(T1230-10-58) AI/ML-Driven In Silico Modeling of the COVID-19 Infection for Accelerated Drug Repurposing and Personalized Therapy

Purpose: The COVID-19 pandemic subjected the global healthcare support system to a tremendous burden. Due to an incomplete and scarce patient dataset, a key challenge lies in the stratification of patients for predicting the severity of disease and successful treatment options. Our goal was to develop a decision-making tool based on our artificial intelligence (AI) and machine learning (ML) platform, which can impute missing key clinical parameters for predicting potential disease severity outcomes in patients infected with SARS-CoV-2.

Methods: Our approach entails an ML/Deep Learning (DL)-driven patient stratification model integrated into a hybrid AI/ML-driven platform, capable of predicting COVID-19 disease outcomes from incomplete clinical datasets. Published clinical datasets were used to train and validate classification and regression models against patient survival and recovery rate. Automated statistical algorithms were deployed to select the most essential features from the input descriptors. Missing values of the patient biomarkers were filled by the multivariate imputation technique. AI/ML algorithms were combined with model stacking to improve the overall performance.

Results: The results demonstrated that DL/ML model was successfully implemented to predict COVID-19 disease outcomes with an accuracy of 0.85, AUC ROC score of 0.88, and F1 Score of 0.89 using patient datasets that were missing 75% of parameter values. Our results demonstrate an innovative approach that allows for AI/ML-driven predictions performed by the computational platform capable of predicting personalized treatment strategies for COVID-19 patients in healthcare units with limited diagnostic capacities.

Conclusion: COVID-19 infections result in widely different outcomes for patients. Infection models can be used to define viral load and tissue damage, and their changes due to therapeutic interventions. For the first phase of this study, our team created such a model, which was validated via modeling of therapy with calcium channel blockers and ACE inhibitors. The second phase of this study will include developing an advanced COVID-19 infection model, which will predict disease progression and best fit therapeutics targeted for very discrete patient classes. This model will use clinical databases as a data source to categorize patients into cohorts with different reactions to infection. The model will be scalable to model progression and outcomes of viral diseases and suggest the most favorable personalized treatment. This study illustrates that a computational platform can accurately model in vivo the efficacy of drug combinations leading to a selective patient dose optimization and patient stratification.

Acknowledgments: The authors of this research are employed by VeriSIM LIfe and have equity positions in the company.
Research performed was funded in part by a SIBR grant from National Science Foundation.

Our general model created during Phase 1 research studies allows to simulate infection progressions in terms of viral load and tissue damage associated with therapeutic concentration. Changes of each of those parameters allow to predict disease outcomes after the use of therapeutic with various potencies regarding main targets among disease pathways. Disease outcomes depend on prediction of the drug PK profile giving understanding of the optimal dosing regimens for each therapeutic depending on its properties and potency.
Implementation of the model on our portal makes possible to make quick modeling of any scenario of the COVID-19 disease progression in patients.



The key component of the COVID-19 module is AI-based patient stratification module that will be use to sort out different patient cohorts depending on the biomarker changes and predicted response to pharmaceuticals. This module will be built in the base of real individual data of COVID-19 patients treated with various types of therapies and integrated with viral infection model and proprietary VSL drug database. Final results of this work will be suggestion for the best fit personalized treatment of various patient cohort or even individual patients.



NSF grant executed by VSL team is devoted to development of comprehensive viral infection model that can simulate disease progression and outcomes as well as best fit therapeutic regimens. Phase 1 was aimed at building a general model of SARS-CoV-2 infection and estimation of the outcomes following a sample therapy with two drug classes inhibiting viral replication and suppressing metabolic pathways responsible for clinical manifestations of the COVID-19 disease.