Application of computational methods in the drug discovery and development of Alzheimer's disease

Abstract

Alzheimer's disease (AD) is a prevalent neurodegenerative disorder marked by the accumulation of amyloid plaques and neurofibrillary tangles in the brain, leading to cognitive decline. As the most common dementia, AD affects over 50 million individuals worldwide, with 10 million new cases each year. Despite many significant progresses in AD research fields, there is still no effective treatment. Advances in molecular modelling and artificial intelligence (AI) algorithm have facilitated the application of computational methods in AD drug development, offering advantages in terms of efficiency, cost–effectiveness, and resource conservation. These strategies, including molecular docking, molecular dynamics simulation, protein structure prediction, virtual screening, and de novo drug design, have been widely employed in the past decades. Additionally, AI pharmaceutical companies have revolutionized the traditional drug development pipeline such as AD biomarker discovery, accurate diagnosis and precision medicine, etc. This review offers a thorough summary of recent progresses in computational-based methodologies in the study of AD drug development, including the elucidation the pathological mechanism at the molecular level, target identification, target 3D structure prediction, lead compound discovery, de novo molecular generation, multi-target directed ligands, drug repurposing, biomarker discovery, etc., showing promising applications of molecular modelling and AI in the development of AD therapy.

Graphical abstract

This review summarizes the cutting-edge research progress of computational methods in Alzheimer's disease drug development, covering areas such as biomarker/target identification, pathological mechanisms, lead compound discovery, drug repurposing, and precision medicine.

1. Introduction

Alzheimer's disease (AD) is a neurological ailment first identified by Alois Alzheimer in 1906¹. Clinically, AD is characterized by progressive memory loss, language and speech dysfunction, cognitive impairment, and severe behavioural abnormalities². Currently, over 55 million individuals globally are affected by dementia, a figure projected to rise to 139 million by 2050, with AD constituting 50%–60% of all dementia cases³. As the global population ages, approximately 9.9 million new cases of dementia are expected to emerge annually⁴. AD is primarily classified into two categories: sporadic and familial. Sporadic AD, which constitutes approximately 95% of all cases, typically manifests after the age of 65 and is referred to as late-onset AD (LOAD)⁵. This AD form is influenced by both environmental factors and genetic susceptibility, with the apolipoprotein E ε4 (APOE ε4) allele being a significant genetic risk factor^5,6. In contrast, familial AD, also known as early-onset AD, is predominantly driven by genetic factors, including mutations in the genes encoding amyloid precursor protein (APP) and presenilin 1 and 2 (PSEN1/2)^5,7. Currently, the etiological factors underlying these diverse pathological changes remain incompletely understood, and the precise mechanisms of pathogenesis are yet to be elucidated.

Despite substantial scientific endeavours, only six small-molecule drugs for AD have received approval from the U.S. Food and Drug Administration (FDA) since 1906⁸, among which are the acetylcholinesterase (AChE) inhibitors such as tacrine⁹, donepezil, galantamine, rivastigmine¹⁰, as well as the N-methyl-d-aspartate (NMDA) receptor antagonist memantine¹¹. The crystal structures of the complexes of these drugs and their targets have been resolved (Fig. 1), providing an important basis for structure-based drug design and discovery. However, these drugs are solely palliative treatment, providing symptomatic relief without offering a cure for AD. Currently, most of the AD drugs in clinical research belong to disease modifying therapy (DMT), accounting for about 76%. Furthermore, drug development targeting amyloid-beta (Aβ) and tau proteins remains the most prevalent area of study, comprising around 34% of DMT-related drugs¹². Notably, the recent approval by the FDA of monoclonal antibodies (mAbs) such as aducanumab, lecanemab, and donanemab, all of which target Aβ, marks significant advancements in the field of DMT13, 14, 15. These therapies can facilitate the removal of β-amyloid plaques from the brain. However, pharmacological interventions targeting amyloid have been associated with significant adverse effects, including brain swelling and amyloid-related imaging abnormalities (ARIA)¹⁶. Consequently, the identification and development of amyloid-independent therapeutic targets and cost–effective treatments for AD are of paramount importance.

Figure 1.

Chemical structures of FDA-approved small molecule drugs for the treatment of AD and their interaction mechanisms with targets at the molecular level. Drugs are represented by magenta sticks and the key residues for drug binding are represented by cyan sticks. Green dashed lines represent the pi–pi and cation–pi interactions between drugs and targets. Yellow dashed lines represent the hydrogen bonds formed between drugs and targets.

The development of drugs for AD has been characterized as a high-risk, high-investment, and high-failure research area¹⁷. A report published by the Pharmaceutical Research and Manufacturers of America indicates that between 1998 and 2017, 146 AD drug candidates failed in clinical trials globally, while only 4 drugs were successfully brought to market, resulting in a clinical success rate of merely 2.7%¹⁸. Traditional drug development processes often need extensive preclinical candidate screening experiments and costly clinical trials to ascertain drug safety and efficacy, rendering the endeavour both protracted and exceedingly expensive¹⁹. The emergence of computer-aided drug design (CADD), including virtual screening (VS), molecular dynamics (MD) simulation, de novo drug design, etc.²⁰, significantly advanced the field of rational drug design²¹. Additionally, the development of artificial intelligence (AI) technology has further propelled the progress of AI-driven drug discovery (AIDD)²², presenting new opportunities for this process²³. The conventional drug discovery process has been optimized through the utilization of advanced algorithms and data processing techniques, thereby enhancing both efficiency and success rates²⁴. Recently, both academia and pharmaceutical companies have utilized these computational strategies to achieve significant progress in AD drug research. Therefore, the purpose of this review is to comprehensively summarize the applications of computational strategies in target identification, biomarker discovery, three-dimensional structure prediction, lead compound discovery, de novo molecule generation, multi-target directed ligands, and drug repurposing, etc. In addition, we also briefly summarized some major breakthroughs achieved by AI pharmaceutical companies in clinical patient stratification, precision medicine, and AD disease diagnosis using AI technology.

2. Overview of the AD pathogenesis, therapeutic targets, and clinical drugs

The challenge in treating AD stems from the incomplete understanding of its pathogenesis. Over the past years, various hypotheses aimed at elucidating the underlying mechanisms of AD were proposed (as shown in Fig. 2). Among these, the cholinergic hypothesis (neurotransmitter)^25,26, the Aβ cascade²⁷, and the tau protein hypothesis²⁸, inflammation²⁹ are the most prominent and widely accepted. This section primarily summarizes the pathogenesis associated with AD and the potential therapeutic drug targets.

Figure 2.

Various AD pathogenesis and potential therapeutic drug targets. The figure was generated by Figdraw (www.figdraw.com).

2.1. Neurotransmission modifying agents for AD

In the AD clinical drug research, agents that modify neurotransmission primarily target the regulation of cholinergic, monoaminergic, glutamatergic neurotransmitters in AD patients²⁶. Among them, the cholinergic hypothesis proposed by Bartus is of great importance³⁰. Based on this mechanism, FDA has approved four AChE inhibitors for AD treatment, namely tacrine, donepezil, galantamine, and rivastigmine. More recently, FDA has also approved Zunveyl (benzgalantamine), a prodrug of galantamine, for the treatment of mild to moderate AD. In addition, the dysregulation of monoamine neurotransmitters, including 5-hydroxytryptamine (5-HT), epinephrine, norepinephrine, and dopamine, is also intricately associated with the pathogenesis of AD²⁶. Smith et al.³¹ reported a significant reduction in 5-HT levels in the brains of AD patients. Among the various subtypes of 5-HT receptors, the primary subtypes implicated in AD are 5-HT_1A, 5-HT_2A, 5-HT₄, and 5-HT₆^32,33. There is also a marked degeneration of glutamatergic neurons in AD patients, which disrupts the excitatory-inhibitory balance and may result in deficits in memory and learning. Elevated levels of glutamate and tau-induced phosphorylation can lead to the overactivation of NMDA receptors, resulting in excessive intracellular calcium influx and subsequent synaptic dysfunction. In 2003, FDA approved the NMDA receptor antagonist memantine for the treatment of moderate to severe AD, demonstrating its ability to enhance memory and cognitive function¹¹. Overall, pharmacological agents targeting neurotransmitters represent a prevalent approach in the treatment of AD. A significant proportion of currently available AD medications belong to this category, and approximately 22% of drugs undergoing clinical trials also focus on neurotransmitter modulation¹². However, these agents primarily alleviate symptoms rather than provide a cure for the disease. Given the intricate interplay of neurotransmitter systems in the brain, the development of multi-target drugs or the repurposing of existing medications may offer promising avenues for future therapeutic strategies.

2.2. Amyloid cascade hypothesis and anti-Aβ drugs

The Aβ cascade hypothesis, first proposed by Hardy and Higgins in 1992, is among the most widely accepted hypothesis³⁴. This hypothesis suggests that under the non-amyloid pathway, the APP is cleaved by α-secretase to produce soluble amyloid precursor protein alpha (sAPPα), which exhibits neuroprotective properties^35,36. However, in the pathological/amyloid pathway, APP undergoes an abnormal metabolic process (Fig. 2). In the amyloid pathway, APP is first cleaved by β-secretase and then by γ-secretase to produce Aβ peptides (Aβ42 or Aβ40), which aggregate into amyloid plaques, leading to neuroinflammation, neuronal loss, and vascular lesions³⁷. The potential therapeutic target could be the various enzymes involved in the APP hydrolysis pathway. Although the inhibitors of β-site APP cleaving enzyme 1 (BACE1, β-secretase)38, 39, 40, 41, and γ-secretase have been extensively studied42, 43, 44, these inhibitors have demonstrated considerable toxicity and inadequate effects during clinical trials. For instance, semagacestat (Phase III, NCT01035138), a non-selective γ-secretase inhibitor, was prematurely terminated due to a significant decline in cognitive function among the trial participants. On the other way, the enhancement of α-secretase activity facilitates the diversion of APP towards the non-amyloid hydrolysis pathway, thereby restoring the levels of sAPPα in the brain. Currently, the most advanced study on targeting Aβ is the direct removal of Aβ plaques using mAbs like aducanumab, lecanemab, and donanemab^13,45,46, which have been approved by the FDA. Based on the same mechanism of action, gantenerumab (NCT01760005), remternetug (NCT05463731), and solanezumab (NCT01760005) are also in phase III clinical trials.

2.3. Tau hypothesis and anti-tau drugs

Tau is a microtubule-associated protein primarily responsible for maintaining microtubule stability. Under pathological conditions, abnormal phosphorylation of tau protein results in the formation of hyperphosphorylated tau (p-tau), which dissociates from microtubules, thereby compromising their structural integrity and ultimately leading to neuronal cell death⁴⁷. Additionally, p-tau misfolds and aggregates within cells to form neurofibrillary tangles (NFTs), which represent a significant pathological hallmark of AD (Fig. 2).The imbalance in the activity of kinases related with tau protein is a critical factor in the aberrant phosphorylation of tau proteins, including glycogen synthase kinase 3β (GSK-3β), cyclin-dependent kinase 5 (CDK5), p38-mitogen-activated protein kinase (p38MAPK), c-Jun N-terminal kinases (JNK), microtubule affinity-regulating kinases (MARK), and cyclic-AMP-dependent kinase, which are also considered as potential therapeutic targets for the treatment of AD^48,49. Additionally, targeting the aggregation and dissemination of pathological tau represents another promising therapeutic strategy²⁸. Currently, the drugs in clinical research are mainly passive immunotherapy targeting tau protein, including a variety of mAbs, such as E2814 (Phase III, NCT05269394). In addition, some small molecule inhibitors can also inhibit the aggregation of tau protein, such as TRx0237 (Phase III, NCT03446001). Unfortunately, no pharmaceutical agent targeting tau proteins has received approval to date. Peters et al.⁵⁰ demonstrated that the deletion of tau protein also resulted in a reduction of Aβ deposition. Consequently, the development of combination therapies that target both Aβ and tau proteins may represent a significant advancement in the treatment of AD⁵¹.

2.4. Anti-inflammation drug for AD

Numerous studies have shown that neuroinflammation plays a key role in the pathogenesis of AD^29,52. Inflammatory activation of microglia and astrocytes in the brains of AD patients, increased levels of proinflammatory cytokines and neuroinflammation53, 54, 55. This persistent neuroinflammation exacerbates the pathological process of Aβ and tau^56,57. Currently, the principal strategies for neuroinflammation drug development encompass the antagonism of inflammatory factors and the improvement of microglial function. Triggering receptors expressed on myeloid cells-2 (TREM2) is an immune receptor expressed on the membrane of microglia, capable of binding to various lipid ligands to initiate cellular immune processes, such as APOE⁵⁸. Studies have shown that the functional loss resulting from TREM2 gene mutations adversely impacts the microglial response to Aβ plaques, significantly elevating the risk of, especially the R47H variant^59,60. AL002 (Phase II, NCT04592874) is a humanized monoclonal IgG1 antibody that binds to TREM2 and activates downstream signalling, thereby inducing microglial proliferation and reducing neuroinflammation and pericellular amyloid plaques⁶¹. In summary, AD is a multifactorial disease characterized by intricate pathological processes. Recent advancements in AD research have identified numerous novel mechanisms and therapeutic targets, including oxidative stress⁶², mitochondrial dysfunction⁶³, neurogenesis⁶⁴, proteostasis, vasculature, etc. (Fig. 2). Additionally, pharmacological agents targeting these pathways have also progressed to clinical trials¹².

3. Misfolding and aggregation mechanism of Aβ and tau proteins

3.1. Introduction to molecular dynamic simulation

Molecular dynamics (MD) simulation is a computational technique employed to investigate the dynamic behaviors of biomolecular systems at atomic level⁶⁵. The fundamental principle underlying MD simulation is Newton's equation of motion, which facilitates the calculation of atomic positions and velocities within the system at specific time intervals. Molecular force fields (FFs) are utilized to delineate the functional relationship between molecular structures and the potential energy of the system. Currently, widely adopted FFs for simulating protein and nucleic acid systems include the AMBER force fields⁶⁶, CHARMM FFs⁶⁷, GROMOS FFs⁶⁸, and OPLS-AA FFs⁶⁹. We have effectively utilized MD simulation techniques to investigate the mechanisms underlying conformational transitions in biomacromolecules70, 71, 72, protein folding73, 74, 75, 76, protein–ligand interactions^77,78, and ligand dissociation kinetics, etc.79, 80, 81. MD simulations are computationally intensive and necessitate substantial computational resources, particularly for long-time simulations. Moreover, the precision of FFs parameters critically influences the accuracy of MD simulations⁸². In recent years, advancements in computer hardware (particularly graphics processing units) and the molecular FFs developed based on machine learning (ML)^83,84, have significantly enhanced the capabilities and efficiency of MD simulations. These advancements have not only extended the time of sampling trajectories but also improved the accuracy of simulations, thereby establishing MD simulations as a crucial tool in the study of biological macromolecular systems. Notably, some enhanced sampling MD simulation techniques optimize the system by lowering energy barriers (such as accelerated MD, Gaussian acceleration MD) or introducing additional bias forces (such as steered MD, adaptive bias force MD), enabling the capture of rare events in the conformational transitions of macromolecules⁸⁰.

Unlike other common globular proteins, Tau and Aβ are intrinsically disordered proteins (IDPs) that are unable to fold into stable or ordered three-dimensional structures. IDPs are characterized by high levels of disorder, local mobility, and dynamic behaviour, which confer distinct functional advantages. Traditional experimental techniques such as X-ray diffraction and nuclear magnetic resonance (NMR) are inadequate for capturing the diverse conformations of IDPs. As an alternative, MD simulations have demonstrated unique advantages in the study of IDP conformations⁸⁵. However, the FFs mentioned above will make the conformational distribution of IDPs more extensive and prone to local energy minimization traps⁸⁶. Therefore, the force field for disordered proteins generally adjusts the dihedral parameters and the tendency of the secondary structure to meet the accuracy of the simulation results⁸⁷. Regarding conformational sampling, replica exchange molecular dynamics (REMD) simulations facilitate the random walk of conformations within the energy space through replica exchanges at adjacent temperatures. REMD approach enables the simulation to overcome energy barriers, thereby improving the sampling of the phase space⁸⁸, which has been widely used in the IDPs simulation^73,75,76. On the other hand, the construction of a Markov state model (MSM) utilizing multiple discrete trajectories facilitates the identification of key conformations of IDPs and the determination of conversion rates between these conformations, thereby offering a robust tool for investigating the conformational transition mechanisms of IDPs^89,90. More details for the MD simulation of amyloid proteins can be found in the Strodel's review⁹¹. In this section, we mainly review the research of MD simulation methods in the misfolding and aggregation mechanisms of tau and Aβ proteins.

3.2. Misfolding and aggregation mechanism of amyloid-beta (Aβ)

Aβ42, the full-length variant consisting of 42 amino acids, exhibits greater toxicity and a more rapid aggregation propensity compared to Aβ40⁹². Over the past two decades, NMR and cryo-electron microscopy (Cryo-EM) were used to determine the spatial structure of full-length Aβ and its multiple fragments. Utilizing MD simulation, an in-depth analysis of the structural dynamics and interactions of Aβ during its aggregation process in solution can be performed. In the past, extensive research has been conducted on the MD simulation of Aβ aggregates. The works on this aspect encompasses investigations into the aggregation mechanisms of Aβ monomers and oligomers in aqueous solutions⁹³, the aggregation mechanism at hydrophilic/hydrophobic interfaces94, 95, 96, as well as the disruption of amyloid fibrils through non-equilibrium MD simulations^97,98.

In the early stages, constrained by computational capabilities and the availability of experimental structures of Aβ, MD simulations were predominantly employed to investigate the aggregation processes of short Aβ segments, including the N-terminus region (Aβ_1–10), the central hydrophobic core (Aβ_17–21), the fibril-turn region (Aβ_22–29), and the C-terminus (Aβ_30–42), etc.⁹⁹. In addition, Aβ_16–22 (KLVFFAE) heptapeptide, one of the shortest fragments, exhibits a higher propensity for aggregation due to the electrostatic attraction between the charged residues K16 and E22 is also studied^100,101. With advancements in computing power and the acquisition of the complete experimental structure of Aβ, simulations focusing on the full length of Aβ40 and Aβ42 have increasingly become prevalent⁹⁹. These MD studies have elucidated the transient structures of various fragments and their roles in Aβ aggregation, including the formation of β-strands and β-hairpin structures at residues 39–42, which facilitate Aβ aggregation and the formation of Aβ oligomers¹⁰². Furthermore, both computational and experimental findings indicate that full-length Aβ40, Aβ42, and Aβ_16–22 accelerate aggregation at hydrophilic/hydrophobic interfaces, such as bilayer lipid membranes¹⁰³, air−water interface⁹⁴ and the surface of monosialotetrahexosylganglioside clusters on neuronal cell membranes¹⁰⁴. On the other hand, Aβ40 is inclined to adopt a β-hairpin conformation at the interface, further enhancing the aggregation process. In a recent study, Li et al.¹⁰⁵ investigated the co-aggregation mechanism of Aβ40 and Aβ42 peptides at the atomic level using REMD simulations. The simulation results indicate that Aβ40–Aβ42 is capable of assembling into a β-sheet-rich heterodimer, with its main chain interactions exhibiting greater strength compared to those of Aβ40–Aβ40 and Aβ42–Aβ42. Furthermore, this study elucidates the critical role of residues within the C-terminal hydrophobic region in facilitating the Aβ40–Aβ42 interaction, thereby offering a molecular basis for the cross-seeding phenomenon observed between Aβ40 and Aβ42.

The development of molecules capable of binding to Aβ peptides and inhibiting the misfolding and self-assembly of Aβ peptides into various toxic aggregates is an important approach in the design of novel anti-AD therapies¹⁰⁶. MD simulations have been employed to investigate the mechanisms by which various small molecules and peptides inhibit Aβ aggregation. The studied small molecules include the anti-AD drug candidate wgx-50^107,108, anti-arrhythmic medication propafenone¹⁰⁹, Norepinephrine¹¹⁰, D-tetrahydrocannabidiol¹¹¹, cyclic disulfide-rich peptides¹¹²,polyphenol¹¹³, epigallocatechin-3-gallate (EGCG)^114,115, and gallate¹¹⁶, etc. Understanding these molecular mechanisms offers valuable insights for the identification of potent Aβ aggregation inhibitors and the design of therapeutic agents targeting AD.

3.3. Misfolding and aggregation mechanism of tau protein

The tau protein exists in six distinct isoforms (0N3R, 0N4R, 1N3R, 1N4R, 2N3R, and 2N4R)¹¹⁷, which are determined by the number of inserted sequences at the N-terminus and the number of repeats within the microtubule-binding repeat domain (MBRD). The largest isoform, 2N4R, comprises 441 amino acids and includes several domains: a negatively charged and disordered N-terminal domain (N1, N2), a more structured and positively charged proline-rich region, a positively charged MBRD (R1-R4, also referred to as K18), and the C-terminus¹¹⁸ (Fig. 3). Native tau is a soluble IDP with little secondary structure, predominantly existing as the full-length 2N4R isoform. In AD patients, the content of insoluble tau is significantly increased (approximately 100-fold higher), primarily comprising the 0N and 4R isoforms, which exhibit a higher propensity for aggregation^119,120. Furthermore, the paired helical filaments PHF6∗ (²⁷⁵VQIINK²⁷⁹) and PHF6 (³⁰⁶VQIVYK³¹⁰), located in the R2 and R3 regions, represent two critical hexapeptide motifs that facilitate tau aggregation through the formation of β-sheet structures. These motifs also serve to differentiate the properties of the R2 and R3 regions from those of R1 and R4¹²¹. The Cryo-EM structure of tau fibrils, as elucidated by Scheres et al., reveals that paired helical filaments (PHFs) and straight filaments (SFs) are the two predominant forms of NFTs, which share a common “C-type” core structure consisting of the R3 and R4 regions (residues V306-F378)¹¹⁸ (Fig. 3). The precise pathological mechanisms of tau protein remain inadequately understood, which is a primary factor contributing to the unsuccessful outcomes of clinical trials involving tau-targeting therapeutics. Consequently, elucidating the processes of tau protein misfolding and aggregation is crucial for advancing our comprehension of tau-related pathology. MD simulations can be used to investigate tau pathology predominantly and study the influence of critical structural domains, post-translational modifications (PTMs), pathological mutations, and tau inhibitors on the misfolding and aggregation behavior of tau protein¹²².

Figure 3.

Schematic diagram of tau and the structures of PHF and SF. The figure was generated by Figdraw (www.figdraw.com).

Owing to the intrinsically disordered nature of tau protein, the majority of experimentally obtained crystal structures are aggregated tau fragments. Consequently, key tau fragments, such as PHF6 and PHF6∗, as initial structures for MD simulations were used to investigate the spontaneous aggregation mechanism of tau. Computational analyses have demonstrated that the R3 region (PHF6) exhibits the highest propensity for aggregation and is pivotal in the spontaneous aggregation process of tau^{73,90,123,124}. Liu et al.⁷³ employed MD simulations and MSM to elucidate multiple β-folding conformations of the R3 region, identifying key residues (such as S324-N327, K331-K334) and intermediate states that play a key role in the β-folding process. Similarly, Santos et al.¹²⁵ performed a 2-μs MD simulation involving 62 PHF6 peptides, highlighting the significance of hydrogen bonds and non-covalent interactions in the aggregation process. Their study further emphasized that, within both antiparallel and parallel aggregation mechanisms, specific hydrogen bonds formed by residues such as I3, Y5, V4, and Q2 are critical to the stability of tau aggregation. REMD simulations were employed to elucidate the molecular mechanisms underlying the co-aggregation of tau and Aβ peptides¹²⁶. The results indicate that the co-aggregation propensity of PHF6 peptides with Aβ_16–22 is significantly greater than that of PHF6∗, which is attributed to the hydrophobic and π–π stacking interactions between Y310 of PHF6 and Aβ_16–22, facilitating the formation of closed β-barrels and larger β-sheets in Aβ_16–22–PHF6 complex. Therefore, inhibiting the formation of tau and Aβ co-aggregates could also be a potential AD therapy. Nevertheless, the simulation of the full-length tau protein has been constrained by the absence of comprehensive experimental structural data. In recent years, advances in deep learning (DL) and AI have markedly enhanced the domain of protein structure prediction. Notably, the AlphaFold algorithm, developed by DeepMind, has demonstrated the capability to predict protein three-dimensional structures with remarkable accuracy by integrating deep neural networks with physical modeling techniques^127,128. He et al.¹²⁹ utilized the full-length structure of tau, as predicted by AlphaFold, and integrated coarse-grained with all-atom simulations to investigate the structural characteristics of the full-length 2N4R tau protein in solution. The radii and chemical shifts of the tau protein derived from MD simulations demonstrated a high degree of consistency with experimental data. Notably, this represents the inaugural MD simulation to validate the characteristics of the full-length tau protein through a comparative analysis of computational and experimental results.

Furthermore, the aggregation process of tau protein is influenced by PTMs, including phosphorylation, acetylation, methylation, N-glycosylation, etc.¹²⁰. Man et al.¹³⁰ demonstrated, through MD simulations and residue free energy decomposition, that abnormal phosphorylation at Ser289 and Ser293 can reduce the binding affinity of tau to microtubules. Concurrently, phosphorylation promotes the order-disorder conformational transition and intramolecular interactions of the monomer R2, potentially accelerating tau aggregation. In the dimer, the phosphorylation of Ser289 facilitates intermolecular interactions and the formation of β-folds, thus expediting the oligomerization of the R2 peptide¹³¹. On the other hand, the acetylation of specific residues can similarly enhance tau aggregation. The study by Shah et al.¹³² demonstrated that the acetylation of PHF6∗ (K280) and PHF6 (K311) results in the formation of antiparallel dimer cores during simulations, thereby promoting the development of highly ordered oligomers enriched with β-sheets. Zou et al.¹³³ demonstrated that acetylation at the K280 residue enhances the interaction among hydrophobic chemical groups within the PHF6∗ peptide, thereby facilitating its aggregation.

Since the preventing the aggregation of tau protein is a viable strategy for the treatment of AD, a diverse array of tau inhibitors has been identified, encompassing small molecules, peptides¹³⁴ and nanoparticles¹³⁵. MD simulations have been employed to investigate the molecular mechanisms through which these compounds, including naphthoquinone-dopamine (NQDA)^136,137, oleuropein aglycone (OleA)¹³⁸, and quercetin¹³⁹, inhibit the aggregation of tau protein monomers (Table 1)^136,138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148. MD simulation has also been used to explore the disaggregation mechanism of inhibitors on already formed fibrils, such as derivatives of 2,4-thiazolidinedione C30¹⁴⁰, proanthocyanidin B2 (PB2)¹⁴¹, purpurin¹⁴² (Table 1). At the molecular level, tau inhibitors primarily engage in hydrogen bonding (Q276, V306, Q307, I308, V309, Y310, K311, V313, etc.), hydrophobic interactions and π–π stacking interactions (I227, K267, V306, I308, V309, Y310, P312, K340, etc.) with key residues that facilitate tau aggregation. In summary, the inhibitory mechanisms of these molecules encompass: establishing interactions with critical residues to hinder peptide–peptide interactions; altering the β-sheet structure of tau into a random coil conformation; preventing the transformation of tau conformation into aggregates rich in extended β sheets; reducing the content and order of extended β-sheet structure (Table 1)^136,138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148.

Table 1.

Summary of MD simulation works to study the molecular mechanism of tau inhibitors in inhibiting tau aggregation.

Molecule	Inhibition molecular mechanism	Interaction	Ref.
Naphthoquinone-dopamine (NQDA)	NQDA prevents the formation of hydrogen bonds between peptides, reducing the ability of PHF6/PHF6∗ to self-assemble into long β chains and large β sheets	NQDA–Q276 (H-bond) NQDA–Y310 (π–π stacking) NQDA–I227 (Hydrophobic interaction)	136
Oleuropein aglycone (OleA)	OleA binds to PHF6 through hydrogen bonds and hydrophobic interactions, forcing PHF6 to exist in a monomeric random coil conformation	OleA–Q307, Y310, K311 (H-bond) OleA–Y310 (π–π stacking) OleA–I308, V309 (Hydrophobic interaction)	138
Quercetin	Quercetin destroys the α-helical structure of tau R2 (PHF6∗) monomer, resulting the increase of native random coil structure content	Quercetin–G273 (H-bond) π–π stacking	139
2,4-Thiazolidinedione derivative (C30)	C30 destroys the ordered structure of PHF6 oligomer, reduces the content of β-sheet, and transforms β-sheet into random coil structure	C30–Q307 (H-bond) C30–Y310 (π–π stacking) C30–V309 (Hydrophobic interaction)	140
Proanthocyanidin B2 (PB2)	PB2 transforms PHF6 oligomer from the ordered β-sheet into disordered one	PB2–Q307, Y310, K311 (H-bond) PB2–V309 (Hydrophobic interaction)	141
Purpurin	Purpurin binds to PHF6 fibrils, causing fibril chain breakage and promoting fibril disassembly	Purpurin–V306, Q307, I308 (H-bond) Purpurin–V306, V309 (Hydrophobic interaction)	142
Adenosine triphosphate (ATP)	ATP binding reduces the oligomerization state of PHF6 and maintains it in a random coil conformation	ATP–V306, Q307, V309, K311 (H-bond) ATP–Y310 (π–π stacking)	143
Trazodone	Trazodone binds to the 1N4R isoform and reduces fibril formation.	Y29, Y197, residues 275–280, residues 306–310, Y394	144
Epigallocatechin-3-gallate (EGCG)	EGCG interacts with 1N4R tau isoform and inhibits the PHFs assembly	EGCG–N265, L266, V313 (H-bond) EGCG– K267, Y310, P312, K340 (Hydrophobic interaction)	145
Baicalein	Inhibits aggregation of repetitive tau	Baicalein–N265, H268, Q338 (H-bond) Baicalein–Y310, V313 (Hydrophobic interaction) Baicalein–H299 (water bridge)	146
Naphthoquinone-Tryptophan hybrids (NQTrp)	NQTrp disrupts the interpeptide hydrogen bonds, resulting in loose packing of the β sheets and chain separation	NQTrp –V306 (H-bond) NQTrp–Y310 (π–π stacking)	147
Palmatine chloride (PC)	PC transforms the β-sheet structure of PHF6 into a random coil and breaks the inter-peptide hydrogen bonds to decompose PHF6 fibrils	PC–V306 (H-bond) PC–Y310 (π–π stacking)	148

3.4. Future direction of MD simulations for Tau and Aβ proteins

While traditional MD simulations have been extensively employed to investigate the intrinsic conformational transitions and potential biological functions of IDPs such as Tau and Aβ, there are still notable limitations and challenges in this field. Simulating the behaviour of Tau and Aβ proteins under real in vivo or experimental conditions remains difficult due to the complex cellular environment. Furthermore, the accuracy of simulation outcomes is often depending on the precision of FF parameters and the duration of the simulations. To address these issues, force fields specifically designed for IDPs, such as ff14IDPs¹⁴⁹, Drude2019IDP¹⁵⁰, CHARMM36m¹⁵¹, environmental specific precise force field (ESFF1)¹⁵², ff14IDPSFF¹⁵³, and a99SB-disp¹⁵⁴, etc., are continually being developed to more accurately capture the true conformational flexibility of IDPs and mitigate the propensity to form α-helices and β-sheets. More information on force fields for amyloid peptide assembly are summarized by Man et al.¹⁰¹. Additionally, the integration of AI models into MD simulations offers novel approaches and directions for enhanced sampling of IDPs¹⁵⁵. AI models, such as variational autoencoders (VAEs), recurrent neural networks (RNNs), generative adversarial networks (GANs), reinforcement learning (RL), diffusion models, and flow-based models, can be utilized to guide the sampling process in MD simulations, steering the system toward more favourable conformational space exploration and improving the structural sampling efficiency of IDPs^156,157. For instance, Zhu et al.¹⁵⁸ introduced Phanto-IDP, an advanced DL model specifically designed to generate accurate protein backbones and enhance conformational sampling for IDPs. Phanto-IDP exhibited the lowest reconstruction root-mean-square deviation and Jensen–Shannon divergence when compared to existing DL models. More AI models for protein conformations generation are listed in Table 2158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173. Recently, the AI²BMD method has efficiently performed all-atom simulations of various proteins with quantum-level accuracy, and its calculation speed is much faster than density functional theory and other quantum mechanics methods¹⁵⁹. Despite the improvements in sampling efficiency and speed offered by AI-based models, their accuracy remains constrained by the scarcity of high-quality data. Furthermore, current AI models predominantly focus on smaller IDPs, necessitating the development of models capable of handling larger IDPs in the future. Additionally, incorporating more experimental data into model construction could enhance model performance.

Table 2.

Representative studies of protein conformations generated based on AI models.

Model	Architecture	Description	Ref.
Phanto-IDP	VAE	A deep learning model combing a graph-based encoder and a transformer-based decoder for generating accurate and diverse protein conformations, particularly for IDPs.	158
AI2BMD	GNN	An AI-driven ab initio biomolecular dynamics system that combines protein fragmentation and ML force fields to efficiently simulate full-atom proteins with quantum-chemical accuracy.	159
idpGAN	GAN	A GAN-based ML model for generating protein conformational ensembles without physics-based iterative sampling, particularly for IDPs.	160
ICoN	VAE	An unsupervised deep learning model for learning conformational changes from MD simulation data, using which the conformational space of Aβ42 is comprehensive sampled.	161
Transformer-backmapping	Transformer	A transformer-based backmapping model for generating protein conformational ensembles from fixed protein backbones.	162
idpSAM	Diffusion	A deep generate modelling approach combined an autoencoder and a denoising diffusion probabilistic model (DDPM) for generating 3D structural ensembles of IDPs with high transferability.	163
IDPFold	Conditional diffusion	A fine-tuned diffusion model for generating IDPs conformational ensembles from protein sequences.	164
taneja_ml_idr	Denoising diffusion	A two-stage supervised ML model for generating 3D conformational ensembles of IDPs using diffusion models based on the predicted 2D properties of IDPs.	165
BGs	Flow based	A flow-based Boltzmann generator for reweighted biomolecule samples from the system's energy function.	166
Str2Str	Score based	An amortized denoising scoring matching model trained using crystal structures for zero-shot protein conformation sampling.	167
AlphaFlow/ESMFlow	Flow matching	Highly accurate single-state predictors (such as AlphaFold and ESMFold) are fine-tuned to generate protein structures in a custom flow framework.	168
ConfDiff	Force-guided diffusion	A force-guided SE (3) diffusion model for generating protein conformations with high diversity and fidelity.	169
DeepDriveMD	Convolutional VAE	A deep learning model combining convolutional VAE and MD for folding small proteins.	170
ITO	Diffusion	A denoising diffusion probabilistic model for learning surrogates of the simulation process with multiple time-resolutions.	171
Timewarp	Flow matching	A normalizing flow generative model for accelerating MD simulation through learning to dynamically increase and optimise time steps.	172
DiAMoNDBack	DDPM	A diffusion-based autoregressive model to backmap coarse-grained protein structures to all-atom resolution, achieving state-of-the-art performance in bond fidelity, clash reduction, and configurational ensemble generation.	173

In summary, MD simulations can model the dynamic processes of amyloid, such as misfolding, aggregation, and conformational transitions. These simulations elucidate the intricate structural dynamics of amyloid and furnish comprehensive dynamic and structural insights, which are essential to understand the misfolding mechanisms of Aβ and tau proteins. Despite the numerous advantages offered by MD simulations, several challenges persist in investigating the misfolding mechanisms of Aβ and tau proteins. Notably, simulating time-scales and larger systems necessitates substantial computational resources and time, often resulting in inadequate sampling. Furthermore, the protein concentrations employed in MD simulations are typically several orders of magnitude higher than those observed in vitro or in vivo, which directly influences the nature of the aggregates formed¹⁷⁴. Furthermore, current force fields exhibit limitations in accurately simulating disordered protein states, often due to insufficient or inaccurate parameterization. To address these challenges, potential strategies include the refinement and optimization of force fields to enhance the reliability of simulations.

4. Target identification and lead compounds discovery of AD

4.1. Drug target identification

The identification of targets represents a crucial initial phase in the early stages of drug discovery, a process increasingly enhanced by the proliferation of AI-based models. These AI technologies leverage extensive bioinformatics datasets, encompassing genomics, proteomics, and drug-target interaction data, to identify potential drug targets¹⁷⁵. Notably, DL models such as neural networks demonstrate exceptional performance in drug target identification, and can discover patterns hidden in complex data and predict the interaction between drugs and protein. For example, Wang et al.¹⁷⁶ constructed an integrative network model that identified ATP6V1A as a pivotal regulator within a neuron-specific subnetwork significantly impacted by LOAD. Based on this regulator, the study identified NCH-51 as a promising candidate compound for LOAD treatment, demonstrating its capability to enhance ATP6V1A expression and partially ameliorate the neurological deficits associated with ATP6V1A deficiency. Furthermore, to tackle the challenges posed by the high-dimensional and sparse nature of human protein–protein interaction network features, a DL-based framework was developed to identify novel potential AD genes, such as DLG4, EGFR, RAC1, SYK, PTK2B, and SOCS1, as well as to discover potential therapeutic agents for AD, including tamoxifen, bosutinib, and dasatinib¹⁷⁷. Despite the identification of numerous susceptibility loci for AD through genome-wide association studies (GWAS), the task of pinpointing high-confidence AD risk genes and potential therapeutic interventions from GWAS and multi-omics data continues to be a formidable challenge¹⁷⁸. Fang et al.¹⁷⁹ developed a network-based AI framework that integrates GWAS with multi-omics data to identify high-confidence AD risk genes and druggable targets. This innovative approach identified 103 genes associated with AD and three drugs (pioglitazone, febuxostat, and atenolol) that are significantly linked to a reduction in AD risk. In addition, in vitro experiments determined that pioglitazone improves AD by downregulating GSK3β and CDK5 in human microglia cells. In a recent research, Lim et al.¹⁸⁰ combined multiomic PandaOmics platform with the FuzDrop to identify potential new AD targets (including MARCKS, CAMKK2, and p62). By leveraging ML and network analysis, researchers can systematically explore vast datasets to more efficiently identify key molecular targets and therapeutic candidates.

4.2. 3D structure prediction

In order to elucidate the biological functions of target proteins and facilitate the rational design of potential therapeutics, it is essential to determine the three-dimensional structure of proteins in specific conformations. Conventional experimental methods for protein structure determination, including X-ray crystallography, NMR spectroscopy, and cryo-EM, offer high accuracy but are often time-intensive and expensive¹⁸¹. Furthermore, these experimental techniques encounter significant challenges when attempting to predict the structures of highly disordered and dynamic proteins, such as tau and Aβ. Consequently, a substantial proportion of protein structures remain unresolved to date. With advancements in computational technology, protein structure prediction has emerged as a method to model the structures of proteins with unknown experimental data based on their amino acid sequences. The earliest protein structure prediction methods mainly relied on sequence alignment and homology modeling, which mainly compared the target protein sequence with the protein sequence of known structure, found the most similar structure from homologous proteins. Homology modeling play an important role in the target-based drug discovery¹⁸². However, when the sequence similarity between the target protein and the template protein is low, the structure prediction error will be larger. In addition, the limited number of protein structures solved experimentally has further restricted the application of homology modeling. Although there are some physics-based de novo prediction methods, they require a lot of time and resources and are only applicable to very small protein structures.

Since the introduction of AlphaFold by DeepMind¹²⁷, AI-based approaches, particularly those utilizing DL, have achieved significant breakthroughs in the field of protein structure prediction. The subsequent release of AlphaFold2¹⁸³ significantly enhanced this capability by accurately predicting the structures of the majority of known proteins. More recently, the latest iteration, AlphaFold3, addresses the challenge of predicting complex structures, including those involving protein–ligand, protein–nucleic acid, and antibody–antigen complexes¹⁸⁴. In comparison to traditional experimental methods, AlphaFold not only substantially reduces the time required for structure prediction but also enables the prediction of three-dimensional protein structures in the absence of homologous templates, particularly excelling in the prediction of small protein structures¹²⁷. The emergence of AlphaFold also provides a powerful tool for predicting the three-dimensional structure of tau and Aβ proteins. A study employed AlphaFold2 to predict the conformation of full-length tau (FLT) and tauC3, emphasizing the critical role of residue 432 in preserving the structural integrity of FLT. In contrast to FLT, tauC3 exhibited an open conformation characterized by the MBRD being positioned distally from the N-terminus¹⁸⁵. Additionally, Santuz et al.¹⁸⁶ utilized AlphaFold2 to predict the structures of Aβ42 oligomers up to hexameric forms, uncovering significant conformational changes that occur during the early aggregation processes associated with AD. The study also demonstrated that the structural evolution of Aβ42 originates from a variety of monomeric configurations, encompassing extended conformations characterized by two helices spanning various residues, as well as random coil structures with turns involving residues G9-H13 and V18-N27. These monomeric forms progress into intricate configurations in dimeric and trimeric assemblies, including helical bundles and β-sheet structures. Higher-order oligomers, ranging from tetramers to hexamers, exhibit a combination of helical and U-shaped conformations, underscoring the dynamic nature of Aβ42 aggregation.

While AlphaFold exhibits robust performance in predicting static protein structures, it possesses inherent limitations in its predictive capabilities, particularly concerning the structural dynamics of proteins, such as the aggregation states of tau and Aβ, and their conformational changes across various environments. Furthermore, AlphaFold is currently unable to accurately capture post-translational modifications (e.g., acetylation, phosphorylation) or covalent modifications, all of which are critical for a comprehensive structural and functional understanding of tau and Aβ in AD. Additionally, AlphaFold encounters challenges in accurately delineating protein–protein and protein–ligand interaction interfaces, a crucial aspect of drug discovery. Although AlphaFold3 has made progress in modeling protein complexes, the accuracy of AlphaFold's predictive outcomes still requires validation through experimental structure determination, especially in contexts involving protein interactions and dynamic transformations. In the realm of AD drug discovery and development, comprehending the conformational flexibility and binding behaviours of targets such as tau and Aβ is of paramount importance. This necessity highlights the limitations of current methodologies and underscores the importance of integrative approaches that combine AlphaFold predictions with experimental techniques like Cryo-EM and NMR spectroscopy.

4.3. Binding site prediction

Protein–ligand binding site prediction is integral to computational drug design, encompassing applications such as structure-based virtual screening, de novo drug design, and the identification of protein functional sites¹⁸⁷. Various strategies have been developed for predicting drug binding sites, which can be classified into structure-based, template-based, traditional ML-based, and DL-based methods^188,189. Among these approaches, commercial software tools like SiteMap¹⁹⁰, SiteFinder¹⁹¹, Molsoft ICM-Pro¹⁹² are extensively employed for pocket prediction and drug design. In a study by Akhoon et al.¹⁹³, a novel donepezil-like natural compound derivative (D1) was identified as a potential AChE inhibitor for targeting AD. Then, the SiteMap was used to predict the dual binding interaction mode of D1 with the catalytic anionic site and peripheral anionic site of human AChE. The characteristics of these dual binding sites of D1 were further elucidated through biological studies, demonstrating its ability to both increase acetylcholine levels and reduce the accumulation of amyloid-beta plaques in Caenorhabditis elegans. Similarly, in the development of a dual-target (AChE and microRNA-15b) inhibitor for AD based on microRNA, SiteMap was utilized to predict binding pockets on pre-miRNA-15b, further showcasing its utility in diverse drug design applications¹⁹⁴.

With the rapid development of AI, some open-source tools have been developed and applied for binding site prediction of ligands, such as P2Rank¹⁹⁵, DeepSite¹⁹⁶, DeepPocket¹⁹⁷, and ScanNet¹⁹⁸. These tools have substantially enhanced the capacity to model and predict protein-ligand interactions, enabling the identification of potentially druggable binding sites on key proteins implicated in AD pathogenesis. Among these, P2Rank is a template-free tool that utilizes ML algorithms to predict ligand binding sites by assessing the ligandability of local chemical environments on the solvent-accessible surface of a protein. DL-based tools such as DeepSite and DeepPocket employ 3D convolutional neural networks (CNNs) to capture the characteristics of binding sites, thereby improving the accuracy of predictions regarding the location and shape of these sites. Meanwhile, ScanNet provides an end-to-end, interpretable DL model that directly learns 3D structural features for binding site prediction. Recently, Tian et al.¹⁹⁹ introduced a user-friendly web platform, PASSer, designed for the rapid and precise prediction of allosteric sites utilizing three distinct ML models. Furthermore, the integration of DL with protein structure prediction methods, such as AlphaFold, has provided researchers with unprecedented capabilities to predict not only the structure of previously unsolved protein targets but also their druggable binding pockets. This integration enables the systematic characterization of key AD-related targets, offering insights into binding pockets on uncommom proteins implicated in the complex disease.

In summary, the significance of binding site prediction should not be underestimated, especially in the context of diseases such as AD, where therapeutic development is particularly challenging. Precise identification and characterization of binding sites are crucial for the efficient and cost–effective discovery of novel drugs. Targeting binding pockets with therapeutics that demonstrate high specificity and minimal off-target interactions can substantially increase the likelihood of clinical success. Additionally, accurate binding site prediction facilitates the design of multi-target or multi-modal drugs, which is particularly advantageous for complex diseases like AD that involve multiple pathological pathways.

4.4. Virtual screening

In the drug discovery process, the identification of lead compounds with strong binding affinity to the target is a critical step. Traditional high-throughput screening methods necessitate the evaluation of a vast array of compounds for biological activity, a process that is often costly, time-consuming, and labor-intensive. In contrast, virtual screening (VS) employs computational tools and drug design software to efficiently identify potential hit compounds from extensive compound libraries, thereby reducing the need for experimental activity assays to a select few hits²⁰⁰. This approach not only reduces experimental costs but also accelerates the lead compound discovery process. In the conventional CADD approach, VS primarily encompasses ligand-based virtual screening (LBVS) and structure-based virtual screening (SBVS)²⁰¹ (Fig. 4). The LBVS approach predominantly relies on the structural attributes of known active molecules and the correlation between molecular structure and biological activity to develop models for the identification of novel compounds. This includes the utilization of pharmacophore models^202,203 and quantitative structure–activity relationship (QSAR) models²⁰⁴. Conversely, the SBVS approach primarily involves techniques such as molecular docking, MD simulations, and de novo drug design, all of which necessitate the availability of the three-dimensional structure of the known target, particularly the drug binding site. The binding site provides critical information, including its size, shape, hydrophobicity, and electrostatic properties, enabling these techniques to predict the interactions of candidate compounds with the target protein. Among them, molecular docking is a critical technique within SBVS. It primarily involves the generation of potential binding conformations of ligands within the binding sites of macromolecular targets using search algorithms. Subsequently, it ranks various compounds and conformations utilizing empirical, force field, and knowledge-based scoring functions²⁰⁵. Despite its utility, molecular docking-based VS process still needs to improve when dealing with the screening of ultra-large molecular databases, such as the substantial computational resources required and the inadequate consideration of compound diversity. Simultaneously, traditional scoring functions exhibit limitations in predicting binding affinity, particularly in accurately characterizing conformational entropy and solvation effects²⁰⁶, as well as in tailoring parameters for specific target proteins.

Figure 4.

Virtual screening approaches for AD lead compound.

With the advent of AI, virtual screening methods using ML and DL models have been effectively developed to explore a wider range of chemical space, such as ligand-based ML models (LBML)²⁰⁷, structure-based ML models (SBML)²⁰⁸ and structure-based deep learning (SBDL) models^209,210. AI-based scoring functions have been developed to evaluate the potential binding affinity between small molecules and biological targets, such as RF-score²¹¹, Pafnucy²¹², DeepDTA²¹³, OnionNet²¹⁴, and DeepPurpose²¹⁵, etc. In comparison to traditional scoring functions, those based on AI and ML exhibit superior performance in predicting binding affinity^216,217. Additionally, AI-driven VS demonstrates significantly greater speed than conventional molecular docking-based VS methods, enabling the efficient screening of extensive databases and facilitating the exploration of a broader array of molecular structures. Nonetheless, these models typically necessitate substantial training datasets and are frequently perceived as black-box models, which can complicate the interpretation of results and diminish confidence in their predictive capabilities. Consequently, there remains considerable potential for the advancement of AI-based VS methodologies, particularly in enhancing the model's generalization capability and accuracy, as well as augmenting its interpretability. This section mainly summarizes some VS researches in the discovery of AD lead compounds.

Over the past decade, numerous VS studies have focused on the primary mechanisms underlying AD, with significant attention given to targets such as AChE, butyrylcholinesterase (BChE), BACE1, and GSK-3β, etc.²¹⁸. AChE, in particular, is a critical target in the acetylcholine degradation hypothesis related to AD and has been extensively investigated for its role in alleviating AD-related symptoms. Recent advancements have led to the identification of novel AChE inhibitors through methodologies such as molecular docking, pharmacophore modeling, MD simulation, and deep neural networks, as detailed in Table 3^219,221,222,224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236. A study conducted by van der Westhuizen et al.²¹⁹ identified several AChE inhibitors with novel scaffolds through docking-based virtual screening and subsequent biological evaluation. MD simulation were performed to elucidate the binding mechanisms, thereby explaining the influence of water bridges on compound activity. Additionally, BChE, another hydrolytic enzyme associated with AD, has been proposed as a crucial strategy for AD treatment, particularly to compensate for reduced AChE levels²²⁰. A related investigation identified the compound Z32439948, a novel BChE inhibitor, through the application of a pharmacophore-based VS approach. The binding sub-pocket of this compound on BChE was characterized using molecular docking and MD simulations. Leveraging the structural insights gained from this sub-pocket, a series of compounds were synthesized and assessed for their biological activity. Among these, compounds 27a and (R)-37a demonstrated significant enhancement of cognitive function in models of scopolamine and Aβ_1–42 peptide-induced cognitive deficits²²¹. In a recent study, Wu et al.²²² developed four DL models to identify potential anti-AD effect of Kaixinsan derived from the traditional Chinese medicine formulation. The research involved training four distinct deep neural networks: one dedicated to predicting AD therapeutics and the remaining three focused on predicting inhibitors for specific AD targets, namely AChE, 5-HT6 receptor, and MAO-A. The compounds identified as promising by these models underwent both in vitro and in vivo validation, demonstrating efficacy in inhibiting key enzymes and enhancing microglial β-amyloid phagocytosis. BACE1 inhibitors are regarded as promising agents for the treatment of AD due to their ability to inhibit Aβ plaque formation²²³. A related study by Das et al.²²⁴ employed a QSAR model to screen potential BACE1 inhibitors from the natural compound libraries of the Timtec and ZINC databases. This study identified two effective scaffolds, isorhamnetin and taxifolin, as BACE1 inhibitors, with EC₅₀ values of 27 and 36 μmol/L, respectively. Additionally, GSK-3β is extensively studied as a critical target for reducing tau aggregation in the treatment of AD. In a study conducted by Lin et al.²²⁵, potential inhibitors of GSK-3β kinase for AD were identified through docking-based VS of over 1.1 million compounds sourced from the ZINC database and in-house databases. Following this screening, 16 compounds underwent evaluation via kinase and cell assays. Among these, compound VB-008 emerged as the most promising GSK-3β inhibitor, exhibiting an IC₅₀ value of 5.4 μmol/L, and demonstrated the capability to enhance tau solubility. Candidate compounds targeting other proteins associated with AD, such as CDK5, APOE, and Nod-like receptor family pyrin domain-containing 3 (NLRP3), have also been identified through computational screening methods. Representative examples are provided in Table 3, where molecular docking emerges as the predominant approach in VS. Overall, both traditional and AI-driven VS methodologies have become indispensable tools in the early phases of AD drug discovery, facilitating the efficient identification and optimization of potential therapeutic agents.

Table 3.

Summary of virtual screening in AD drug discovery.

Pathway	Target	Method	Inhibitor	Activity	Ref.
Neurotransmission	AChE	Molecular docking, MD simulation	1-Ethyl-3-methoxy-3-methylpyrrolidine and 1H-pyrrolo[3,2-c] pyridin-6-amine scaffold	IC50: <50 μmol/L	219
	AChE	Molecular docking	S-I 26	AChE IC50: 14 ± 1 μmol/L	226
	BChE	Pharmacophore model, molecular docking, MD simulation	Z32439948	hBChE IC50: 1.4 ± 0.6 μmol/L	221
	AChE, MAO-A, 5-HT6	Deep neural network	2,4-di-tert-Butylphenol; elemicin	AChE IC50: 4.512 ± 0.191 μmol/L; MAO-A IC50: 62.24 ± 1.08 μmol/L;	222
Aβ	BACE1	Pharmacophore model	ZINC53276039	Antiamyloidogenic activity	227
	Aβ42/BACE1	MD simulation, molecular docking	AE-848 (a dual inhibitor)	Aβ42 and BACE1 IC50: 36.95 and 22.70 μmol/L	228
	FoxO1	Molecular docking, MD simulation	Compound D	FoxO1 agonist	229
	BACE1	QSAR model	Isorhamnetin, taxifolin	EC50: 27, 36 μmol/L	224
Tau	GSK3β	molecular docking	VB-008	IC50: 5.4 μmol/L	225
	MARK4	Virtual high-throughput screening	NAG	IC50: 4.11 μmol/L	230
	CDK5/p25	Pharmacophore model, molecular docking	Thieno[3,2-c]quinolin-4(5H)-one	IC50: 7 μmol/L; Ki: 3.38 μmol/L	231
	p38MAPK-CREB	Molecular docking, MD simulation	Puerarin	Anti-AD by maintaining the synaptic plasticity	232
Inflammation	APOE4	High-throughput virtual screening protocol, MD simulation	CBA2	Protects against Aβ42 aggregation	233
	NLRP3	Molecular similarity, docking, MD simulation	DS15	IC50: 3.85 μmol/L	234
	NLRP3	Pharmacophore model, molecular docking	NIC7	IL-1β IC50: 9.0 μmol/L; IL-18 IC50: 3.5 μmol/L; LDH IC50: 9.1 μmol/L	235
	Fatty acid amide hydrolase (FAAH)	Pharmacophore model, molecular docking	NCI80, NCI794, NCI328	IC50: 1.76, 2.14, and 2.10 nmol/L	236

4.5. De novo molecular generation

Although the number of molecules currently available for virtual screening is as high as tens of millions or even hundreds of millions, it is only the tip of the iceberg compared to the real chemical space. Therefore, generating new chemical molecules with novel structures is crucial for drug discovery. Traditional de novo drug design can generate molecules that are complementary to the receptor in shape and energy based on its activity. However, this method relies on the binding pocket of protein and often ignores the flexibility of the pocket²³⁷. The advent of AI-based molecular generative models has precipitated a paradigm shift in drug discovery, enabling the rapid and efficient exploration of extensive chemical spaces and facilitating the design of target-specific active molecules. Classical DL architectures including VAEs, RNNs, GANs, RL, and flow-based models, which are foundational technologies in generative modelling to generate novel structures^238,239.

Over the past few years, a diverse array of generative models both ligand-based and structure-based has been developed to advance drug discovery and development²⁴⁰. Ligands-based generative models typically use representations such as 1D SMILES strings, 2D molecular graphs, or even 3D conformations to generate potentially active molecules with similar or improved properties. Here, we present a comprehensive overview of AI-based molecular generation models pertinent to AD, as summarized in Table 4241, 242, 243, 244, 245^,249, 250, 251, 252. For example, Li et al.²⁴¹ proposed a conditional graph generative model to generate drug-like molecules with specific properties, such as drug-likeness and synthetic accessibility, as well as dual inhibitors of JNK3 and GSK3β. Furthermore, a multi-objective approach, termed RationaleRL was developed to generate molecules exhibiting multiple properties. JNK3 and GSK-3β were evaluated to generate dual inhibitors of AD with drug similarity and synthetic accessibility by using this model²⁴². Recent advances in natural language processing models have significantly expedited the progress of de novo drug design utilizing chemical language models. In a study by Ballarotto et al.²⁴³, a chemical language model was employed to design novel Orphan nuclear receptor NURR1 agonists, a neuroprotective transcription factor that has emerged as a promising target for the treatment of AD, Parkinson's disease, and multiple sclerosis. Recently, AI model was developed for the de novo generation of BACE1 inhibitors for AD. This application facilitated the generation of 1,000,000 candidate inhibitors, from which 14 candidates were identified through docking and MD simulation²⁴⁴. CNSMolGen, one RNN-based generative model, has been specifically developed for central nervous system-related drug discovery. This model has demonstrated the capability to generate synthesizable and biologically relevant molecules with favorable pharmacological profiles, as validated by computational studies²⁴⁵. However, it is evident that these generative models primarily rely on ligand information without incorporating the binding pocket details of the target protein or chemical knowledge during the molecular generation process.

Table 4.

Summary of AI-based molecular generation models in AD drug discovery.

Model	Description	Architecture	Target	Source code	Ref.
Molecule_generator	A novel multi-objective de novo drug design framework using a conditional graph generative model	Conditional graph generative model	JNK3, GSK3β	https://github.com/kevinid/molecule_generator	241
RationaleRL	A novel approach for multi-objective molecule generation by utilizing interpretable substructures	RL	JNK3, GSK3β	https://github.com/wengong-jin/multiobj-rationale	242
CLM	A novel approach for de novo design of Nurr1 agonists using a fragment-augmented conditional language model	RNN with LSTM	Nurr1	https://github.com/ETHmodlab	243
WGAN-GP	The first AI application to the de novo generation of BACE1 inhibitors for AD	Wasserstein GAN with Gradient Penalty, GA	BACE1	https://github.com/EvX57/GenAI-DrugDiscovery-for-BACE1	244
CNSMolGen	A bidirectional RNN model for de novo design of central nervous system drugs	Bidirectional RNN	SERT	https://github.com/xueww/CNSMolGen	245
Moo-denovo	A pair-based multi objective optimization approach to generate molecules with specified properties	RNN	NA, AChE, SARS-CoV-2	https://github.com/alberdom88/moo-denovo	249
MARS	Multi-objective drug discovery that utilizes Markov chain Monte Carlo sampling and graph neural networks to efficiently generate diverse molecular candidates with desired chemical properties	GNN and Markov chain Monte Carlo sampling	JNK3, GSK3β	https://github.com/yutxie/mars	250
MolSearch	Search-based multi-objective molecular generation and property Optimization	Monte Carlo tree search (MCTS)	JNK3, GSK3β	https://github.com/illidanlab/MolSearch.	251
MolRL-MGPT	A RL algorithm with multiple GPT agents for drug molecular generation	GPT, reinforcement learning	SARS-CoV-2, GSK3β, JNK3	https://github.com/HXYfighter/MolRL-MGPT	252

Since drug molecules exert their therapeutic effects by binding to target proteins, incorporating structural information of proteins is crucial for enhancing the accuracy of drug design. Structure-based molecular generation approaches consider the structural details of the target protein, particularly the binding pocket, to guide the molecular design process. Ragoza et al.²⁴⁶ investigated the generation of 3D molecular structures using receptor binding site-conditional deep generative models. This model is trained using a conditional VAE applied to an atomic density grid representation of cross-docked protein–ligand structures. An inference procedure is subsequently employed to derive valid molecular conformations from the generated atomic density. Jiang et al.²⁴⁷ developed an autoregressive flow-based generative model, termed PocketFlow, which is designed to generate novel molecules within protein binding pockets while incorporating chemical knowledge during the molecular generation process. This model successfully produced molecules that were 100% chemically valid and exhibited high drug-likeness, leading to the identification of active compounds targeting histone acetyltransferase 1 and YTH domain-containing protein 1. Recently, the Interaction Prior-guided Diffusion model (IPDIFF) was introduced, leveraging prior knowledge of protein–ligand interactions for protein-specific 3D molecular generation²⁴⁸. Although these structure-based generative models can generate efficient molecules based on binding pockets, they are limited by the information of the structures of target protein. Furthermore, structure-based generative models may generalize poorly to unseen protein binding pockets, which may cause the model to perform less well than expected in real applications. In general, the molecular generation model is still in its infancy in the field of AD drug discovery and encounters some challenges, including the rationality and synthesizability of the generated molecules, which require further refinement. Nonetheless, it is indisputable that both presently and future, AI-based molecular generation models will introduce significant innovation and opportunities to the field of AD drug discovery. These models are anticipated to advance the development and dissemination of novel therapeutics, thereby making positive contributions to the treatment of AD and enhancing the quality of life for patients.

5. Multi-target directed ligands of AD

Given the complex nature of AD pathogenesis, traditional approaches focusing on a single target have achieved limited success in developing effective treatments. As a result, multi-target drug discovery has emerged as a promising strategy for addressing multifactorial complex diseases, attracting considerable interest from the research community²⁵³. Multitarget-directed ligands (MTDLs) have been proposed as a novel approach for drug discovery in complex diseases such as AD, aiming to integrate multiple therapeutic actions within a single molecule^254,255. To minimize time-intensive processes at the experimental stage, a range of computational techniques have been developed for multi-target drug discovery, including target prediction, similarity search, virtual screening, pharmacophore modelling, inverse docking, QSAR analysis, and network pharmacology²⁵⁶. In this context, we present several examples of MTDLs for AD, along with their associated target combinations, as depicted in Fig. 5.

Figure 5.

Multi-target directed ligand design strategies for 9 targets associated with Alzheimer's disease.

Over the years, the discovery and design of multi-targeted drugs targeting AChE and other AD-related targets have remained pivotal in the therapeutic strategies for AD²⁵⁷. The synergistic inhibition of AChE/BuChE and MAO-A/B may have an enhanced effect in the treatment of AD²⁵⁸. The discovery of ladostigil further underscores the potential of inhibitors that concurrently target AChE and MAO to combat AD²⁵⁹. In another research, Bautista-Aguilera et al.²⁶⁰ employed 3D-QSAR modeling to guide the design of donepezil and tacrine derivatives, which inhibit MAO-A/B and AChE/BuChE. More recently, Mateev et al.²⁶¹ have repurposed the FDA-approved antiretroviral drug dolutegravir as a dual inhibitor of MAO-B and AChE, utilizing in silico consensus molecular docking techniques and biological evaluation. Furthermore, screening of natural compounds was performed to assess their potential as MTDLs against AD utilizing in silico methodologies to evaluate their inhibitory effects on AChE and BACE1 enzymes. The compound withanolide A emerged as effective, demonstrating an IC₅₀ of 107 μmol/L for AChE, 57.1% inhibition of BACE1 activity, and exhibiting no toxicity to SH-SY5Y cells²⁶². Additionally, antagonists of the histamine H3 receptor (H3R) have been implicated in neurodegenerative diseases associated with impaired cholinergic function, such as AD, while agonists of the 5-HT4 receptor (5-HT4R) have also been identified as promising therapeutic agents for AD^263,264. Lepailleur et al.²⁶⁵ utilized a pharmacophore-based methodology to screen the CERMN compound library, successfully identified an active compound with high affinity for H₃R antagonism and 5-HT4R agonism. In their study, compound 17 exhibited affinity K_i values of 41.6 and 208 nmol/L for H₃R and 5-HT4R, respectively, indicating its potential as a dual-target therapeutic agent for AD. Over the past few decades, extensive research has demonstrated that the Aβ and tau hypotheses significantly contribute to neurodegeneration in AD. Consequently, dual inhibitors targeting both Aβ and tau are increasingly recognized as promising therapeutic agents for AD treatment^256,266. BACE1 and GSK3β serve as principal targets for Aβ and tau, respectively, and have been identified as optimal candidates for multi-target drug development. A ligand-based virtual screening protocol was proposed to identify multifunctional fragments that target both BACE1 and GSK3β²⁶⁷. Among the identified fragments, compound 4 exhibited the most potent activity, demonstrating low micromolar dual inhibitory effects against BACE1 (IC₅₀ = 30.3 ± 3.5 μmol/L) and GSK3β (IC₅₀ = 0.7 ± 0.2 μmol/L). ML and QSAR methodologies have been effectively employed to discover ligands capable of binding to multiple targets in AD. A study utilized naive Bayesian and recursive partitioning algorithms to predict MTDLs for 25 key targets associated with AD by employing multitarget-quantitative structure–activity relationships (mt-QSAR). Through in vitro experimental validation, several potent MTDLs were identified, among which DL040 demonstrated nanomolar activity as a dual cholinesterase inhibitor targeting AChE and BuChE, as well as functioning as an H3 receptor antagonist²⁶⁸. Currently, multi-target drugs have achieved notable success in the treatment of complex diseases, including cancer and disorders of the central nervous system. Nevertheless, the complex interactions among multiple targets can result in off-target side effects²⁶⁹. The integration of computer-aided and AI-assisted drug design has significantly expedited the development of multi-target therapeutics. However, the precision and validation of computational models remain a challenge. Future progress in the field of multi-target drug discovery should prioritize advancing the understanding of disease biology and refining computational models to enhance their accuracy and validation, thereby minimizing off-target effects.

6. Drug repurposing in AD

Drug repurposing typically involves the investigation of a wide array of pharmaceutical compounds, including those that are outdated, unsuccessful, currently undergoing clinical trials, or already approved by the FDA, to identify novel therapeutic applications. This process employs diverse computational and experimental methodologies to expedite the initiation of new clinical uses. The transition from laboratory research to clinical application is thereby accelerated, resulting in significant reductions in both time and cost associated with drug development²⁷⁰. This approach is particularly advantageous in the realm of AD drug research and development. In the most recent AD drug development pipeline (2024), 39 agents have been repurposed across 52 clinical trials¹². Physics-based and AI-based approaches are widely used in drug repurposing studies. Next, some representative examples will be described for AD drug repurposing.

In the study conducted by Hassan et al.²⁷¹, the binding affinities and interactions of potential drug candidates were assessed using molecular docking, pharmacogenomics analysis, and MD simulations. The results of enzyme inhibition and kinetic mechanism analyses suggested that cinitapride may serve as a promising candidate for the treatment of AD²⁷¹. Furthermore, advances in high-performance computing hardware and multi-omics data have rendered AI a potent tool for identifying potential AD therapeutics from existing drug libraries²⁷². For example, Vignaux et al.²⁷³ utilized Bayesian ML models to investigate small molecules for the treatment of AD. Through this model, Ruboxistaurin, originally a therapeutic agent for Diabetes Mellitus, was repurposed as an inhibitor of GSK3β for AD. Furthermore, to identify novel indications for existing drugs, a ML framework known as drug repurposing in AD (DRIAD) was developed for drug repurposing in AD by analysing mRNA expression profiles and differential gene lists from diverse datasets. By analysing gene expression data from human brains at various stages of disease progression, the DRIAD framework can identify associations between gene activity and AD, thereby potentially informing future therapeutic strategies²⁷⁴. Apart from the DRIAD, NETTAG is an interpretable network topology-based DL framework to identify disease-associated genes and drug targets through GWAS and multi-omics approaches. By integrating human interactome network-based predictions and validations with patient data derived from electronic health records, the framework also identified gemfibrozil as a potential drug candidate associated with a reduced incidence of AD²⁷⁵. In a recent study, a novel affinity prediction model named multi-perspective neural network ensemble learning drug–target affinity (MNNEL-DTA) was developed utilizing a multi-perspective neural network ensemble learning approach for drug repurposing in AD. This model successfully identified a medin inhibitor (glutethimide) and a dual agonist of sigma-1 receptor and spleen tyrosine kinase (SYK), Memantine. Notably, Memantine, an approved drug for AD, and glutethimide, which has shown potential in preliminary studies for AD treatment, further substantiate the model's reliability in drug repurposing applications for AD²⁷⁶.

Currently, several candidate compounds identified through computational methodologies have entered clinical or pre-clinical trials, as detailed in Table 5^42,277,279, 280, 281, 282, 283, 284, 285. Notably, Fluoropezil represents a new generation of AChE inhibitors, designed utilizing a kinetically driven fast binding and slow dissociation theoretical calculation approach²⁷⁷, and is presently undergoing Phase II clinical trials in China (CTR20212054). Additionally, a compound identified by Sumitomo Pharma using Exscientia's AI technologies, DSP-0038, functions as a serotonin 5-HT_2A receptor antagonist and a serotonin 5-HT_1A receptor agonist^278,279. In 2021, DSP-0038 entered Phase I clinical trials for the treatment of AD-related psychosis in the United States. Nevertheless, the majority of drug candidates discovered through computational methods remain in the preclinical research phase. For instance, Brown et al. employed a structure-based drug design strategy to develop HTL9936, a selective M1-muscarinic acetylcholine receptor agonist²⁸⁰. Concurrently, the pharmacological properties of HTL9936 were assessed both in vivo and in vitro during early clinical studies. These findings indicated that HTL9936 holds promise as a candidate drug for addressing memory loss associated with AD. Additionally, several preclinical drug candidates, including KAI-11101²⁸¹, DK02²⁸²,SSZ²⁸³, and Q17²⁸⁴ were identified through virtual screening, as detailed in Table 5. Overall, the utilization of computational methods in the development of anti-AD drugs remains in its nascent stages, necessitating further time to evaluate the efficacy of these approaches. Although AI techniques have the potential to expedite the drug discovery process, their success in clinical trials still needs further vaidation. Nevertheless, this does not undermine the prospective role of AI in drug development. Furthermore, most computational methods serve as auxiliary tools and must be integrated with experimental techniques, such as biological activity testing and chemical synthesis, to achieve comprehensive drug development.

Table 5.

Drug molecules in clinical or pre-clinical study discovered by computational methods.

Drug name	Chemical structure	Drug design strategy	Target	Phase study	Ref.
PF-06648671		Pharmacophore model	γ-Secretase	Pre-clinical	42
Fluoropezil		Binding kinetics-driven drug design strategy	AChE	Phase II	277
DSP-0038	/	Exscientia's AI technologies	5-HT1A/5-HT2A	Phase I	279
HTL9936		Homology model, fragment-based drug design	M1-muscarinic acetylcholine receptor	Pre-clinical	280
KAI-11101		Molecular docking, free energy perturbation (FEP)	Dual leucine zipper kinase (DLK)	Pre-clinical	281
DK02	/	In silico virtual screening	AChE	Pre-clinical	282
SSZ		In silico virtual screening	Multi-targets	Pre-clinical	283
Q17		In silico virtual screening, MD simulation	DYRK1A	Pre-clinical	284
Compound 548		De novo generation, ML	iNOS	Pre-clinical	285

7. AI-driven biomarker discovery and precision medicine for AD

Within the framework of precision medicine, the integration of AI into AD research has exhibited considerable transformative potential, particularly in the areas of biomarker discovery and patient stratification. These advancements are crucial for overcoming the limitations associated with traditional diagnostic approaches, which are often invasive, expensive, and inaccessible. The identification of biomarkers capable of predicting responses to specific therapies is vital for improving drug efficacy, avoiding ineffective treatments, and advancing personalized medicine²⁸⁶. Furthermore, the capacity to predict which patients are most likely to benefit from particular treatments or experience lower risks is instrumental in optimizing therapeutic decision-making. AI technologies have emerged as powerful tools for addressing these challenges by facilitating the analysis of complex multi-omics data, including genomics, transcriptomics, proteomics, and metabolomics, as well as real-world data obtained from sources such as electronic health records, insurance claims, and patient registries²⁸⁷. AI-driven methodologies have significantly advanced the identification of biomarkers associated with AD progression and therapeutic response. For example, Hampel et al.²⁸⁸ conducted the first genome-wide search for biomarkers associated with AD treatment response through clinical studies of the selective sigma-1 receptor agonist blarcamesine. They identified key genomic variants and baseline mini-mental state examination scores as significant predictors of clinical outcomes, demonstrating the potential of precision medicine to tailor AD treatments. In another study, Guo et al.²⁸⁹ employed AI to examine extensive cerebrospinal fluid protein datasets from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Their analysis identified the protein YWHAG, along with several protein combinations such as YWHAG/SMOC1/PIGR/TMOD2, as exhibiting superior diagnostic accuracy compared to traditional biomarkers like Aβ₄₂ and p-tau. Similarly, Latifi-Navid et al.²⁹⁰ utilized AI-driven integrated multi-omics analysis to identify 13 critical AD hub genes and molecular pathways associated with AD progression in the brain, cerebrospinal fluid, and plasma. These discoveries facilitate the development of non-invasive biomarkers and elucidate key molecular processes underlying AD. To further advance biomarker and therapeutic target discovery in AD research, a network-based multi-omics data integration platform, termed “AD Atlas”, has been developed. This platform integrates high-quality multi-omics data through advanced graph-based analysis and AI techniques, highlighting the transformative potential of AI in expediting progress toward more effective and personalized interventions for AD²⁹¹. Collectively, studies utilizing AI highlight its capacity to reveal latent patterns within data, thereby enhancing our comprehension of AD and significantly advancing diagnostic precision and the development of targeted therapeutic strategies.

In conjunction with biomarker discovery, AI has proven exceptionally useful in patient stratification, a critical component of precision medicine aimed at addressing the clinical heterogeneity of AD. Stratification enables the categorization of patients into subgroups based on unique disease characteristics, genetic profiles, and anticipated treatment responses, thereby supporting the development of personalized therapies^292,293. By utilizing AI models that synthesize biomarker data, clinical information, and neuroimaging results, researchers have significantly advanced the understanding of disease progression and enhanced the efficiency of clinical trials. For example, Birkenbihl et al.²⁹⁴ employed DL techniques to examine the cognitive and functional trajectories of 283 early-dementia patients from the ADNI database, identifying two distinct subgroups characterized by “slow” and “fast” disease progression. These findings were externally validated in an independent cohort of 2779 patients from the National Alzheimer's Coordinating Center (NACC) database. A ML model trained on cross-sectional data achieved high accuracy in predicting progression subgroups, highlighting the potential of AI in patient stratification for clinical trials. By employing this classifier to simulate enrichment trials, a 13% reduction in trial costs was observed, highlighting its potential to accelerate drug development for cognitive impairment. Beyond the optimization of clinical trials, stratification at the molecular and subtype levels has unveiled disease mechanisms that were previously obscured by patient heterogeneity. For instance, Wen et al.²⁹⁵ utilized the unsupervised ML method, Surreal-GAN, to identify two distinct subtypes of AD: one subtype is characterized by diffuse atrophy and systemic pathologies, such as cardiovascular issues and inflammation, associated with non-APOE4 genetic factors; the other subtype is marked by medial temporal lobe atrophy driven by amyloid, tau, and APOE. These subtypes demonstrate distinct progression patterns, thereby elucidating subtype-specific therapeutic targets for precision treatments. In addition to stratification based on omics data, AI models have also achieved significant advancements in utilizing clinical records and neuroimaging for the early prediction and classification of AD risk. For instance, Tandon et al.²⁹⁶ identified eight cerebrospinal fluid peptides capable of stratifying individuals according to their risk of progressing from asymptomatic to symptomatic AD. Utilizing ML techniques, these peptides were able to accurately differentiate between control-like and AD-like subgroups among biomarker-positive, cognitively normal individuals. The validity of these findings was confirmed in an independent cohort, underscoring their robust potential for early risk assessment and informing the development of future DMTs.

In a recent study, an AI model employing bidirectional gated recurrent units demonstrated superior performance in predicting the progression from mild cognitive impairment to AD, surpassing other ML methods without the need for expert-driven feature selection. This model offers a clinical tool for early risk stratification using readily available healthcare data, thereby facilitating targeted interventions for high-risk patients²⁹⁷. Additionally, multimodal frameworks that integrate imaging and clinical data have shown the efficacy of combining diverse data types to capture the complex nature of AD. For instance, Chen et al.²⁹⁸ developed a multimodal medical data fusion framework that integrates the vision transformer and CNN to model brain magnetic resonance imaging (MRI) data and clinical records, respectively, resulting in enhanced performance for early AD diagnosis and recognition. A multi-task fusion language image question answering model (FLIQA-AD) was also introduced, which combines electronic health records with 3D MRI images using self-attention and cross-attention mechanisms within the ADFormer module, achieving state-of-the-art performance in AD detection and mini-mental state examination cognitive score prediction, while exhibiting robust zero-shot recognition capabilities on large datasets²⁹⁹. Guan et al.³⁰⁰ further demonstrated the efficacy of the CD-Tron clinical language model, which achieved high accuracy in the ultra-early detection of cognitive decline, with interpretability ensured through SHAP analysis. Recently, Lee et al.³⁰¹ developed a predictive prognostic model to determine whether patients with mild cognitive impairment would progress to AD, and validated it through longitudinal outcomes from various memory clinics. The framework was shown to be highly accurate and outperformed traditional clinical indicators, highlighting the ability of AI to generalize from controlled research datasets to real-world clinical settings.

8. Biotech start-ups and pharmaceutical companies leveraged AI and ML for AD drug discovery

As computing power and biological data advance, biotech startups and pharmaceutical companies are increasingly leveraging AI and ML to accelerate AD drug development and drive rapid innovation in this field. In 2019, the tensorial reinforcement learning model (GENTRL) developed by Insilico Medicine company was used to design and validate potent discoidin domain receptor 1 kinase inhibitors³⁰², thereby illustrating the potential of AI to expedite drug discovery processes³⁰³. In order to generate potential candidate molecules that conform to specified drug-like properties, a reinforcement learning-based approach is employed for the de novo design of targeted chemical libraries. This is achieved by optimizing molecular characteristics through multi-objective strategies. The methodology is demonstrated through its application in generating novel inhibitors for therapeutic targets such as neuraminidase, AChE, and the main protease of SARS-CoV-2²⁴⁹. Additionally, another AI pharmaceutical startup, BenevolentAI, has developed the Judgment Augmented Cognition System (JACS), an advanced AI-driven platform designed to extract and synthesize information from extensive biomedical datasets to develop novel clinical candidates targeting neurodegenerative diseases. JACS has been instrumental in identifying clinical candidates for AD and PD, as well as other rare conditions. Notably, the system identified two candidate compounds for AD as early as 2014, highlighting the transformative potential of AI-driven approaches in drug discovery³⁰³. Additionally, several companies are using AI approaches to enhance diagnostic and prognostic tools in AD research³⁰⁴. IGC Pharma, a pharmaceutical company committed to the development of innovative treatments for AD, is incorporating AI into its research efforts by developing a transformer-based model that diagnoses AD and dementia through the analysis of multi-modal clinical data sourced from databases such as the NACC, ADNI, and the Parkinson's Progression Markers Initiative (PPMI). This model improves diagnostic accuracy by examining a range of biomarkers and patient information, and it is currently being employed in the IGC-AD1 Phase II clinical trials to facilitate personalized treatment for AD. Nevertheless, the inherent complexity of AD presents significant challenges, and many startups in the field encounter a high risk of failure. Overall, while AI technology has demonstrated considerable potential in the realm of AD drug discovery, it remains in its nascent stages and offers substantial opportunities for future development.

9. Conclusion and future prospects

AD is caused by multiple pathological mechanisms and is one of the complex neurodegenerative diseases that is being tackled worldwide. Although a variety of pathological mechanisms related to AD have been proposed, it is still unclear which is the fundamental factor in the formation of AD. Currently, the number of AD drugs approved by the FDA is limited and cannot completely cure AD. Therefore, finding safer and more effective AD drugs remains an urgent task. In recent years, the development of structural biology, molecular modelling and AI technology has brought new hope to the field of AD research, including early diagnosis of the disease³⁰⁵ and drug development pipeline from target discovery to clinical research. AI has emerged as an indispensable tool for accelerating biomarker discovery and refining patient stratification in AD. By integrating multi-omics, neuroimaging, and clinical data, AI not only enhances diagnostic precision but also stratifies patients into biologically and clinically meaningful subgroups, enabling individualized interventions. These breakthroughs hold significant promise for optimizing clinical trials, accelerating drug discovery, and ultimately delivering precision medicine for a complex, heterogeneous disease like AD. Considering the diversity of AD pathogenesis, multi-target drugs will still be a hot research area in the future development of AD drugs³⁰⁶. CADD and AIDD-based methods have unique advantages in multi-target drug design, such as multi-target QSAR, multi-target molecular docking, and pharmacophore merging, etc.^307,308. However, the development of multi-target drugs still faces some challenges, such as improving their binding affinity with different targets, optimizing their pharmacokinetic properties, bioavailability, etc.³⁰⁹. In addition, the design of PROTACs targeting misfolded proteins provides possibilities of developing drugs against non-druggable AD targets, such as tau, Aβ protein, etc.^310,311. At the same time, PROTAC molecules can also target some kinases relevant with AD such as GSK3β, AKT, etc.³¹². On the other hand, allosteric modulators and the covalent inhibitors of tau and Aβ are also very attractive and feasible in AD drug development, providing more diverse strategies for the development of AD drugs³¹³. With the development of structural biology, more oligomer structures of tau and Aβ are constantly being resolved, which will provide important structural basis for drug design and understanding the mechanism of disease^314,315.

The development of accurate AI models necessitates high-quality, representative, unbiased, and comprehensive labelled data to enhance the accuracy of data feature learning and improve generalization. However, the confidentiality of experimental data within many pharmaceutical companies has resulted in “data islands”. Consequently, data sharing and resource integration are essential for AI technology to achieve significant advancements³¹⁶. Furthermore, the architecture of AI-based DL models often functions as a black box, lacking interpretability, which raises concerns regarding their credibility. Although computational virtual screening methods have expedited the drug discovery process, their success rate still needs improvement. Furthermore, hit compounds identified through virtual screening typically necessitate extensive structural modifications, as well as comprehensive in vivo and in vitro efficacy and toxicology evaluations, prior to advancing into preclinical research. Currently, only a limited number of drug molecules in clinical trials are designed using AI-based methodologies³¹⁷. Consequently, the AI-driven drug discovery process still faces significant challenges.

Author contributions

Qianqian Zhang and Xiaoqing Gong collected the literatures and wrote the original draft. Xiaojun Yao and Huanxiang Liu reviewed and revised the draft.

Declaration of generative AI in scientific writing

During the preparation of this work the author(s) used AI to polish the language. After using this tool, the author(s) reviewed and edited the content as needed and took full responsibility for the content of the publication.

Conflicts of interest

The authors declare no conflicts of interest.