Background: During the I-SPY2 clinical trial, ganitumab (G) immunotherapy (IO) was co-administered with metformin (GM) to counteract G-induced hyperglycemia. The GM + standard of care (SOC) combination showed promising results but did not meet the statistical threshold for phase III. Predicting individual patient response to IO is a key limitation for clinical use and a reason why IO phase II to III oncology trial transitions fail. Thus, we used SimBioSys TumorScope (TS), a 4D spatiotemporal multiscale biophysical model, to identify predictive IO biomarkers of pCR in response to GM+SOC. Methods: Patients that received GM+SOC therapy (n=41) and SOC (n=41) matched controls from the I-SPY2 trial were included in analyses. Using SOC clinical data and DCE MRIs, we generated biophysical simulations of each individual patient’s response to SOC therapy. Derived from the TS simulation object of each patient, our “simul-omics” data represent tumor morphology, breast tissue proportions, drug response and delivery, and microvasculature-related feature sets. We also had access to the pre-treatment tumor biopsy-derived transcriptome microarray data. Predictive modeling was performed in a modular 3-step fashion: 1. clinical-only data to predict pCR, 2. clinical data plus candidate simul-omics data to predict pCR, and 3. available feature space for features that interact with treatment to provide the strongest predictive model of pCR. Results: The SOC data model (HR status, grade, treatment, age, race) predicted pCR with an accuracy (acc) of 0.67, sensitivity (sens)=0.71, specificity (spec)=0.66, and Cohen’s kappa=0.27. Simul-omic features median tumoral Kt at sim week 1 (acc=0.73), and tumoral pre-contrast small area emphasis (acc=0.77) improved model performance over SOC. When included in the same model, they increased predictive acc of the model to 0.79 (sens=0.64, spec=0.82, kappa=0.41). Simulation-derived features improved predictive power beyond SOC data. Exploratory search of the feature-space identified simul-omic features that would generate the most predictive models with a significant treatment interaction effect. Two simul-omics features added predictive strength to the model, with near-nominal interaction effects: phi_quant_90_regimen_wk22 (late timepoint mv density) (acc=0.84, kappa=0.58), and slopemap_tumor_glcm_clustershade (acc=0.86, kappa=0.60). We tested transcript expression as interaction terms in the model. The best performing model with an interaction effect was SHISA4 (acc=0.88, sens=0.86, spec=0.88, kappa=0.67). Conclusions: Our results show value in: 1. predicting pCR, the additive value of TS simulation data over just clinical/SOC data, 2. identification of potentially mechanistic drug targets facilitating GM response, and 3. potential to stratify patient populations into responder vs. non-responder categories prior to treatment.