Artificial Intelligence-Driven Designer Drug Combinations: From Drug Development to Personalized Medicine

Abstract

Artificial intelligence holds great promise in transforming how drugs are designed and patients are treated. In a study recently published in Science Translational Medicine, a unique artificial intelligence platform makes efficient use of small experimental datasets to design new drug combinations as well as identify the best drug combinations for specific patient samples. This quadratic phenotypic optimization platform (QPOP) does not rely on previous assumptions of molecular mechanisms of disease, but rather uses system-specific experimental data to determine the best drug combinations for a specific disease model or a patient sample. In this commentary, we explore how QPOP was applied toward multiple myeloma in the study. We also discuss how this study demonstrates the potential for applications of QPOP toward improving therapeutic regimen design and personalized medicine.

Comment on: Rashid, M.; Toh, T. B.; Hooi, L.; et al. Optimizing Drug Combinations against Multiple Myeloma Using a Quadratic Phenotypic Optimization Platform (QPOP). Sci. Transl. Med. 2018, 10, eaan0941.

Artificial intelligence (AI) holds great promise in transforming how drugs are designed and patients are treated. In a study recently published in Science Translational Medicine, a unique AI platform makes efficient use of small experimental datasets to design new drug combinations as well as identify the best drug combination for specific patient samples. ¹ This quadratic phenotypic optimization platform (QPOP) does not rely on previous assumptions of molecular mechanisms of disease, but rather uses system-specific experimental data to determine the best drug combinations for a specific disease model or a patient sample. In this commentary, we explore how QPOP was applied toward multiple myeloma in the study. We also discuss how this study demonstrates the potential for applications of QPOP toward improving therapeutic regimen design and personalized medicine.

Bortezomib (Bort) has been the standard of care for first- and second-line treatments for multiple myeloma. Even though it has improved the overall survival rate of multiple myeloma, most patients eventually become resistant to it, resulting in treatment failure. Secondary lines of treatment currently used in the clinic were not specifically optimized in the context of Bort resistance. Instead, these regimens were often based on observed enhanced synergy in established Bort-sensitive multiple myeloma models. As researchers and clinicians learn more about the complex signaling pathways that interact to drive cancer progression, the need for more effective drug combinations becomes more apparent. Coupled with the huge pool of combinations and permutations for combinatorial drug regimens, rapid optimization of personalized regimens is impossible with the current methods of regimen design. In this study, an AI-based technology platform, QPOP, used small experimental datasets to develop new drug combinations against drug-resistant multiple myeloma. Because of the efficiency of this platform, QPOP was also able to identify patients that may be more sensitive to these optimized drug combinations from experimental data with patient samples.

QPOP overcomes hurdles associated with conventional drug combination design by identifying optimal drug–dose combinations in an experimental data-driven deterministic manner independent of molecular mechanistic assumptions. This platform is based on the notion that the influence of higher-order terms is negligible in adaptable and robust complex systems, including complex biological systems.^2–4 QPOP thus identifies the most optimal drug combinations without any prior knowledge of the system of interest using a second-order linear regression analysis. This approach includes the linear, dual, and quadratic contributions of the drugs on the system of interest. By mapping the quantitative phenotypic effect that a set of drug combinations have on a specific patient sample or disease model, QPOP can quickly identify the best drug combinations for that specific patient sample or disease model, as projected from the parabolic response surface maps. When applied toward Bort-resistant multiple myeloma, QPOP identified a combination of decitabine (Dec) and mytomycin C (MitoC) that outperformed standard-of-care regimens in Bort-resistant multiple myeloma xenograft mouse models. When QPOP was applied toward ex vivo patient samples, QPOP was able to identify which primary patient samples responded better to Dec/MitoC than standard-of-care regimens. The parabolic response surface maps of this Dec/MitoC combination from the three preclinical stages of QPOP analysis were also different, highlighting the importance of recalibrating drug combinations along the entirety of the drug development pipeline. This is a departure from current methods where single-drug ex vivo chemosensitivity assays are extrapolated onto drug combinatorial regimens based on the assumption of drug additivity.^5,6 Collectively, these show the robustness of QPOP in identifying and optimizing drug combinations that target drug-resistant multiple myeloma, as well as the potential to personalize combination therapy.

Extrapolating from this work, QPOP could be harnessed as a clinical support decision platform. Considering interpatient heterogeneity, the diversity in patient response treatment remains a huge hurdle. Because of the efficiency and speed of QPOP ex vivo primary tumor analysis, QPOP may be a useful tool in helping clinicians to determine the most effective drug combination regimen from a small amount of patient sample. We hypothesize that QPOP-optimized drug combinations will likely collect into distinct groups of patients. Given enough patient samples analyzed, the results of QPOP-directed drug combination ranking could lead clinicians toward biomarkers that identify subsets of patients that are more responsive to specific treatment regimens. Further implementation of QPOP and other small dataset-driven AI platforms into clinical studies related to combinatorial therapy should provide greater insight into the efficacy of these platforms as personalize medical tools.

This study presents QPOP as a comprehensive optimization platform with a range of applications from drug development to personalized medicine. The ability to identify optimal drug combinations on the specified system of interest rapidly, in a rational and deterministic manner, with significant reductions in time and cost, is ideal for both the clinic and the laboratory. As drug development trends toward more specific molecularly targeted therapeutics, identifying the most effective drug combination for these new drugs will become increasingly vital to their success in getting clinical approval. Furthermore, as more drug combinations become approved for clinical use, identifying the best drug combination for each patient will prove more and more difficult by conventional methods. QPOP appears to be able to address both of these problems through the ability to use a small number of tests from a large-parameter space to maximize drug combination efficacy and safety independent of disease mechanism or predetermined drug synergy. AI platforms such as QPOP, which can maximize the use of small datasets directly from patient samples or patients themselves, will ultimately improve the efficacy and outcomes of clinical drug combination development and personalized medicine.