IDentif.AI: Rapidly optimizing combination therapy design against severe Acute Respiratory Syndrome Coronavirus 2 (SARS‐Cov‐2) with digital drug development

Abstract The emergence of severe acute respiratory syndrome coronavirus 2 (SARS‐CoV‐2) led to multiple drug repurposing clinical trials that have yielded largely uncertain outcomes. To overcome this challenge, we used IDentif.AI, a platform that pairs experimental validation with artificial intelligence (AI) and digital drug development to rapidly pinpoint unpredictable drug interactions and optimize infectious disease combination therapy design with clinically relevant dosages. IDentif.AI was paired with a 12‐drug candidate therapy set representing over 530,000 drug combinations against the SARS‐CoV‐2 live virus collected from a patient sample. IDentif.AI pinpointed the optimal combination as remdesivir, ritonavir, and lopinavir, which was experimentally validated to mediate a 6.5‐fold enhanced efficacy over remdesivir alone. Additionally, it showed hydroxychloroquine and azithromycin to be relatively ineffective. The study was completed within 2 weeks, with a three‐order of magnitude reduction in the number of tests needed. IDentif.AI independently mirrored clinical trial outcomes to date without any data from these trials. The robustness of this digital drug development approach paired with in vitro experimentation and AI‐driven optimization suggests that IDentif.AI may be clinically actionable toward current and future outbreaks.

States, Singapore, Taiwan, Japan, European Union, India, and Australia. 4 After demonstrating promising open-label study results in China, 5 FPV has been approved in India and Russia for treatment of mild and moderate COVID-19 patients, with additional clinical trials (NCT04402203 and NCT04402203) have been initiated for further validation. The majority of trial outcomes are either pending or have not shown clinical benefit over standard of care (SOC) or placebo. As such, while drug repurposing enables rapid intervention against COVID-19, thus far, it has not led to clarity with regard to how to best treat this disease.
Traditional methods for implementing combination therapy and monotherapy based on drug repurposing rely on mechanism of action (MOA)-based drug selection and standard clinical dosing guidelines to achieve drug synergy and therapeutic efficacy. For example, a preclinical study showed that RDV as well as high-dose chloroquine (CQ) were efficacious toward Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) in vitro. 6 While this is an established approach that has led to promising candidate therapies, many of these regimens were not able to translate their in vitro outcomes into successful clinical results. Therefore, optimal efficacy that is clinically relevant is a different objective that presents substantial challenges to traditional drug screening and repurposing methods. For example, if candidate effective drugs are given in combination at suboptimal respective doses, resulting efficacy is moderate or even absent. At the same time, the relative doses between drugs within a combination can substantially impact treatment efficacy and toxicity due to unpredictable drug interactions. Another common hurdle is that repurposed drugs in vitro demonstrate the desired antiviral activity only at the high concentrations not achievable in a human body at safe dosing regimens. Therefore, drug dosing has a critical role in identifying which drugs belong in the optimal combination in the first place, and optimizing treatment outcomes, particularly in combination therapy, ultimately relies on simultaneously selecting the right drugs at the right respective doses. 7,8 Reconciling drug-dose parameters also requires leveraging unpredictable drug interactions in order to mediate maximal efficacy of combination therapies. Unfortunately, simultaneously pinpointing these parameters is an extraordinarily complicated task. For example, a parameter space of 1 trillion (10 12 ) possible combinations would be created from a pool of only 12 candidate therapies interrogated at 10 dose levels. This is an insurmountable barrier for traditional drug screening. Important studies have previously sought to leverage drug synergy interactions to predict multidrug combinations. 9 Other strategies have investigated higher order drug interactions to develop antimicrobial drug combinations. 10 Bridging these findings with clinical validation remains a challenge due to the size of the experimental search space.
In this study, we sought to overcome these challenges in developing effective combination therapies against SARS-CoV-2 infection using the IDentif.AI platform and an in vitro SARS-CoV-2 infection model with a live virus derived from a patient sample. IDentif.AI harnesses a quadratic relationship between clinically relevant therapeutic inputs (e.g., drug and dose) and biological outputs (e.g., quantifiable measurements of efficacy, safety) to experimentally pinpoint clinically relevant optimal combinations from large parameter spaces accounting for unpredictable interactions with a marked reduction in the number of required biological experiments ( Figure 1).
IDentif.AI is not purely computational and does not use preexisting training datasets. Instead, it uses an orthogonally designed set of calibrating regimens and in vitro experimentation to simultaneously identify effective drugs, their unpredictable interactions and corresponding, clinically relevant doses that optimize treatment outcomes from prohibitively large drug-dose parameter spaces that cannot be reconciled by brute force drug screening. 7,11 In effect, IDentif.
AI leverages these calibrating regimens to crowdsource SARS-CoV-2 live virus responses to experimentally drive the efficacy toward an optimal outcome. An earlier version of IDentif.AI was previously rapidly developed as a proof of concept strategy to pinpoint an optimal combination for vesicular stomatitis virus. 7 Here, we report a clinically actionable IDentif.AI with a streamlined workflow that incorporates clinically relevant dose design, an artificial intelligence (AI)-based strategy that prospectively and experimentally crowdsources the patientderived live SARS-CoV-2 virus to drive the optimization process, as well as a follow-on validation process that has resulted in a ranked list of drug combinations that are simultaneously optimized for drug composition and the dose of each respective therapy. This has resulted in results that broadly and independently align with clinical trial outcomes without requiring any data from these studies, thereby resulting in a platform that can be used as a first-line approach toward clinical decision support and therapeutic guidance with any number of additional drug options to address the COVID-19 pandemic as well as future outbreaks. In this study, this AI-driven digital medicine approach was applied to a 12-drug set of candidate therapies added to a cellular infection model to pinpoint unpredictable drug interactions and clinically actionable combination therapy regimens against the live SARS-Cov-2 virus isolated from a nasopharyngeal swab of a patient in Singapore. 12 The 12-drug set included a broad spectrum of repurposed agents that were evaluated in clinical studies for treatment of COVID-19 or were administered in conjunction with these therapies, including RDV, FPV, RTV, LPV, oseltamivir phosphate (OSV-P), DEX, ribavirin (RBV), teicoplanin (TEC), LST, AZT, CQ, and HCQ. Noteworthy, the drugs' concentrations were clinically relevant, that is, did not exceed one-tenth of the levels observed in the patient blood in response to standard dosing. Based on prior studies of minimal resolution experimental design, 3 clinically relevant dosing levels were employed with these 12 drugs, creating a combinatorial space of 531,000 regimens. 13 With a three-order of magnitude reduction in required tests, we identified a clinically actionable list of two-, three-, and four-drug combinations ranked based on viral inhibition efficacy in vitro with accompanying safety data against kidney epithelial cells (Vero E6), liver epithelial cells (THLE-2) and cardiomyocytes (AC16).
The identified drugs in the combinations were all at clinically relevant concentrations, not higher than one-tenth of the drug levels in blood in response to established clinical dosing. The top-ranked combination was comprised of RDV, RTV, and LPV which mediated a 6.5-fold increase in efficacy (viral inhibition %) compared to RDV alone due to an unforeseen drug interaction. Further demonstrating the clinical actionability of IDentif.AI, HCQ, and AZT combination was shown to be a relatively ineffective regimen in vitro at clinically relevant doses, mirroring recent clinical results. Importantly, the IDentif.AI-pinpointed relative efficacy of the combinations and monotherapies at the clinically relevant doses that did not use any preexisting antiviral clinical data was independently confirmatory of many of the clinical trial endpoints to date. These outcomes, coupled with the fact that foundational precursors to IDentif.AI have been clinically validated for infectious disease, oncology, and organ transplantation human studies, support the potential application of IDentif.AI as a clinical decision support platform for the optimized design of actionable combination therapy regimens. 14-16 2 | RESULTS

| Screening drug pool and experimental model
A pool of drug candidates was first chosen and evaluated for downstream IDentif.AI analysis and drug combination optimization. The pool of candidate therapies for IDentif.AI-driven optimization contained 11 drugs that were hypothesized to inhibit SARS-CoV-2 viral infection via affecting: viral entry into the host cell-CQ, HCQ, AZT, LST, TEC; viral replication-RTV, LPV; viral RNA synthesis-RDV, FPV, RBV; viral release-OSV-P. [17][18][19][20] To create combinations actionable within the current clinical guidelines we aimed to investigate drug interaction space between the antiviral and concomitant medications.
DEX has been proposed for treating acute respiratory distress syndrome resulting from COVID-19 (NCT04381936), LST is a common hypertension drug whose dosing should not be paused while undergoing COVID-19 treatment. 21 TEC is a wide spectrum antibiotic prescribed for pulmonary infections, potentially including those occurring as COVID-19-related complications. 22 IDentif.AI is a dynamic optimization AI-based platform that utilizes orthogonal array composite design (OACD), consisting of a resolution IV two-level (drug concentrations) factorial design and a three-level orthogonal array, to efficiently screen for influential factors and determine optimal drug-dosage combinations within the SARS-CoV-2 in vitro, cellular infectious disease model. Aliasing and F I G U R E 1 Project IDentif.AI workflow. Project IDentif.AI has four phases: (1) clinically relevant concentrations are established for each drug based on dose-response curves and maximal plasma concentration (C max ) of clinically administered dosages, (2) combination therapies determined with an orthogonal array composite (OACD) design are experimentally tested in an in vitro, cellular infectious disease (ID) model, (3) IDentif.AI analysis of the drug dose parameter space identifies drug-drug interactions and ranks optimal, clinically relevant drug-dosage combinations, and (4) biological validation of clinically relevant combinations designed by IDentif.AI-designed or already in trials confounding are addressed for each independent drug's linear, bilinear (drug-drug interaction), and quadratic effects by the resolution IV design, factor screening, and deterministic nonlinear relationships. 11,13,23 IDentif.AI interrogates drug-dose relationships in order to pin-   (Table 1). These results indicated low cellular effects of the selected monotherapies at the tested concentrations. No effect of the maximum vehicle concentration (0.1% DMSO) was detected on viral CPE inhibition or on cytotoxicity (Student's t test, N = 12, p > 0.05). The EC 50 and CC 50 of HCQ, CQ, RDV, FPV, and RBV were different from previously reported values, attributable to differences in the experimental conditions (e.g., SARS-CoV-2 strain, assays, incubation periods). 6,24 Regardless of the monotherapy antiviral activity, all drugs were considered for the combinatorial optimization process in order to identify possible unpredictable drug interactions that could markedly impact treatment efficacy and safety.
Accounting for a common source of failure in translating in vitro results to clinical trials, the high ratio of EC 50 to maximum plasma concentration (C max ) achieved in the human body, 25 C max was included as a crucial consideration for selecting drug concentrations at Levels 1 and 2 for each drug that ensure none of the drugs were overrepresented in relation to other drugs and to human pharmacokinetics (Table 1). Additionally, evidence has emerged that SARS-CoV-2 infection causes pathology of the vascular system and may require a treatment maintaining sustained drug level in the blood. 26,27 We examined C max for each drug as specified in clinical data after reaching a steady state at an established dosing regimen given to a population without drug metabolism impairment and at dosing regimens listed on a drug label specified by a national regulatory body for RTV given at a high dose of 600 mg twice daily (bid), with and without other antiviral drugs, varies between 11 and 14.7 mg/L. 30-32 LPV requires a pharmacokinetic enhancer. When given at 400/100 mg bid LPV/RTV, reported C max for LPV reaches 12.3-12.9 mg/L. 28, 33 The reported C max for RBV administered orally bid at a total daily dose of 800, 1000, or 1200 mg, was 4.23 mg/L at Week 4. 34 The reported C max for CQ was 0.73352 mg/L when given at an initial 450 mg dose followed by two 300 mg doses. 35 In accordance with the FDA label, HCQ reaches a C max of 2.436 mg/L after a single intravenous high dose of 310 mg. 36 The FDA label reported steady-state C max of AZT is 0.24 mg/L at a standard once daily (qd) 250 mg dose, following a 500 mg initial dose. 37  In order for IDentif.AI to determine optimized drug combinations from this 12-drug set, 100 drug-dose combinations were generated according to OACD (Table S1,      did not induce as much viral CPE inhibition as compared to RDV alone.

| Experimental validation of IDentif.AI results
These data confirm that IDentif.AI can accurately reflect the unsatisfactory outcomes observed in those clinical trials, without incorporating any prior clinical data or drug mechanism assumptions as inputs.

Drug-drug interactions were investigated with an additional
IDentif.AI interaction reanalysis of the OACD experimental data. % Inhibition IDentif.AI response surface plot mirrored well-documented and experimentally confirmed synergy between RTV and LPV ( Figure 3(a)). In contrast, IDentif.AI identified an antagonistic interaction between RTV and OSV-P (Figure 3(b)), a combination that is currently being investigated in clinical trials (NCT04303299). It is important to note that combining RDV with LPV only at clinically relevant concentrations, which to our knowledge has not been explored clinically as a registered trial, doubled their individual viral CPE inhibition when added together. Accordingly, the corresponding %Inhibition IDentif.AI response surface plot identified a previously unknown synergistic interaction between RDV and LPV (Figure 3(c)). Further confirming IDentif.AI rankings and validation experiments, it is important to note that the RDV/RTV interaction was not significant, but when given in three-drug combination, RTV boosted the RDV/LPV interaction almost two times (Figure 3(d)). These results further highlight the ability of IDentif.AI to leverage unexpected drug-dose interactions to identify optimal drug combinations at clinically relevant concentrations from a massive drug-dose search space.

| Design of drug combinations
Drug combinations for 12 drugs at three concentration levels (0, 1, 2) were generated using an OACD as described by Xu et al. 23 The OACD combines resolution IV two-level factorial design and a three-level orthogonal array to provide the least number of combinations required for factor screening of each independent drug's linear, bilinear (drug-drug interaction), and quadratic effects. 11,13,23 The resolution IV OACD used for this study had 100 combinations: 36 combinations based on the orthogonal array combined with 64 combinations based on the factorial design (Table S1, Supporting Information).

| SARS-CoV-2 virus
All experiments involving live virus were performed in a biosafety level-3 (BSL-3) laboratory. SARS-CoV-2 was isolated from a nasopharyngeal swab of a patient in Singapore with ethics approval and con- Briefly, the reagent was added into each well and incubated for 10 min at room temperature prior to measurement of luminescence readout using microplate reader (Tecan).

| Cell cultures
African green monkey kidney Vero E6 cells (C1008) were plated at 2 × were dissolved in sterile-filtered water.

| Viral inhibition and cell cytotoxicity of drug monotherapies
All virus infection experiments were performed in a BSL-3 laboratory.
The drugs were diluted in Vero E6 culturing media before dispensing into wells of 96-well plates. The laboratory staff performing the subsequent experimental work was blinded to the well content arrangement on the plates. The Vero E6 cells (2 × 10 4 cells/well) and media with and without SARS-CoV-2 treatment (100 TCID 50 ) were added to the plates containing the drugs and the controls. The drug concentrations ranged between: Z'-factor has been calculated to assess the assay quality across all experiments and in each viral experimental set to ensure it can generate reliable information. Z'-factor is a statistical coefficient that incorporates dynamic range and data variability of the positive and negative controls: where σ c+ and μ c+ represent the SD and mean of the luminescence signal of the positive control (control cells) and σ c-and μ c-represent the SD and mean of the luminescence signal of the negative control (cell + virus control), respectively. 0 < Z' < 0.5 represents a "do-able assay" and 0.5 ≤ Z' < 1 represents an "excellent assay." 56 Luminescence data were normalized to the average readout from the vehicle control cells on the same plates. Cytotoxicity and viral CPE inhibition were calculated as follows 57,58 :  IDentif.AI analysis correlated drug combinations experimental results into a second-order quadratic series. Each independent drug combination inhibition and monotherapy inhibition replicate was used in the optimization process. The second-order quadratic model is as follows: where y represents the desired biological response output (%Inhibition), x n is the nth drug concentration, β 0 is the intercept term, β n is the single-drug coefficient of the nth drug, β mn is the interaction coefficient between the mth and nth drugs, and β nn is the second-order coefficient for the nth drug, while m ≠ n. This second-order quadratic analysis and parabolic response surface plot analysis were conducted using the built-in "stepwiselm" function in MATLAB R2020a (MathWorks, Inc.

| Statistical analysis
All experiments were performed in at least triplicate biological repeats. To account for uncertainties propagated in the process of normalization, %Inhibition and %Cytotoxicity are presented as mean ± propagated SD, with the propagated SD derived from the following equation 60 : where and σ T and σ I represent the propagated SD for the mean value of %Cytotoxicity and %Inhibition, and σ c+ , σ c− and σ E+, σ E− represent the SD of the luminescence signal of the positive control (control

| CONCLUSIONS
Following the emergence of SARS-CoV-2, a global effort to clinically assess a broad spectrum of repurposed and novel compounds was initiated. In order to fully optimize the development of a treatment regimen against SARS-CoV-2 or any future epidemic/pandemic, it is important to move beyond traditional drug selection approaches, since mechanism-of-action-based drug selection alone will unlikely yield sufficient efficacy for broadly favorable clinical outcomes. This is because globally optimized combination design will rely on simultaneously optimal drug and dose identification, which is a major challenge for traditional drug screening and repurposing approaches due to an insurmountably large drug-dose parameter space. This work has addressed this challenge using IDentif.AI, an AI-based digital drug development platform that rapidly crowdsourced the patient-derived live virus to experimentally pinpoint and validate ranked combinations within 2 weeks. Unpredictable drug interactions were harnessed by IDentif.AI to pinpoint unforeseen, top-ranked combinations, and the IDentif.AI rankings independently aligned with broadly reported clinical trial outcomes without requiring data from these studies. Therefore, IDentif.AI can be potentially deployed as a first line of defense to rationally pinpoint optimal drug-dose combination therapy regimens for rapid clinical validation while also potentially deterring the assessment of regimens that are unlikely to yield suitable clinical outcomes.
Collectively, these capabilities may serve as a foundation for global accessibility to clinically actionable and optimized therapeutic responses to current and future pandemics.

DATA AVAILABILITY STATEMENT
The data supporting the findings of this study are available from the corresponding author upon reasonable request.