. Both methods are part of the Stats package in R [2].
Results: We have classified 2038 individual target lesions, selected and measured according to the WHO criteria, of 642 mCRC patients from two Phase II studies. CCs have been estimated for 216 patients with multiple classified individual lesions, and clustered. Results from both studies are consistent across tested scenarios and highlight a similar tumor dynamics in about the 60%-70% of classified lesions. The degree of similarity decreases when computed without considering any delay between lesion dynamics.
Conclusions: The proposed methodology, by integrating knowledge from other fields, provides a novel and suitable workflow for the non-parametric analysis of individual target lesions prior to any modeling step. Our approach is flexible enough to be applied to any case study. Moreover, by coupling the information on the target tumor metastases along with the lesion dynamics, it enables the modeler to precisely evaluate the maximum information gain obtainable by considering individual tumor lesions in tumor size modeling of resistance to anticancer drugs.
This work was supported by the DDMoRe project (www.ddmore.eu).
References:
[1] Bedard, Philippe L., et al. Tumour heterogeneity in the clinic. Nature (2013) 501 (7467): 355-364.
[2] MacQueen, J. B., et al. Some Methods for classification and Analysis of Multivariate Observations. Proceedings of 5th Berkeley Symposium on Mathematical Statistics and Probability (1967) 1: 281–297.