HAI-VECT(SCD): AI-Humanoid Enabled Virtual Clinical Trial for Sickle Cell Disease

1a. Updating the aiHumanoid Simulation to v8.2

The previous version 8.1 of the aiHumanoid [8] underwent modest updates to v8.2. The main differences are that the revised version includes new simulations of Quality of Life (Physical and Psychological) and important updates to Erythropoiesis, and RBC physiology subsystems. The total number of integrated organoid simulations remains at 21. Regarding literature validation of the SCD simulations, the same approach used in version 8.0 and 8.1 was applied to the new and updated simulations comprising v8.2. A detailed overview of this process is summarized in Appendix 1.

1b SCD Validation Profile in the aiHumanoid Simulations

To confirm the presence of SCD in the aiHumanoid simulations, the following 8 phenotypic features assembled from the literature were evaluated: HbA, HbF, Quality of Life-Physical, Quality of Life-Psychological, Reticulocytes, RBC Hemolysis, Pain, and Vaso-Occlusion Events (VOE) in addition to the HBB LOF mutation.

The final panel of SCD disease markers was determined by statistical analysis using a Z Score estimate of >3 when patient features profile changes from control were positive direction and <-5 when they were negative. The final panel of predictors for the healthy virtual patient profile and 25 virtual SCD patients were compared first with a correlation coefficient and a Bonferroni corrected p value <0.006 to determine whether SCD patient profiles were significantly different from the healthy controls.

As before [8] the group of 25 virtual SCD patient profiles was created by GPT4 from recent literature to cover a wide range of patient demographics and phenotypic characteristics than could be achieved with a single common patient simulation. All patient profiles include Age, Gender, Risk factors, Symptoms, HBB mutation status and VOE history. A representative group of virtual patient profiles is summarized in Appendix 2.

2. Rationale for the drug combinations

The validated aiHumanoid based SCD simulations were used to evaluate the therapeutic response of each of the 25 virtual patients to six unique two drug combinations. The current standard of care for serious SCD is generally supportive plus hydroxyurea. Hydroxyurea was approved by the FDA for the treatment of adults with Sickle Cell Anemia (SCA) on December 31, 1998, and it was later approved for use in children with SCD in 2017. The approval was based on evidence showing its efficacy in increasing fetal hemoglobin levels and reducing the frequency of vaso-occlusive crises.

This paper describes a virtual pivotal clinical trial demonstrating the efficacy of hydroxyurea in reducing the frequency of painful crises in patients with Sickle Cell Anemia, which contributed to its FDA approval.

The rationale for evaluating two-drug combinations of currently approved drugs for treating Sickle Cell Disease (SCD) is driven by the need for improved therapeutic strategies to alleviate the manifold complications of this chronic disorder [1]. SCD is characterized by several pathological processes including hemolysis, vaso-occlusion, inflammation, and endothelial dysfunction, and often requires a multifaceted treatment approach [9].

Each of the four approved drugs targets different aspects of the pathophysiology of SCD, with potential synergistic or additive effects when used in combination. For example, hydroxyurea induces fetal hemoglobin, which can inhibit the polymerization of sickle hemoglobin [3], while Crizanlizumab primarily acts to reduce the frequency of painful crises by blocking P-selectin-mediated multicellular adhesion [4]. Combining diverse mechanisms of action has the potential to treat the disease more comprehensively, potentially enhancing efficacy and improving patient outcomes in terms of reducing vaso-occlusive events, alleviating pain, and improving quality of life [5,6]. Evaluating these combinations in a rigorous scientific manner can provide valuable insights into the most effective multidrug regimens, paving the way for personalized treatment approaches and potentially unveiling novel therapeutic paradigms for managing SCD.]. The selection of the specific drug combinations was based on the individual drugs’ distinct modes of action, the potential to target different aspects of SCD pathophysiology simultaneously, and preliminary evidence from clinical settings suggesting enhanced therapeutic benefits when used in combination.

3. Study Design and Objectives

This is the first of its kind, aiHumanoid based multi-arm, dose-escalation virtual clinical trial. The primary objective is to determine the potential ability of unique two drug combinations to reduce vaso-occlusive events (VOE)compared with standard of care, hydroxyurea, in virtual patients with SCD. Important secondary outcomes include cumulative toxicity, pain relief, and impact on quality of life. An overview of the study design is presented in Table 1a.

The virtual patients: The profiles for 25 unique virtual patients diagnosed with SCD were synthesized by GPT-4, OpenAI’s advanced language model (September 2023 version) https://chat.openai.com/. GPT4 used its extensive training data encompassing medical literature, patient profiles, and related clinical information, to synthesize diverse representative patient backgrounds, histories, genetic markers, and disease stages. This trial design permits us to create a virtual clinical trial with 6 (groups)X4(doses)X25 virtual patients) the equivalent of 600 patients. The virtual patients used in this study serve as hypothetical but commonly encountered examples of SCD but do not represent real individuals or precise medical histories.

(Note: GPT-4, developed by OpenAI, is a state-of-the-art language model capable of generating human-like text based on extensive training data. In this study, it was employed to generate virtual patient profiles, harnessing its vast knowledge from medical literature to ensure the profiles’ accuracy and relevance.)

Inclusion Criteria:

1. Diagnosis of SCD: Confirmed diagnosis of homozygous sickle cell anemia (HbSS) or compound heterozygous sickle cell disease (HbSC, HbSβ^0, or HbSβ^+ thalassemia).

2. Age: Individuals aged 12 years or older.

3. History of VOE: Patients must have experienced at least one VOE in the 12 months prior to inclusion.

4. Data Availability: Availability of sufficient historical clinical data to enable meaningful analysis.

5. Stable Treatment Regimen: Participants should be on a stable treatment regimen for at least 3 months prior to inclusion, with no changes to any other concurrent medication for SCD intended to prevent VOEs.

Exclusion Criteria:

1. Other Hemoglobinopathies: Individuals with other hemoglobinopathies or medical conditions that might confound the study results.

2. Pregnancy or Lactation: Women who are pregnant, planning to become pregnant, or breastfeeding during the study period.

3. Recent Blood Transfusion: Patients who have received a blood transfusion within 30 days prior to the start of the study.

4. Renal or Hepatic Dysfunction: Individuals with significant renal or hepatic dysfunction that might impact drug metabolism or excretion.

5. Severe Concurrent Illness: Any severe concurrent medical condition or illness that, in the investigator’s judgment, could compromise participant safety or compliance, interfere with consent, study conduct, or interpretation of study results.

6. Prior use of Study Drugs: Patients with prior exposure to the study drugs in combination within 3 months before the start of the study.

7. Inability to Comply: Individuals deemed unable to comply with study procedures and requirements.

4. Assessment of Response to Therapy Using a Specific Biomarker Subset

The optimal composite biomarker profile of the 8 features defined above plus multiorgan toxicity was used to assess the response to therapy in simulated SCD patients. Changes in the levels of these biomarkers post-treatment with the two drug combinations, relative to the hydroxyurea treatment group, serve as indicators of therapeutic efficacy.

Virtual patient demographics

The gender specific characteristics of the 25 virtual SCD patients are summarized in Table 1b. The data indicate that the patients are well matched. There was only one statistically significant difference between male and female attributes at the p = 0.05 level. The virtual male patients appear to be heavier than female subjects.

The SCD feature profile

Data for the 25 virtual patients were generated and analyzed for predictive value by comparing the profile with that from an unaffected aiHumanoid. A Z test was used to identify a subset of 8 features best predicted the presence of SCD. A two tailed paired T test was then used to confirm the predictive value of the final profile. The original p value of 0.05 was modified for multiple comparisons using the Bonferroni correction. This process created a corrected p value of 0.006. Using this method 8 of the original 14 features were selected for inclusion in the final disease profile. These eight factors are HbA, HbF, QoL-Physical, QoL-Psychological, Reticulocytes, Hemolysis, Pain, and VOE. A summary of this analysis appears in Table 2.

Dose 1: 25% of Maximum Dose

The initial phase of formal testing started with a dose equivalent to 25% of the maximal dose for all drugs in both treatment arms. We utilized Cohen’s d to estimate the treatment effect. A two-tailed paired T-test was also used to determine the significance of the predicted differences observed between the hydroxyurea treatment group and the two drug combination groups. While T-test p-values provided insight, our primary metric for discerning a true treatment effect was Cohen’s d estimate, as recommended by sources including [10,11]. When Cohen’s d 95% Confidence Interval (CI) crosses 0 it is interpreted to mean that there is no effect.

For assessing toxicity, our analysis mirrored the above approach, but a Z-score was added to gauge hydroxyurea toxicities in comparison to the predicted toxicities for the other two drug treatment groups. A Z-score of ≥2 was set as a benchmark to identify potential toxic outliers in the two drug treatment groups. Using this approach, any treated patient with a Z-score of ≥2 was considered to have experienced a dose-limiting toxicity (DLT).

Summary: The combination of hydroxyurea plus Crizanlizumab at 25% of maximum dose produced a small to medium beneficial treatment effect on VOE based on Cohen’s d values (95% CI: -0.61+/- 0.34).

Summary: Based on Cohen’s d there is a trend toward a reduction in cumulative organoid toxicities involving multiple drug combinations. These Cohen’s d values suggest a trivial to small-medium beneficial effect across the drug combinations at 25% of maximum dose. The most effective combination at reducing cumulative toxicities appears to be Endari plus Crizanlizumab based on Cohen‘s d value (95% CI: -0.36 +/- 0.20).

Summary: Importantly, no individual organoid toxicity measurement including the cumulative organoid toxicity estimate deviated by more than one standard deviation from the mean predicted toxicity for the hydroxyurea group.

Assessment: The analysis above reveals possible small reductions in anticipated organoid toxicities between the treatment groups and hydroxyurea. In the absence of evidence of DLT, we proceeded to increase the dosage for all drug combinations to 50% of the maximal dose in all groups.

Dose 2: 50% of maximum dose

The analysis of all patient data for all treatment groups was conducted in an identical manner to that outlined above for Dose 1 Efficacy and Toxicity.

Summary: The combination of Endari plus Crizanlizumab at 50% of maximum dose produced a medium beneficial treatment effect on VOE with a Cohen’s d value of -0.53. A close second was the combination of Voxelotor plus Crizanlizumab with Cohen’s d value of -0.51.

Summary: Based on Cohen’s d there is a trend toward a reduction in cumulative organoid toxicities involving multiple drug combinations. These Cohen’s d values suggest a trivial to small-medium beneficial effect for most of the drug combinations at 50% of maximum dose. The most effective combination at reducing cumulative toxicities appears to be Endari plus Crizanlizumab with a Cohen ‘s d value of -0.72 (95% CI: -0.72+/-0.41). This suggests at least a medium beneficial effect on cumulative toxicity.

Summary: Again, no individual organoid toxicity measurement including the cumulative organoid toxicity estimate deviated by more than one standard deviation from the mean predicted toxicity for the hydroxyurea group.

Assessment: In the absence of evidence of DLT, and the indication of medium reduction in VOE, we proceeded to increase the dosage for all drug combinations to 75% of the maximal dose in all groups.

Dose 3: 75% of maximum dose

The analysis of all patient data for all treatment groups was conducted in an identical manner to that outlined above for Dose 1 and Dose 2 Efficacy and Toxicity.

Summary: The combination of Endari plus Crizanlizumab at 75% of maximum dose produced a medium beneficial treatment effect on VOE with a Cohen’s d value of -0.69 (95% CI: -0.69+/-0.39).

Summary: Based on Cohen’s d there is a strong trend toward a reduction in cumulative organoid toxicities involving multiple drug combinations. These Cohen’s d values suggest a small to large beneficial effect for most of the drug combinations compared to hydroxyurea at 75% of maximum dose. The most effective combination at reducing cumulative toxicities appears to be Endari plus Crizanlizumab with a Cohen ‘s d value of -0.82 (95% CI: -0.82+/-0.47). This suggests a large effect on reducing cumulative toxicity compared to hydroxyurea.

Summary: The combination of Endari plus Crizanlizumab at 75% of maximum dose produced a large beneficial treatment effect on cumulative toxicity with a Cohen’s d value of -0.82.

Assessment: In the absence of evidence of DLT, and the indication of medium reduction in VOE, we proceeded to increase the dosage for all drug combinations to 100% of the maximal dose in all groups.

Dose 4: 100% of maximum dose

The analysis of all patient data for all treatment groups was conducted in an identical manner to that outlined above for Dose 1, Dose 2, and Dose 3 Efficacy and Toxicity.

Summary: The combination of Endari plus Crizanlizumab at 100% of maximum dose produced a large beneficial treatment effect on VOE with a Cohen’s d value of -0.87 (95% CI: -0.87+/-0.39).

Results Summary

aiHumanoid based virtual clinical trials vs Mendelian Randomized (MR) virtual clinical trials

Mendelian Randomization (MR) is a valuable statistical method that uses genetic variants, like SNPs, as instrumental variables to infer causal relationships [21]. Due to the random assignment of genetic variants at conception, MR mirrors properties like randomized controlled trials, addressing issues such as confounding [22]. This method harnesses natural genetic variation to explore the causal role of a modifiable risk factor on disease [23]. Challenges of MR include pleiotropy, linkage disequilibrium, and the potential for imprecise estimates of causal effects [24]. Moreover, MR isn’t designed to encapsulate dynamic system behaviors like feedback loops [25]. The ongoing development of MR is striving to resolve most of the current shortcomings of MR without contravening its basic assumptions [26].

In comparison, FCM based aiHumanoid simulations and the DeepNEU database have a foundational reliance on concepts, which can include genes and other biological entities, and their causal relationships. While the specific intricacies of the aiHumanoid simulations and the DeepNEU database are proprietary, the general principles of Fuzzy Cognitive Maps (FCMs) and their application in biological systems have been discussed in the literature [7,8,27]. FCMs are directed cyclic graphs that capitalize on unsupervised learning while asynchronously updating relationship weights [32]. Feedback loops, epigenetics, and miRNAs are seamlessly integrated as concepts within FCMs and similar AI simulations [27]. Pleiotropy is counteracted in sophisticated computational simulations by allowing multiple gene interactions [26,28]. The issue of pleiotropy is addressed in the aiHumanoid simulations in that each gene concept has more than 9 inputs from and directed outputs to other genes and phenotypic features [7.8]. Both methodologies demand substantial data inputs, and the constant evolution of data platforms is a testament to the rapid advancements in the field [29].

We conclude that the data supports the assertion that both FCM/DeepNEU and MR stand as pivotal technologies for the simulation of early-stage clinical trials. While MR provides valuable insights through genetic variants, allowing for causal inference in scenarios where randomized controlled trials may not be feasible, the aiHumanoid system offers a more dynamic and comprehensive approach. By simulating direct interventions, exploring dosing levels, and predicting individual virtual patient responses, aiHumanoid bridges the gap between early-stage clinical trial requirements and practical applications. Therefore, when we compare the static nature and lack of pleiotropy of MR with the versatile aiHumanoid simulations, the latter emerges as more aligned with the needs and focus of Phase 1 and 2 clinical trials.