Artificial intelligence and open science in discovery of disease-modifying medicines for Alzheimer’s disease

Summary

The high failure rate of clinical trials in Alzheimer’s disease (AD) and AD-related dementia (ADRD) is due to a lack of understanding of the pathophysiology of disease, and this deficit may be addressed by applying artificial intelligence (AI) to “big data” to rapidly and effectively expand therapeutic development efforts. Recent accelerations in computing power and availability of big data, including electronic health records and multi-omics profiles, have converged to provide opportunities for scientific discovery and treatment development. Here, we review the potential utility of applying AI approaches to big data for discovery of disease-modifying medicines for AD/ADRD. We illustrate how AI tools can be applied to the AD/ADRD drug development pipeline through collaborative efforts among neurologists, gerontologists, geneticists, pharmacologists, medicinal chemists, and computational scientists. AI and open data science expedite drug discovery and development of disease-modifying therapeutics for AD/ADRD and other neurodegenerative diseases.

Graphical abstract

This paper presents a comprehensive, authoritative, critical, and accessible review of general interest to the drug discovery and development community concerning use of data science (genomics, transcriptomics, proteomics, metabolomics, interactomics, radiomics, and phenomics [electronic health records]) and artificial intelligence (AI) tools for Alzheimer’s disease (AD) and AD-related dementia (ADRD).

Introduction

Alzheimer’s disease (AD) is a neurodegenerative disorder characterized by multiple pathophysiological processes.¹ AD and AD-related dementias (ADRDs) are a major global health challenge, with an estimated 6.7 million Americans age 65 and older living with AD/ADRD today, and the AD population in the US alone is expected to reach 13.8 million by 2050.² New and highly effective treatments for AD are a crucial clinical need. Unfortunately, the attrition rate for AD clinical trials remains 98%,^3,4 and the costs of clinical trials are rapidly rising. A monoclonal antibody targeting aggregated amyloid, lecanemab⁵ (Leqembi), was approved as the first disease-modifying medicine by the US Food and Drug Administration (FDA) in 2023. Recently, donanemab significantly slowed cognitive and functional decline in a large phase 3 study of early AD.⁶ However, the need for monthly infusions and safety vigilance (such as amyloid-related imaging abnormalities and brain bleeding) may challenge its widespread use.^7,8 The amyloid cascade explains many findings in AD, but amyloid-β is important in early AD and not late in AD pathogenesis. Consequently, anti-amyloid drugs are unlikely to act after symptom onset. Thus, it is essential to identify additional, amyloid-independent drug targets and affordable oral medication for the treatment of symptomatic AD.

The high failure rate of randomized controlled clinical trials in AD is due in large part to a lack of understanding of the pathophysiology of AD and its therapeutic modulation.⁴ This deficit may be addressed by applying artificial intelligence (AI) to “big data” to rapidly and effectively expand therapeutic efforts in new directions, predominantly outside of the prevailing amyloid hypothesis that has historically driven the field.⁹ Recent rapid growth in computing power and memory storage, an unprecedented wealth of big biomedical data, and development of advanced analytical processing algorithms are leading to significant breakthroughs in AI.^10,11 These advances hold promise for optimizing clinical trials and drug development in AD, as well as guiding the field to novel therapeutic targets. Indeed, AI is already improving drug discovery and design of clinical trials,^9,11,12,13 in some cases delivering new candidate therapeutics within months rather than the many years typically required for traditional discovery methods.^11,12 Thus, broad application of AI tools holds promise to accelerate research and development of new AD medicines (Figure 1) while reducing cost and aiding in strategic clinical design.

Figure 1.

Open multiome data-driven precision medicine and personalized treatment in Alzheimer’s disease (AD)

Integration of the multiome datasets, including genome, transcriptome (including single-cell transcriptomics), proteome, metabolome, phenome, radiome, and the human interactome (i.e., protein-protein interactions), is essential for development of precision medicine and personalized treatment in AD using artificial intelligence approaches. In theory, the researchers can build AI or other in silico models (including machine learning and network medicine models), including cell-type-specific models (such as microglia, astrocytes, and neurons), from the multiome profiling datasets. These in silico models can be used for patient stratification, target identification, drug discovery/repurposing, and personalized treatments for AD and other complex brain disorders if broadly applied.

Open big data for AD target identification

The recent increase in generation of “omic” data, including genetics, genomics, epigenomics, transcriptomics, proteomics, metabolomics, lipidomics, radiomics, and phenomics, as well as data digitalization in patient care and the pharmaceutical sector present both challenges and opportunities (Figure 1). For example, the growing availability of big data prompts challenges for personalized clinical diagnosis and treatment of AD, based on the FAIR (findable, accessible, interoperable, and reusable digital objects) principle. These challenges have motivated the application of advanced AI and other in silico tools to help scientists identify promising drug targets and accelerate progress toward effective treatments for AD and other challenging diseases. The details for target identification from genetic and genomic findings, including AD,¹⁴ can be found in recent reviews.^15,16

Databases that coalesce high-throughput experimental data are crucial for developing AI-based solutions to identify drug targets, disease pathobiology, and a drug’s mechanism of action (Table S1). High-throughput DNA/RNA sequencing technologies have generated robust genomic data in multiple national/international genome sequencing projects in AD, including the Alzheimer’s Disease Sequencing Project (ADSP),¹⁷ the Alzheimer’s Disease Neuroimaging Initiative (ADNI),¹⁸ and the Genetics of Alzheimer’s Disease Data Storage Site (NIAGADS; https://www.niagads.org). The NIAGADS is a national genetic and genomic data repository for AD research that has collected 102 datasets, 183,003 samples, and over 3.3 billion genotypes (accessed September 27th, 2023). The ADSP, launched in 2012, has sequenced and analyzed genomes to identify a wide range of risk and protective genetic variants in AD, including whole-exosome sequencing (WES) data from ∼20,000 samples, as well as WGS data from over 36,000 samples.¹⁷ The Accelerating Medicines Partnership-Alzheimer’s Disease (AMP-AD) was formed to identify biologically relevant therapeutic targets, as well as new biomarkers. AMP-AD data and analysis results are stored in the AMP-AD Knowledge Portal (https://agora.ampadportal.org) and include genomic, transcriptomic, proteomic, metabolomic, and phenotyping data from more than 15,000 subjects. The AD Knowledge Portal¹⁹ contains upwards of 100,000 data files from over 80 studies of people with AD and related model organisms. In addition, several human longevity translational projects, such as the Longevity Consortium, Long-Life Family Study, Longevity Genomics, and Integrative Longevity Omics, provide multi-scale omics and aging-related datasets (Table S1), which could translate into possible therapeutic target discovery for aging-related disorders, including AD/ADRD.

AI for AD drug discovery

In the process of drug discovery, AI exploits a theoretical framework of graphical models that represents biological and chemical structures (Figure 2). These AI and graphical models characterize different data science domains and promote development of mathematical algorithms. This section describes emerging AI tools for accelerating drug repurposing and emerging therapeutic development in AD.

Figure 2.

Molecular design and optimization using artificial intelligence (AI) and machine learning (ML) approaches in drug discovery and development

(A) Typical molecular representation using graph methods, including SMILES-based molecular presentation for building encoder and decoder models.

(B) Molecular property prediction and optimization using AI and ML approaches. The users can build AI/ML models from molecular representations (A) to evaluate various drug biological endpoints, including drug-target binding affinities, drug-gene expression profiles, drug’s pharmacokinetics properties (i.e., brain penetration), and others.

(C) A diagram illustrating the theoretical process of graph-based molecular representation and learning process.

De novo drug discovery

There has been substantial recent progress in de novo drug design and optimization through AI and deep learning techniques, facilitating strategic design of novel molecular structures both de novo and based on information gleaned from large databases (Table S1), such as ZINC²⁰ and GDB-17.²¹ This technology has brought a new direction to an otherwise vast chemical space, aiding in the design of molecules with desired properties, such as high binding affinity with protein targets and low predicted toxicity.²² Generative models,^23,24,25 such as variational autoencoders, generative adversarial networks, and normalizing flows, have been applied to generate realistic and diverse molecules with drug-like properties and ideal brain penetration properties (Table S2). An international team proposed a generative tensorial reinforcement learning (GENTRL) model for de novo small-molecule design.²⁶ Via optimizing synthetic feasibility, novelty, and biological activity under the GENTRL model,²⁶ the team identified potent inhibitors of discoidin domain receptor 1 (DDR1), a kinase target implicated in fibrosis and neurodegenerative diseases.²⁷ In addition, an AI-designed dual-targeted 5-HT1A agonist and 5-HT2A antagonist by Exscientia (UK), DSP-0038, is under a phase 1 trial for possible treatment of AD psychosis.²⁸ The details for de novo drug discovery using AI tools can be found in a recent review.^11,29

Retrosynthesis prediction

Once a desired molecular structure is identified, the next question is whether it can be readily synthesized from already-existing molecules. This is known as retrosynthesis prediction, which infers a set of reactants for classification into two categories: (1) template-based approaches³⁰ and (2) template-free approaches.^31,32 Templates are transformation rules involving different substructures or functional groups, which are usually derived from hand-coded rules from domain experts or automatically extracted with algorithms. For template-based approaches, the key problem for retrosynthesis prediction is selecting the optimal template for a query molecule. Recently, a conditional graph logic network (CGLN) was proposed for ranking the templates for a query molecule.³⁰ While this provides interpretable approaches, it is slow and unable to generalize to reactions outside of the domain knowledge. Thus, there is growing interest in developing template-free approaches. Liu et al. have formulated retrosynthesis prediction as a machine translation problem by representing molecules as simplified molecular-input line-entry system (SMILES) strings³¹; Shi et al. proposed a graph-to-graph framework (G2Gs) for retrosynthesis prediction by treating both the product and reactant molecules as graphs.³² A new chemical synthesis strategy, termed “pharmacophore-directed retrosynthesis,” has been reported to generate a natural product, gracilin A, for neurodegenerative diseases like AD.³³

Optimization of pharmacokinetics (PK) and pharmacodynamics (PD)

A fundamental challenge in drug discovery and lead optimization is evaluating molecular characteristics of drug candidates, such as ADME (absorption, distribution, metabolism, and excretion) and brain penetration (i.e., blood-brain barrier [BBB] penetration) properties. Traditionally, molecules are represented as molecular fingerprints³⁴ that portray the presence or absence of particular substructure molecules. Deep learning has recently helped to enable major progress in this arena. Gómez-Bombarelli et al. have used recurrent neural networks to learn molecular representations with the SMILES strings as input³⁵; yet, SMILES strings are unable to capture the complex relationships among atoms, and more efforts to represent molecules as graphs and graph neural networks are underway.^36,37,38 Graph neural networks are specific types of neural architecture designed for graph-structured data. Gilmer et al. proposed a general framework called the message passing neural network (MPNN) to extract molecular representations.³⁶ Here, the feature representation of an atom is iteratively updated by combining the representations of its neighbors with its own representations.³⁶ These architectures are usually trained in a supervised fashion, which requires a large number of molecules with known properties. Sun et al. proposed an approach called InfoGraph for learning molecular representations in an unsupervised and a semi-supervised fashion, which reduces the number of labeled molecules required for training.³⁹

ImageMol, an unsupervised pretraining deep learning framework was pretrained on 10 million molecular images of unlabeled drug-like, bioactive molecules to predict molecular targets and properties (including PK/PD) of candidate compounds.⁴⁰ ImageMol outperformed existing state-of-the-art approaches in various benchmark datasets. Recent methodological advances in the integration of machine learning and mechanistic models gave rise to another set of powerful tools that can be used to infer mechanism-of-action, safety, efficacy, and PK/PD profiles of candidate molecules.⁴¹ In AD drug development, poor BBB penetration is possibly one of the critical causes for the high failure rate of clinical trials.⁴² To address this challenge, scientists utilize AI and machine learning models (Table S2) for assessment of BBB properties^43,44,45,46 before experimental assays, which can be applied to AD drug development programs.

Drug repurposing/repositioning

Drug repurposing/repositioning reduces the time and cost of drug development,⁴⁷ in particular for challenging diseases, such as AD/ADRD⁴⁷ and coronavirus disease.^12,48,49 For example, anti-inflammatory agents for arthritis might be repositioned for development of treatment for AD/ADRD.⁵⁰ There are ∼50 repurposable drug trials (∼40 unique repurposed agents) based on the latest AD drug development pipeline (2023).⁵¹ To expand drug repositioning efforts, a recent machine learning-based framework named DRIAD (drug repurposing in AD) has been developed to quantify the potential associations between AD biological processes and linked genetic datasets, thereby prioritizing drug candidates for repurposing.⁵² DRIAD prioritized baricitinib as a candidate AD drug, and baricitinib is being tested in an open-label, biomarker-driven basket trial (ClinicalTrials.gov: NCT05189106) in people with AD and amyotrophic lateral sclerosis (ALS).

In addition to DRIAD, AlzGPS is a systems biology platform with over 100 multi-omics datasets that capture molecular profiles underlying AD pathobiology⁵³ (Table S1). This tool enables network-based discovery and prioritization of potential targets for AD drug repurposing.⁵³ NETTAG is a network-topology-based deep learning framework to identify AD-associated genes and prioritize candidate drugs. There are four steps to utilize NETTAG for AD target and drug discovery: (1) assemble noncoding AD genome-wide association study (GWAS) loci effects on expression quantitative trait loci (eQTLs), protein QTLs, and other types of QTLs; (2) cluster protein-protein interactions (PPIs) from databases (Table S1) into multiple functional network modules; (3) quantify a node’s (gene’s) score by integrating its functional similarity to each gene identified with multiple brain-specific gene regulatory evidence via influencing GWAS loci and prioritize likely causal genes for AD by their aggregated gene regulatory features; and (4) prioritize repurposable drugs for potential treatment of AD by evaluating network proximities between NETTAG-inferred AD gene products (proteins) and known drug targets (such as AlzGPS⁵³; Table S1) under the human protein-protein interactome network model. Using NETTAG,⁵⁴ the team successfully identified that gemfibrozil (an approved lipid regulator) was significantly associated with reduced risk of AD compared to simvastatin using an active-comparator design from real-world patient data. Other deep learning methodologies (deepDR⁵⁵ and deepDTnet⁵⁶) were developed for new target identification and drug repurposing in heterogeneous drug-gene-disease networks (Table S1).

Via analysis of real-world health insurance claims from 7.2 million patients from the IBM MarketScan Medicare Supplemental database, two FDA-approved p300/CBP inhibitors, salsalate and diflunisal, were found to be associated with decreased incidence of AD, and neuroprotective efficacy was also validated in mice.^57,58 Using an endophenotype-based in silico network medicine approach,⁵⁹ a team identified that sildenafil usage was significantly associated with reduced likelihood of AD and further experimentally validated the findings using an AD patient induced pluripotent stem cell (iPSC)-derived neuron model.⁵⁹ Another team demonstrated that bumetanide (an FDA-approved oral diuretic) provides a potential treatment of apolipoprotein APOE4-related AD.⁶⁰ The details of computational methods for AD drug repurposing can be found in recent reviews.^47,61,62

Real-world-data-derived drug development

Although biofluid and genetic biomarkers are more important in defining the pathologic process underlying AD/ADRD,^63,64,65,66 real-world data (RWDs), such as electronic health records (EHRs) or insurance/pharmaceutical claims, contain practice-based evidence for patient care and management as well. In recent years, there has been substantial research on building AI models for extracting insights from RWDs to facilitate drug development.⁶⁷

Clinical trial emulation

Randomized controlled trials (RCTs) are the gold standard for drug development. However, due to the stringent definition of eligibility criteria, the number of eligible trial participants is typically small. This makes the trial cohort “ideal” rather than representative of real-world patient populations. RWDs are comprised of practice-based observations from real-world patients and thus provide a valuable resource for drug development. However, RWDs are challenged by confounding factors, like gender, race, and lack of detailed clinical, biomarker, and genetic information (i.e., APOE genotypes). Recently, there have been attempts to replicate the treatment effects obtained from RCTs in RWDs with the help of AI. RWD cohorts with features like those of the trial participants were identified, and if individual data were available for RCT participants, then propensity score matching approaches were applied. Due to complicated confounding situations in RWDs (e.g., high-dimensionality and temporality), conventional logistic regression approaches are unable to estimate the propensity scores with high accuracy. However, high-dimensional propensity score estimation, doubly robust methods, and deep learning approaches can resolve this problem through the target trial emulation method.^68,69 This method was applied in a study of millions of claims and EHR and found zolpidem (an approved anti-insomnia medicine) as a candidate drug in slowing or reversing Parkinson’s dementia.⁷⁰ If individual data are not available in the trial, then several alternative approaches can be applied to match the summary statistics of the RWD population to ones in the trial. One such example is matching-adjusted indirect comparisons (MAICs),⁷¹ in which exponential weights are calculated on RWD patients to match the features reported in specific trials. In addition to treatment efficacy, safety is an important consideration for RCTs. The feasibility of using RWDs to simulate RCTs⁷² has been demonstrated with trial ClinicalTrials.gov: NCT00478205 (phase 3 double-blind, parallel-group trial comparing 23 mg donepezil sustained release with 10 mg donepezil immediate release formulation in moderate to severe AD), in which RWDs from the OneFlorida Clinical Research Consortium⁷³ were used to develop reasonable estimations of the safety of RCTs on both one- and two-arm trials.

RWDs can also help broaden participation in clinical trials by relaxing eligibility criteria. Liu et al. explored this idea in trials for lung cancer in which they collected the eligibility criteria for a set of selected trials and evaluated the individual impact on the clinical endpoint (mortality) of eliminating the criteria serially.⁷⁴ Shapley value, a game theoretic technique that has been used for black-box machine learning model interpretation, was adopted for quantifying the impact for each criterion. The authors demonstrated that dropping certain eligibility criteria may not impact treatment effectiveness but can greatly enlarge the number of eligible trial participants. Similar approaches hold promise for expanding subject inclusiveness in AD trials.

Subphenotyping

Like other complex diseases, AD is highly heterogeneous. Subphenotyping refers to identifying subgroups of patients with AD with similar phenotypic or molecular characteristics (Figure 3). This is critical for understanding the clinical complexity of AD and designing effective customized treatment plans. Within the field of AI, the problem of subphenotyping is usually referred to as the patient clustering or segmentation problem, for which the goal is to identify patient clusters such that patients within the same cluster exhibit more homogeneous characteristics and disease progression profile.⁷⁵ The core problem is how to define appropriate patient similarities. Current approaches define such similarities based on patient representations, such as Euclidean distance for vectors,⁷⁶ dynamic time warping for temporal sequences,⁷⁶ or more advance temporal models capable of discovering latent disease states.^77,78 The recently developed EHR transformers such as BEHRT⁷⁹ and med-BERT⁸⁰ are examples of models that capture the temporal nature of longitudinal patients records. EHR transformers belong to the same family of deep generative models mentioned above. By applying the models to patient data that scale across a range of diseases, not only can predictions on multiple downstream tasks be achieved but novel, meaningful patient presentations are also derived. For example, the study of the BEHRT transformer predicts occurrence of dementia in the next 6 months after last visit with a relatively high performance.⁷⁹

Figure 3.

A diagram illustrating a framework for AI-based drug repurposing and rational drug combination design via targeting molecular networks of multiple proteinopathies (including Aβ and tau) and microglia abnormalities (disease-associated microglia, inflammation microglia, interferon microglia, and other microglial subtypes) derived from multi-omics profiles

Each network graph represents molecular network built from multi-omics profiles from individuals with different proteinopathies (i.e., amyloid and tau) or microglia abnormalities (i.e., inflammation, stress, and proliferation, major histocompatibility complex [MHC] class I and II, interferon microglial subpopulations, and other types of microglia) in various public databases, such as the Alzheimer’s Disease Knowledge Portal, The Alzheimer’s Cell Atlas, and protein-protein interactome networks (Table S1).

One promising direction in AD subphenotyping is understanding the differential effect of drugs on different disease subtypes. With the techniques mentioned above, AD subphenotypes can be defined from RWDs,⁷⁶ and then the average treatment effect of individual drugs or drug combinations within each AD subphenotype can be estimated (Figure 1). This can generate precisely stratified treatment effect estimations as well as refine clinical trial eligibility criteria to improve efficiency and efficacy. In addition to clinical-data-based subphenotyping, multi-modal analysis of omics data can be employed. Using multi-modal analysis of more than 1,500 transcriptomes from patients with AD, researchers identified three major subtypes of AD characterized by different combinations of molecular signatures, regulator genes, and dysregulated pathways,⁸¹ emphasizing the importance of precision medicine drug discovery in AD/ADRD (Figure 3). For example, a recent study identified brain-region-specific molecular changes associated with late-onset AD (LOAD) using multi-omics analyses.⁸² Another study presented a multi-modal deep learning approach to incorporate imaging (magnetic resonance imaging [MRI]), genetic, and clinical data to classify patients into AD, mild cognitive impairment (MCI), and healthy controls; they demonstrated that integrating multi-modality data outperforms single-modality models.⁸³ A similar multi-modal machine learning framework that integrates clinical and neurocognitive data, structural MRI, and polygenic risk scores was created to predict patients with psychosis in AD, a symptom frequently associated with AD/ADRD.⁸⁴

Clinical image analysis

As AD is a neurodegenerative disorder, neuroimaging plays a vital role in characterizing the disease state and progression.¹⁸ Different types of neuroimaging, including MRI, fluorodeoxyglucose positron emission tomography (PET), amyloid PET, tau PET, and single-photon emission computed tomography, have been leveraged in AD research to characterize phenotypic features of the brain.⁸⁵ These neuroimaging data contain disease information that is complementary to information contained in EHRs. In the context of AD/ADRD drug discovery, there are two types of applications: (1) using neuroimaging endophenotypes to identify endophenotype-specific targets or pathways for personalized drug discovery and drug repurposing studies and (2) utilizing neuroimaging biomarkers to serve as outcome measures or input features for stratifying patients and evaluating drug efficacy or side effects during clinical trials. A recent study presented network-based disease-progression-specific drug repurposing for AD based on neuroimaging-derived disease stages.⁸⁶ Using three types of PET brain imaging for microglial activation ([¹¹C]PBR28), amyloid-β (Aβ) ([¹⁸F]AZD4694), and tau ([¹⁸F]MK-6240), a recent study showed that the co-occurrence of Aβ, tau, and microglia abnormalities was the strongest predictor of cognitive impairment.⁸⁷ This study highlighted that synergistically targeting microglial abnormality, tau, and Aβ endophenotypes could offer more clinical benefits compared to targeting each endophenotypes alone, supporting previous endophenotype-based drug repurposing in AD.⁵⁹

Co-occurrence of proteinopathies (including Aβ, tau, TDP-43, and α-synuclein) and microglia abnormalities is a common pathophysiology of AD/ADRD.^87,88,89 There are multiple microglial subpopulations involved in microglia abnormalities.⁹⁰ Multi-omics can be used to repurpose known drugs able to target microglia subpopulations. Using multi-modal analysis of single-cell/-nucleus RNA sequencing data from AD patient brains, two approved asthma drugs (fluticasone and mometasone) were found to be significantly associated with reduced likelihood of AD by targeting AD-associated microglia.⁹¹ Combination therapies via targeting molecular networks or pathways derived from multi-omics data from individuals with well-characterized proteinopathies and microglia abnormalities may offer more effective treatment approaches compared to monotherapies targeting one protein/microglia subtype. Researchers can build specific molecular networks for each proteinopathy or AD-relevant microglial subtype by integrating PPI network findings from multi-omics data available from the AD Knowledge Portal or TACA⁹² and prioritize repurposable drugs for each specific proteinopathy or microglia abnormality.⁹³ Researchers will then prioritize drug combinations that synergistically block major proteinopathies and microglia abnormalities (Figure 3) to identify more effective treatment for patients with AD/ADRD. More details about AI-based or network-based drug combination design can be found in recent studies.^94,95,96

A major role for imaging biomarkers in current drug development programs is to evaluate target engagement and provide evidence in support of disease modification for clinical trials of disease-modifying therapies.⁹⁷ For example, the pivotal trials of aducanumab included patients with early AD who had confirmed brain amyloid using amyloid PET.⁹⁸ In addition, MRI has been recommended prior to initiating therapy of aducanumab, during the titration of the drug, and at any time the patient has symptoms suggestive of amyloid-related imaging abnormalities.⁹⁸ There are large national and international collaborative consortiums, such as the ADNI⁹⁹ and Enhancing NeuroImaging Genetics through Meta-Analysis (ENIGMA),¹⁰⁰ aimed at studying large cohorts with longitudinal neuroimaging, genetics, and clinical profiles for use in modeling clinical trials.

Language and other digital markers

To address concerns about privacy and reliability in obtaining important clinical data, lower-cost and more frequently obtained markers of patient status with digital technologies have been developed. For example, a chatbot with reinforcement learning technologies to identify patients with potential MCI through conversations with clinicians has been developed.¹⁰¹ Another example is the demonstration of machine learning technology ability to detect patients with AD from voice recordings, relying on subtle changes in patients’ patterns of speech due to a language disorder that might not be detected by human ears.¹⁰² Additional efforts are underway in developing sensor technologies to monitor health and activities at home. The digital markers derived from sensor data can encode human behavioral information in daily life, which correlate with health status.¹⁰³ Event-related potentials (ERPs) and electroencephalograms (EEGs) are technologies that provide information-rich data in AD and dementia research and diagnosis.^104,105 For example, the P300 brain potential is sensitive to AD processes during its early stages,¹⁰⁴ while EEG is commonly used to assess disease progression.¹⁰⁵ DeepMAge, a deep learning DNA methylation aging clock, accurately predicts biological age relevance to exact ages of individuals with various age-related healthy conditions.¹⁰⁶ Spatial navigation has become an important factor in identifying preclinical AD, and using spatial navigation big data (i.e., the Sea Hero Quest game) offers useful tools toward personalized cognitive diagnostics of at-genetic-risk AD.¹⁰⁷ Multi-sensory gamma stimulation approaches ameliorate AD-associated pathology and improve cognition.¹⁰⁸ In addition, digital technologies such as wearable sensors and mobile apps can monitor medication adherence to improve efficacy of clinical care and clinical trials.¹⁰⁹

Discussion, perspective, and future directions

Addressing challenges in de novo drug discovery

Modern machine learning can address data gaps that currently impede successful drug development. Existing machine learning pipelines for de novo drug discovery typically rely on machine learning to provide the properties of new compounds in the search process. These predictors are trained with a small set of biologically labeled data and need to be regularly updated as new labeled data become available. This is expensive and time consuming. To accelerate the process, meta-learning facilitates interactively querying a user to label new data points for training high-quality machine learning predictors. Graff et al. utilized this approach to prioritize molecule screening in virtual drug screening,¹¹⁰ providing supportive evidence for generating informative labeled datasets to enable training high-quality model predictors. Closing the loop between data generation and model training is a promising future direction, and there is growing industry interest in utilizing robots to automatically collect high-quality data.¹⁰

Critically, drug discovery depends on efficiently searching the vast chemical space, and the optimal approach is rapidly evolving. Existing search algorithms based on reinforcement learning usually generate molecules atom by atom. However, in traditional drug discovery, drugs are designed by combining different meaningful fragments, which are functional groups identified by domain experts. Existing molecular design and optimization algorithms mainly focus on single-property optimization, while molecules satisfying multiple properties are usually sought. A multi-modal machine learning infrastructure (Figure 4) could offer potential solutions to effectively conduct multiple-property optimization as well as to search meaningful molecular fragments and synthesize promising molecular candidates using a robotics-based platform.¹¹¹

Figure 4.

A proposed multi-modal ML infrastructure for precision medicine drug discovery in AD

Traditional reductionist paradigm overlooks the inherent complexity of human diseases and has often led to treatments that are inadequate or fraught with adverse effects. Existing data resources, including genomics, transcriptomics, proteomics, metabolomics, and interactomics (protein-protein interactions), have not yet been fully utilized and integrated to explore the roles of AD drug discovery. A multi-modal ML infrastructure can provide comprehensive toolboxes for precision medicine drug discovery in AD: (1) multiome-driven target identification; (2) disease subphenotyping to accelerate clinical trial testing of disease-modifying treatments; (3) automatic drug discovery and development (including de novo molecular design), pharmacodynamics and pharmacokinetics (in particular brain penetration) property evaluation and optimization (such as absorption, distribution, metabolism, excretion, and toxicity [ADMET] properties), and robotics-based chemical synthesis, selection of candidate molecules for preclinical evaluation using various AI and ML approaches (such as deep generative models, reinforcement learning, and graph networks); (4) real-world patient data observation to verify clinical efficacy; and (5) selection of promising molecules for personalized clinical trial evaluations.

As drug development is a complex process involving many steps, multi-modal machine learning tools can significantly reduce the time and cost of drug development. For instance, multi-modal machine learning approaches^83,84 improve accuracy of patient subphenotyping during clinical trial design by assembling neuroimaging, genetic, and multi-omics profiling data. With the help of deep learning, effective representations can be learned for different data modalities,^26,112 which can then be fused by simple concatenation or more complicated nonlinear transformation¹¹³ to perform downstream tasks such as molecular design, PK and BBB property evaluation and optimization, and robotics-based chemical synthesis, which can greatly accelerate the drug discovery and development process. If broadly applied, AI-based tools will accelerate the development of disease-modifying treatments for AD.

Practical and methodological challenges for AI-readable data

Effective drug discovery pipelines require multi-omics, multi-modal (imaging, EHR, biofluid markers, etc.), and in vivo and in vitro validation in both animal and human models. AI can facilitate each step involved in this complex procedure/pipeline. However, there are several practical and methodological challenges to collecting, curating, and applying these multi-modal omics and clinical data. Clinical and multi-omics data harmonization plays an essential role in facilitating the quality of AI models in real-world drug development, which is challenging in most current AI-based drug discovery projects. Second, data sharing is a barrier for AI-based drug discovery. Unique integration of biomedical data from both industrials and public domains, such as pretraining models and federated learning technologies, may significantly increase the impact of AI-assistant drug development. In vivo and in vitro validations of AI-predicted targets and drugs are still challenging due to the lack of highly reproducible experimental models for AD/ADRD research. Population-based drug-AD/ADRD outcome analyses from massive EHR or health insurance claim data offer potential validation approaches for AD drug repurposing findings.^59,91,96

Reducing ethnic disparity and increasing inclusiveness in clinical trials

Because of the complexity of AD, large-scale and heterogeneous datasets covering genomic, cellular, clinical, and behavioral aspects of study subjects are needed to advance disease understanding (Figure 1). Important sources of heterogeneity in AD include age, gender, race, and ethnicity.^114,115 Collaborations across different entities and institutions therefore need to involve diverse patient populations as well as participants with diverse disease characteristics.¹¹⁶ This will ensure inclusiveness in trial participation, diversity of trial data, and equity for trial opportunities and access to treatments that address the complex biological and social aspects of AD. AI can also be used in forward-looking applications to increase participant diversity by supporting decision-making, such as determination of the likely accuracy of diagnosis across ethnic groups and predicting side effects that might affect trial participation and retention.¹¹⁷

It is important to recognize that AI models may involve some components of bias in their training process since models tend to bias toward the data from which they were derived. These biases could disproportionately affect minority patients. In response, there have been significant efforts on methods to assess and mitigate bias in existing data.¹¹⁸ Going forward, it is crucial to construct diverse datasets when training an AI model. Effective execution of this task will require local, national, and international collaboration. For example, the global biobank meta-analysis initiative will power genetic and target discovery while reducing ethnic disparities.¹¹⁹ Several large-scale, ethnically diverse genetic studies—including African Americans,⁶⁵ Amish,¹²⁰ and Hispanic/Latinos¹²¹—have made a great impact on AD research and drug development.

Model underspecification is another potential concern for deep learning. This is because multiple routes exist to map the inputs to the same predictions, and the learned model from one dataset may not work well on another. For example, the model learned from patient data in clinical trial site A may not work well in site B due to the limited size of available training samples and high model complexity. This is a common situation for AI machine learning with local healthcare systems data.

In conclusion, development of transparent and comprehensive reporting guidelines for AI and machine learning models in medicine is vitally important,^122,123 as this will help identify the circumstances under which the model can be appropriately applied. Collaborations across different healthcare systems within large consortiums will increase the size of training datasets and improve the diversity of the training data, which would be expected to improve model generalizability.⁶⁷

Model generalizability

Due to the disease complexity, the AI and machine learning models are likely to suffer from “dataset shift”^124,125; for instance, the dataset used for model training is different than the data the model will be deployed on, which impacts model generalizability. A related problem is model underspecification,¹²⁶ which corresponds to the situation where multiple routes exist to map the inputs to the same predictions, and thus a learned model from one dataset may not work well on another. This is particularly true for complicated models such as deep learning due to the large number of model parameters and relatively small sample size, which is common for local healthcare systems data. This is another reason why we need open data science—collaborations across different healthcare systems within large consortiums will greatly increase the sample size of training datasets and improve the diversity of the training data, which would be expected to improve model generalizability.

Regulatory science challenges of AI implications

Data protection is key to inclusiveness and maintaining the trust of diverse participants. Data transmission standards, such as the Fast Healthcare Interoperability Resource (FHIR)¹²⁷ for clinical data transmission, are key to data credibility and confidentiality. Encryption techniques, such as homomorphic encryption and semantic hashing, are critical to protect sensitive patient data. Another potential solution is federated learning,¹²⁸ which protects individual patient data through collective learning from multiple local sites without transferring the original raw data. However, the general setup of federated learning still needs a central server to aggregate and update the model parameters, which may face the risk of information leak. Recently, a new paradigm called swarm learning^129,130 was proposed that combines edge computing, blockchain-based peer-to-peer networking, and coordination in a decentralized way without the need for a central coordinator. Specifically, all local sites are treated as participants in a block chain. A certain site will initiate a “transaction” through the smart contract mechanism when it updates the model parameter so that all other participants need to agree and transmit their local model parameters. Swarm learning makes the patient data collection and analysis more secure in a decentralized way, which is critical in clinical trials.

Another important direction is to improve transparency and interpretability of AI and machine learning models so that the user can understand the decision-making process and to ensure proper assessments of usability and potential failure cases. Specifically, model interpretability should be contextualized,¹³¹ with different levels of model transparency required for different applications. Focusing only on model transparency may sacrifice high-performance models in many scenarios and lead to misplaced trust in others. Model interpretation methods have potential uses.¹³² For example, knowledge distillation is a popular interpretation technique that aims at advancing a secondary interpretable model to approximate the prediction results of the first model.¹³³ These models could be vulnerable to adversarial attacks, which could manipulate the model explanations deliberately and make them unreliable. Deriving robust, reliable, and secure model interpretations is therefore an important future research direction.

AI-based acceleration of clinical trials

AI technologies combined with RWDs can help identify optimal participants for clinical trials and refine inclusion eligibility criteria for more efficient trial design.¹⁰⁹ There are at least three ways of achieving this goal. One is trial matching. Once the eligibility criteria are established, they can be transformed into computable phenotypes, and RWDs can be searched for identification of participants that match these target phenotypes. This approach increases diversity in trials by identifying potential candidates across ethnic groups. The potential challenges of this approach are that (1) not all eligibility criteria are computable and searchable and (2) some of the computable phenotypes are not readily available in the structured fields of RWDs. In this latter case, clinical natural language processing is needed to extract information from unstructured clinical notes. The second application of AI to RWDs involves patient similarity analytics, which can help identify similar patients according to RWDs. This increases the candidate pool by including subjects who do not exactly match the computable phenotypes transformed from the eligibility criteria. The third way in which AI can accelerate clinical trial design is to estimate the potential treatment effects from RWDs and then translate patient cohort characteristics into eligibility criteria and sample size estimates that improve trial success rates. For a new chemical entity, the potential treatment effects from similar drugs can be estimated.

Effective interpretation technologies (i.e., knowledge distillation) are needed to interpret patient characteristics as eligibility criteria.

Digital health technologies, such as mobile and wearable devices, can be used to collect patient-based outcomes during trials to better understand drug adherence, safety, and effectiveness. AI technologies for analyzing structured forms (such as surveys) and continuous physiological signals can also further automate this process and increase efficiency.

Open data science initiative for AD drug discovery

AI tools are becoming increasingly common in scientific discovery and AD research.^134,135 Despite the enthusiasm for AI-enabled drug discovery for AD, questions and issues abound. For decades, translational science has been facing a challenge: how to translate research findings into new effective medicines and how to couple these medicines to new technologies that enable efficient delivery. Inaccessible datasets mean that AI cannot contribute value by applying data mining and predictability. Most AI models in drug discovery pipelines require large volumes of data for training and validation, especially for deep learning application. Lack of adequate-quality data and robust sharing practices are key challenges in drug discovery, and inadequate data quality and lack of diverse populations can lead to AI models with poor generalizability, biases, and other adverse characteristics. Data harmonization that improves data quality and utilization via the use of domain knowledge and machine learning techniques thus provides a crucial contribution to the development and application of AI-based drug discovery.

To facilitate collaboration and open big data science, making machine learning models publicly available is essential. Funded by the Gates Foundation, the Alzheimer’s Disease Data Initiative (ADDI; https://www.alzheimersdata.org/) has provided a data-sharing platform that is advancing scientific breakthroughs and progress toward new treatments and cures for AD/ADRD. Data standardization and model standardization are equally important. In addition, reliable and robust model interpretation technologies are important. Yu and Kumbier¹³⁶ proposed the concept of “veridical data science,” in which predictability, computability, and stability are defined as the three cores of data science. Here, stability is key not just for prediction but also for interpretation.¹³⁷ Causal inference approaches¹³⁸ are empowered to infer hidden disease mechanisms.

To effectively identify AD-relevant information in RWDs, such as EHRs, computable phenotyping approaches and natural language processing are both critical. Because of the potential effort involved in the clinical note annotation process, active learning is promising to improve annotation efficiency.¹³⁹ The same strategy can improve the efficiency of labeling or annotating other types of data, such as neuroimaging.

How to effectively predict the chemical or biological properties of compounds is another key problem for drug design. However, as noted, it is time consuming and expensive to obtain the labels of compounds to train a predictor model. Therefore, an important future direction for AD drug design is to minimize the labeling efforts for training an effective property prediction model. There are a few potential directions for addressing this challenge. One is based on reinforcement active learning.¹⁴⁰ Instead of randomly picking a compound for labeling, for example, active learning techniques can select the most informative molecules for modeling (e.g., the ones with the largest uncertainty in prediction). The second direction is leveraging multi-task or transfer learning methodologies,^141,142,143 which aim to identify and leverage shared knowledge across different data to improve the model performance. Though the number of labeled molecules for each property is limited, a model for predicting multi-properties simultaneously can be created, or supervision from the source tasks can be transferred from abundant labeled molecules to tasks in which the labeled molecules are limited.

Clinical validation is another crucial element of successful AD drug development. There are different types of clinical and biomarker data, such as structured and unstructured fields in EHRs, cognitive assessments, neuroimaging, and physiological signals. Different data modalities contain different types of information that characterize patient health status, which could be helpful for understanding drug safety and effectiveness. To obtain rigorous evidence from clinical data, we envision a need for building a comprehensive multi-modal causal inference framework by integrating all heterogeneous information. As shown in Figure 4, a multi-modal machine learning infrastructure could provide comprehensive toolboxes for precision medicine drug discovery in AD. Users can utilize AI-assistant tools for AD/ADRD target discovery and validation from unique integration of multi-omics and clinical data (Table S1). Researchers and trialists may identify novel targets or test disease-modifying treatments in specific AD/ADRD subphenotypes (amyloid- or tau-PET-positive patients^5,6) based on patient subphenotyping findings from multi-omics, fluid biomarkers, and clinical neuroimaging. Importantly, researchers could conduct automatic drug discovery and development (including de novo molecular design), property evaluation and optimization (such as ADME and PK/PD properties), robotics-based chemical synthesis, and selection of candidate molecules for preclinical evaluation using various AI and machine learning approaches (Table S2; Figure 4), including deep generative models²⁹ and reinforcement learning technologies.¹⁴⁴ Finally, trialists will utilize large real-world patient data observation studies to verify clinical efficacy and select promising molecules for personalized clinical trial evaluations.

Outlook and summary

In summary, the application of AI solutions to AD drug discovery and development, as well as clinical trial design, is possible via cross-disciplinary, open science collaborations to address current gaps and challenges. If broadly applied, AI could facilitate decision-making for first-in-class AD drug discovery among academia, the pharmaceutical industry, and healthcare systems.¹⁴⁵ AI can accelerate AD drug development at the levels of molecular design and synthesis, choice of molecule, side effect and drug interaction prediction, and trial sample size requirements guided by clinical and biological heterogeneity, analysis of trial data, and application of trial outcomes to patients who may be most helped by the intervention (Figure 4). Altogether, AI is an indispensable aspect of the future of accelerating drug development and personalized medicine for patients with AD/ADRD.