Identification of paclitaxel resistance through a new statistical approach based on a random forest of perfect trees classifcation

International audienceBackground: Predictors of paclitaxel sensitivity in breast cancer published ten years ago, are still pending. The authors showed that paclitaxel pathological complete response (pCR) was in one hand, encountered in aggressive breast tumor with immune response and in another hand, paclitaxel resistance in less aggressive tumor. We have developed a new analysis paradigm, mixing neurons into nodes of trees classification and news class of statistical information based on free-error trees classification. We proposed to reanalyze the Bauer et al’s dataset using this novel approach. Methods: GES22513 dataset including 14 duplicated observations and 54675 anonymized probes was analyzed. A random forest of one million trees whose nodes were composed of neurons including 15 probes, was developped. We selected probes for which a free-error classification was obtained and ranked them according to the inverse of the probability of being a confounding factor and to the inverse of the probability of interacting with another probe. We compared the sets of probes which were necessary to obtain an error-free classification between those associated with a decrease and those associated with an increase of the probability of pCR. Results: Our 15 best ranked predictors were free-error classification for all observations. This includes gene expression of TLCD2, BRCC3, CHI3L2 and PROX1. Their over-expressions were associated with an increase in the probability of pCR, and gene expression of APH1B, ARFGEF1, ARID2, BPGM, CAMK2N1, CCNY, PARM1, PHKA1, PSMD9, SUDS3 (two probes) whose over-expressions are associated with a decrease in the probability of pCR. Ten out of these probes were concordant with Bauer et al’s conclusion. Four probes ( BPGM, PHKA1, CCNY and ARFGEF1) are in contradiction with it. The limited biological information were available for TLCD2. The statistical analysis also showed that TLCD2, BBRCC3, CHI3L2 and PROX1 were altogether positively modulated by eight genes/probes ( CKS1B, ADIG, NCR3, RIN3, NIPAL1, 234422_at, DCLRE1C, SLC17A4). At the opposite, the modulation of genes associated with a decrease in the probability of pCR, was rather heterogeneous and involves many more genes. Conclusions: This preliminary work shows that our statistical approach allows a perfect classification of tumors with and without pCR. Also, it proves that the selected probes/genes are respectively associated with aggressiveness/basal and less aggressiveness/luminal phenotypes. These results need to be validated on an independent cohort