Mutation of TP53, a critical tumor suppressor protein, is a marker of breast cancer prognosis and can potentially predict preoperative treatment response.¹ Currently, TP53 mutational status is determined through DNA sequencing of biopsied tissue. Radiogenomics is defined as the attempt to identify associations between radiomic (computer extracted image features from radiographic imaging) and genomic features for the prediction of pathological biology and outcomes. In this work, we explore a radiogenomic approach to non-invasively predict TP53 mutational status from DCE-MRI. TP53 inactivation creates disorder across multiple spatial scales, e.g. uncontrolled cytoskeleton formation² distorts micro-architecture, while unregulated division creates disordered cell clusters. Standard DCE-MRI pharmacokinetic (PK) features – computed by modeling contrast agent uptake from intensity of post-contrast DCE-MRI phases – quantify tumor-associated changes in permeability, but offer no means of capturing multi-scale disorder patterns. In this work we explored the ability of a computer-extracted radiogenomic descriptor, Co-occurrence of Local Anisotropic Gradient Orientations (CoLlAGe), to predict the TP53 mutational status from routine clinical breast DCE-MRI alone. Additionally we also compared the approach with the predictive power of PK parameters obtained from DCE-MRI. CoLlAGe involves computing pixel-wise gradient orientations, determining dominant orientations via Principal Component Analysis, and extracting second order statistical features from the co-occurrence matrix of dominant orientations. The hypothesis behind our approach is that CoLlAGe will identify subtle microarchitecture differences on MRI associated with breast cancers with and without the TP53 mutation through a multi-scale quantification of disorder across local image gradients.The ability of CoLlAGe to predict TP53 mutational status was compared against PK features on 23 DCE-MRI breast cancer cases with known TP53 mutational status. Breast tumors were obtained from patients on a preoperative clinical trial and whole-exome DNA sequencing was performed with an average coverage of 100X, followed by alignment, variant calling,³ and variant filtering to eliminate potential germline variants. 13 breast cancers harboring non-synonymous TP53 mutations were annotated as TP53^MUT, while the remaining 10 samples were classified as TP53^WT. 1.5 or 3.0 T STIR and T1w fat saturation axial scans were obtained with an 8 or 16 channel dedicated breast coil. Patients were imaged after administration of gadolinium contrast agent through intravenous injection, and tumor extent was manually annotated onto images by a radiologist. DCE-MRI PK parameters reflecting vascular permeability to contrast agent (Ktrans: volume transfer coefficient, Kep: flux rate constant, Ve: extracellular volume ratio) were computed from DCE-MRI post-contrast phases using Toft’s model⁴ for comparison. A supervised hierarchical clustering of CoLlAGe parameters was used to identify features capable of distinguishing TP53^MUT and TP53^WT. The ability of CoLlAGe and PK features to predict mutational status was assessed through separation of tumors corresponding to DNA sequencing-confirmed mutational status in a furthest distance unsupervised hierarchical clustering. Through visual inspection of supervised cluster expression maps, skewness, and kurtosis of the information measure of correlation were identified as discriminating radiogenomic features. The skewness and kurtosis of the information measure of correlation alone were found to be 74% accurate (77% sensitivity, 70% specificity) in distinguishing TP53^MUT and TP53^WT (Figure 1). In figure 2, the difference in expression of this feature is shown with TP53^MUT and TP53^WT examples. Three additional distinguishing features for skewness (inertia, sum average, difference variance) and kurtosis (energy, entropy, correlation) were identified and included in an 8 feature unsupervised clustering. We identified two distinct clusters, representing TP53^MUT and TP53^WT, with accuracy of 78% (85% sensitivity, 70% specificity) (Figure 3, a.). TP53^WT over-expressed skewness and kurtosis of information measure of correlation, while these features were under-expressed in TP53^MUT. Interestingly, the TP53^MUT cluster contained two sub-clusters, which differed in expression of skewness of the sum variance, inertia, and difference variance. All three misclassified WT tumors were sorted to the TP53MUT sub-cluster with high expression of these features. A potential explanation for these two mutant phenotypes is differing inter-tumor TP53 inactivation.⁵ Meanwhile, PK parameters produced two main clusters: the first bearing 38% of mutants and 44% of WT, the second 54% of mutants and 56% of WT. A third, separate cluster comprised a single mutant study (Figure 3, b.).We reported preliminary success in the prediction of TP53 mutational status in breast cancer from DCE-MRI through analysis using CoLlAGe features. CoLlAGe demonstrated higher accuracy in sorting TP53 mutant and WT studies than standard DCE-MRI PK parameters. A non-invasive means of discerning TP53 mutational status may allow clinicians to easily determine prognosis, assess treatment response, and inform treatment strategy.