Background: Homologous recombination and repair (HRR) deficiency (HRD) is characterized by genomic instability associated with dysfunction in BRCA1/2 or other HRR genes. Patients with breast cancer harboring an HRD phenotype with or without HRR mutations have derived clinical benefit from PARPi therapy. We have previously shown that GuardantINFINITY, a novel genomic and epigenomic liquid biopsy assay, can identify HRR SNVs, indels, rearrangements, copy number loss, reversions, and BRCA1 promoter methylation for patient selection and resistance monitoring, a major challenge for PARPi treatment. Here, we present a method of predicting HRD status by cfDNA using GuardantINFINITY in patients with advanced breast cancer. Methods: We developed a probabilistic genomic model to predict HRD status, inferred from genome-wide somatic SNV, indel, and CNV signatures indicative of BRCA1/2 deficiency, including large-scale state transitions (LST), whole-genome tumor loss of heterozygosity (LOH), telomeric allelic imbalance (TAI). A second probabilistic model, based on targeted measurements of genomic and epigenetic changes was learned from a subset of clinical samples to enhance the genomic model. The model was trained and tested on a cohort of over 12,000 GuardantOMNI and GuardantINFINITY clinical breast cancer samples to assess the sensitivity of accurately detecting deficiency in select HRR-genes (BRCA1/2, PALB2, RAD51D). The aggregated predictive model was validated on an independent cohort of breast cancer samples. Results: The model based on genomic and epigenetic signals identified HRR-gene deficiency in a breast cancer cohort with a estimated 95% LoD of 22.5% tumor fraction and an AUC of 0.75 across all tumor fractions, an improvement in sensitivity over when only the genomic model is used (AUC of 0.7, 95% LoD of 28%). Specificity remained high for both models (100%, n = 83) in cancer free samples. Application of this model in a cohort of 1,101 patients with unselected breast cancer identified 390 patients with an HRD phenotype, of whom 192 patients had a known HRR-gene mutation. Of the 201 patients with a known pathogenic mutation in an HRR gene, 49 also had co-occurring LoH in the same gene, indicating biallelic loss. Conclusions: In this analysis, we demonstrate that a probabilistic model of genomic and methylation predictors can detect HRD status in patients with breast cancer from cfDNA using GuardantINFINITY. Additional analytical and clinical studies to further evaluate this model are ongoing. With HRD prediction, GuardantINFINITY provides a comprehensive minimally-invasive solution for PARPi and DNA damage treatment selection, longitudinal monitoring, and an exploratory platform for investigating epigenetic signals that may underpin resistance.