Background
Zr-Df-Crefmirlimab is a humanized, engineered, 80-kDa minibody (an antibody fragment), with high affinity to human CD8 (kd of 0.4 nM). It has been evaluated as an imaging agent in a Phase-II open-label multi-dose study (NCT03802123) in patients with metastatic solid malignancies (figure 1) scheduled to receive standard of care immunotherapy (Nivolumab, Pembrolizumab and Ipilimumab-Nivolumab). Supervised machine learning (ML) combined with tumor growth-inhibition (TGI) modeling was applied to predict clinical response using various baseline patient characteristics including CD8 density (via biopsy) and CD8 PET imaging SUV (Standardized Uptake Value).
Methods
Modeling framework for prediction of response to immunotherapy was developed leveraging multimodal data (figure 2) including CD8 PET imaging readouts. Random forest was used for classification of response using only baseline characteristics from 32 patients. Data was randomly assigned into training and test data (65:35). Performance of the model on test data was evaluated using area under the receiver operating characteristic curve (ROC AUC). Rates of tumor growth and kill were estimated from a TGI model developed using early tumor kinetic data (first three time points). Baseline CD8 density and SUV were used along with estimated TGI model parameters to classify patient response using supervised ML. Explainable ML techniques like partial dependence and individual conditional expectation were leveraged to further explore contribution of features of interest towards model outcome.
Results
Using only baseline characteristics, AUC of 0.75 was achieved on the test data with an overall prediction accuracy of 82%. Using partial dependence of model features, increase in the likelihood of patient response was observed with increasing CD8 density and SUV at baseline. Early tumor kinetic data was described reasonably well by the developed TGI model. AUC of 0.88 was achieved on the same test data (as above) with an overall prediction accuracy of 91% using individual estimates of tumor growth and kill from the TGI model and baseline CD8 density and SUV. Decrease in likelihood of response was associated with increasing rate of tumor growth and smaller baseline CD8 density.
Conclusions
Baseline CD8 density and PET SUV data were used along with other patient characteristics to predict clinical outcome to immunotherapy with a reasonable degree of accuracy. Using a combined approach of tumor growth inhibition modelling and supervised machine learning, high precision in prediction of clinical outcome was achieved leveraging baseline CD8 PET information and early tumor kinetic data.
Abstract 1281 Figure 1
Abstract 1281 Figure 2