Harnessing the power of AI-based models to accelerate drug discovery against immune diseases

KEYWORDS:

• Artificial intelligence
• immune disease
• machine learning
• precision medicine
• virtual patients

1. Introduction

Traditional drug discovery often involves laborious and time-consuming processes. However, emerging artificial intelligence (AI)-based models offer new ways to expedite the design and development of novel therapeutics, with numerous ongoing applications to the management of immune diseases (IDs). Herein I discuss the current impact and future prospects of AI driven approaches in this field, particularly through the implementation of a precision medicine approach taking into account patient specificities.

AI aims to develop intelligent machines based on artificial neural networks, inspired from primate brain physiology. AI represents a convergence of technologies that recapitulate several dimensions of human intelligence, namely perception, analysis, action, and learning, and more recently expressive communication, with generative AI like ChatGPT [1]. Together with machine learning, AI allows for the extraction of patterns in patient data to create predictive models to inform human decisions. Applied to IDs, AI can produce models used to categorize patients, choose relevant therapeutic targets, identify and optimize interesting drug candidates, and select patients likely to benefit from the drug (Table 1).

Table 1. AI-based predictive models applied to immune diseases.

Predictive models	Applications to immune diseases
Representation of the heterogeneity of chronic and acute- immune disorders	Disease models integrating multi-omics profiling data from large cohorts to stratify patients with either Systemic Lupus Erythematosus, primary Sjögren’s syndrome, Rheumatoid Arthritis, Inflammatory Bowel Diseases in homogeneous subgroups. With the emergence of generative AI form based on large language models, patients can also be categorized following text mining of their medical records, thereby improving diagnostic of many IDs, including primary immunodeficiencies and hematologic malignancies. Identification of specific molecular signatures as biomarkers to categorize patients. Models of the acute respiratory distress syndrome to identify and manage patients with a risk of developing severe forms of COVID-19.
Selection of relevant therapeutic targets	Models of autoimmune diseases documenting a complex pathophysiology, involving multiple dysregulated pathways, both disease-specific and shared across several diseases. Network computing analyses to establish causal disease models of Systemic Lupus Erythematosus, primary Sjögren’s syndrome and Systemic Sclerosis, highlighting the components of the perturbed system predicted to be involved in disease causality. Various models of COVID-19 to identify targets for drug repurposing.
Identification and optimization of drug candidates	Models predicting in parallel multiple properties (binding, pharmacological and ADMET properties) of small synthetic molecules applied to various IDs. Deep learning methods to estimate how well monoclonal antibodies bind to the target, how specific they are, and how much they activate the target, given their amino acid sequences and structures. This also involves determining how close they are to natural immunoglobulins to reduce the risk of immune reactions and improve their developability. Machine learning approaches to predict drug toxicity (eg carcinogenicity) or drug interactions. Models to select combination therapies of interest to treat autoimmune diseases.
In silico evaluation of drug performance in virtual patients	Virtual patients created by AI-enhanced QSP to predict the efficacy of an anti Interferon type I monoclonal antibody in cutaneous Lupus. QSP models are constructed from data from different sources, relating to biological processes in blood or tissue, clinical symptoms, and pharmacokinetic/pharmacodynamic features of the drug candidate. Digital twins as replicas of the human immune system, under physiological or pathological conditions. Causal disease models of Lupus Erythematosus and primary Sjögren’s syndrome used for computational perturbation of the disease interactome to mimic drug activity.

2. Fostering precision medicine for chronic immune diseases

IDs represent a very diverse group of disorders characterized by dysregulation of the immune system, resulting in aberrant or altered immune responses [2,3]. They encompass both autoimmune diseases when the immune system recognizes its own tissues as foreign and attacks them, as well as inflammatory bowel or skin diseases, in which the immune system is chronically hyperactivated, possibly due to commensal bacteria. It further includes allergies and asthma in which the immune system overreacts to substances usually not harmful, primary immunodeficiencies linked to inborn genetic errors, as well as hematologic malignancies affecting the immune system, such as leukemias and lymphomas. In addition, numerous infectious diseases can lead to severe perturbations of the immune system, with a recent example being COVID-19 [4,5].

Beyond the obvious differences between the above IDs in terms of pathophysiological mechanisms involved, those various conditions share many features. The latter include a complex etiology and heterogeneous nature consequent to the natural history of these chronic and recurrent diseases [3,6]. AI-modeling is useful to represent patient heterogeneity frequently observed within a given disease, both in terms of dysregulated immune compartments, organ involvement, systemic manifestations, and responses to existing treatments.

As the diagnosis and treatment of many IDs remain challenging, there is also a growing interest in precision medicine approaches [7–10]. To this aim, AI is being used to create disease models to categorize patients into homogeneous clusters based on shared underlying pathophysiological mechanisms, with the goal to design and implement targeted therapies well adapted to defined patient subsets (Table 1). The various predictive models based on AI enabling better targeted therapeutic approaches in autoimmune and inflammatory bowel diseases, as well as in asthma and allergies are discussed below. As of today, AI-based models of primary immunodeficiencies and hematologic malignancies are mainly considered to facilitate diagnostics [11,12].

3. AI-based models to support better targeted therapies for immune diseases

3.1. Representation of the complexity of immune disorders

To implement precision medicine, the first step is to produce models to understand and represent patient heterogeneity. Recent advances in high-throughput multi-omics technologies allow for the complete molecular profiling of cohorts of thousands of patients to document dysregulated biological processes in blood and target organs [6,8,9]. With the maturation of AI and machine learning, the integration and analysis of these high-dimensional data is now possible, a challenge that could not previously be met with conventional bioinformatics. From this, researchers can stratify patients into homogeneous subgroups and identify distinctive molecular signatures associated with each individual cluster. These signatures may eventually serve as biomarkers for precision medicine approaches, not only to support patient selection, but also to predict disease progression, treatment response and safety risks.

3.2. Selection of relevant therapeutic targets

Following patient stratification, a comprehensive mapping of genes or proteins that are dysregulated in blood or target organs relative to healthy controls is performed for each individual cluster. Pathway enrichment methods allow the assembly of disease-associated molecules into canonical molecular pathways [1]. Subsequent network computational analyses are then performed to identify among dysregulated pathways those predicted to be involved in the causality of the disease. To this aim, an interactome can be created that documents levels of influence of each single gene or protein on all other components of the perturbed biological system (Table 1). This allows the identification of hubs, master regulators and driver mutations critical in regulating the whole disease system, suggesting a causal role in the disease [13].

3.3. Identification and optimization of drug candidates

After selecting therapeutic targets, AI is then used to design and optimize new drugs against these targets. AI-enhanced computational chemistry relies upon quantitative structure – activity relationships (QSAR) to predict target-binding and pharmacological activities among virtual libraries of small chemical molecules. Artificial neural networks are trained to associate and predict pharmacological properties of billions of possible drug molecules, based on their chemical structure [1]. With AI, chemists can predict both the capacity of these molecules to mediate an agonistic or antagonistic activity, their ADMET properties (absorption, distribution, metabolism, excretion, toxicity), as well as their solubility and stability. Whereas artificial neural networks are becoming broadly used to predict the properties of small chemical molecules and synthetic antisense oligonucleotides, they are also applied to monoclonal antibodies, an interesting therapeutic modality for IDs (Table 1) [14].

Furthermore, recent modeling studies of various immune diseases, including Systemic Lupus Erythematosus, Sjogren’s disease, inflammatory bowel diseases and allergies have revealed that multiple proinflammatory molecular pathways are often dysregulated in patients. This raises a strong interest in developing drug combinations to treat patients with such complex IDs [15]. Computational methods can thus be used to design and select combination therapies to be tested in confirmatory clinical studies, with a better chance of success.

3.4. In silico trial simulation in virtual patients

AI can help in evaluating how effective and safe drug candidates for IDs are. AI enables designing clinical trials, choosing patients as well as integrating and analyzing data from the studies. In addition, AI has recently been used to simulate clinical trials by predicting in silico drug performance [13]. To this aim, virtual patients are generated from electronic health records or real-world data obtained from actual patients. Alternatively, virtual patients are also generated by using AI-enhanced quantitative system pharmacology (QSP) models of autoimmune diseases to estimate the effectiveness of experimental drugs (Table 1) [13].

A different method is based on digital twins, a concept that was first introduced by the aerospace and automotive industries to create a virtual reality that mimics a physical system of components and how they interact. In health applications, digital twins typically represent a model of a specific organ that is derived from medical imaging and physiological data at multiple levels (i.e. cell, tissue, whole organ). Digital twins, for example, in the form of a living human heart model, do not use extensive molecular data so far, and are therefore more suitable to evaluate medical devices that have a mechanical effect rather than drugs that affect biological processes [13]. Whereas it remains extremely challenging to model the complexity of the immune system, first steps have been undertaken to create an immune digital twin to replicate immune responses in both healthy individuals as well as patients with IDs [9].

Lastly, AI-based causal models of the pathophysiology of IDs can be used to mimic drugs interacting with the most influential drivers of the perturbed biological system. To this aim, the potential effect of a drug on a target is replicated by either abrogating or enhancing computationally the influence of the targeted gene, while further assessing whether the disease system moves toward the interactome related to healthy controls [13]. Importantly, although virtual patients as well as computational models based on system biology are mainly used today to predict drug efficacy, they can also help to evaluate safety risks related to drug candidates that target known adverse outcome pathways.

4. Expert opinion

Many concrete applications of AI-based models are currently transforming our approach to the management of IDs (Table 1), including:

• AI-based predictive models used to (i) stratify patients into homogeneous clusters, (ii) represent the pathophysiology as a perturbed biological system with inferences of causality, (iii) design and optimize drug candidates or combination therapies with desired properties, and (iv) evaluate the efficacy and safety of drug candidates in virtual patient models [3,6,8,9].

• Evolution toward a computational precision medicine to design more personalized treatments against IDs. This evolution driven by big data and advanced computational analytics will accelerate the discovery and development of more personalized treatments to the benefit of the patients [1,7–9].

Several points should be considered when implementing AI-based models to inform drug development against IDs, including:

• Enhancing the reliability of computational models by verifying the data that is used to train predictive algorithms for quality and completeness, adhering to good simulation practices, and validating the predictive outputs with experimental data from preclinical and clinical trials. This will improve the trustworthiness of those models and eventually facilitate acceptance of predicted outputs by regulatory authorities.

• In this regard, clear guidelines fulfilling standardized requirements should be established jointly by regulators and researchers, with respect to data management, curation, and preprocessing, as well as best practices of model validation.

• Better and more predictive models will be needed to overcome the current complexity to integrate massive, multimodal and high-dimensionality data (e.g. multi-omic, imaging, environmental and lifestyle data) [13].

• Specific models are also needed to represent ‘healthy’ immune responses, e.g. in the context of natural exposure to infectious pathogens or vaccines, while taking into account inter-individual variability. Such models will be useful as a benchmark to determine pathological alterations specifically linked to IDs [9].

5. Conclusion

AI today enables considerably strengthening hypotheses regarding the selection of therapeutic targets, the choice of optimal drug candidates, and the identification of patients to whom the drug will ultimately be targeted. In light of these potentialities, AI is poised to significantly accelerate the evolution toward a precision medicine in all therapeutic areas, including IDs. With the perspective of better targeted and more personalized treatments against these complex diseases, these current and future advances can greatly improve the lives of patients with various IDs by offering better adapted treatments, which should prove more effective and better tolerated.

Declaration of interest

The author is an employee at Servier. The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

Reviewer disclosures

Peer reviewers in this manuscript have no relevant financial or other relationships to disclose.

Additional information

Funding

This paper was not funded.