Artificial intelligence revolutionizing CNS drug discovery and development

Abstract

Central nervous system (CNS) drug discovery faces high attrition rates, long timelines and substantial costs due to complex disease biology and the difficulties in safe drug delivery. Conventional CNS processes remain slow and trial-and-error driven. These challenges often result in poor brain penetration, off-target toxicity or limited efficacy after years of development. Recently, the integration of artificial intelligence (AI) with computer-aided drug design (CADD) has enabled more precise and scalable approaches for therapeutic development. AI-powered tools prioritize high-value analogs, streamlining design and optimization. This review provides an overview of how AI technologies are redefining early-stage CNS drug discovery, particularly for complex and underserved neurological diseases.

Introduction

Central nervous system (CNS) drug discovery remains one of the most formidable and persistently high-risk domains in current pharmaceutical research, whereas in many therapeutic fields, a single target or pathway dominates disease progression.^(p1) CNS disorders arise from multilayered interactions spanning genomics, proteomics, synaptic circuitry and emergent cognitive and behavioral processes.^(p2) Neurodegenerative and neuropsychiatric disorders, including Alzheimer’s disease (AD), Parkinson’s disease (PD), schizophrenia, amyotrophic lateral sclerosis (ALS) and psychiatric conditions, are not defined by a single drug target; rather, multiple complex and dynamic networks of dysregulated molecular targets are responsible for disease progression and incurable conditions.^(p3),(p4) This multiscale, high level of molecular complexity has historically limited the success of classical drug development processes in identifying effective therapeutics.^(p5)

In addition to biological complexity, the blood–brain barrier (BBB) is another major factor that impedes CNS drug discovery.^(p6) The BBB is a highly selective physiological barrier that tightly controls molecular entry into the brain. Many compounds with strong in vitro potency fail to achieve therapeutic exposure in the CNS due to low penetration, poor permeability or unfavorable physiological profiles.^(p7) This single interface eliminates a substantial proportion of CNS candidates during preclinical or early clinical evaluation.

The convergence of biological complexity and drug delivery constraints is reflected in the field’s poor historical performance in overall drug-development success rates. Approximately 85% of drugs fail in clinical development, with CNS programs showing the lowest probability of regulatory success and disproportionately high failure rates in Phase II and Phase III, when clinical efficacy must be demonstrated.^(p8) These failures are commonly attributed to induced neuronal toxicities and a lack of clinical benefit in large patient populations. CNS trials also require long treatment durations, specialized clinical end points and large heterogeneous populations. Owing to late-stage failures, CNS drug development typically exceeds 12–15 years and costs approximately USD 2.8 billion from early discovery to market approval.^(p9),(p10) The difficulty of examining CNS drugs directly in the living human brain, combined with the lack of predictive animal models and reliable molecular biomarkers, further complicates the process.^(p5),(p11) Together, these scientific and economic challenges have led many pharmaceutical companies in the USA and globally to reduce investments or even withdraw from CNS drug discovery and development.^(p12) This trend underscores the urgent need for innovative scientific and methodological frameworks to mitigate uncertainty and reduce attrition in late-stage CNS drug development.

Artificial intelligence (AI) has emerged as a transformative and essential tool for addressing long-standing challenges in CNS drug discovery and development.^(p13) It integrates large-scale biological, imaging, chemical and clinical datasets to create system-level models of disease complexity that traditional processes cannot achieve.^(p14) Within AI, machine learning (ML) algorithms are increasingly being used to analyze multiomics data, identify disease modules and map protein–protein and protein–drug interaction networks underlying neurodegeneration. More recently, deep learning (DL), a specialized branch of ML based on multilayer neural networks, has enabled the modeling of highly complex, nonlinear relationships in large biological and chemical datasets.^(p15) In parallel, reinforcement learning (RL) focuses on goal-driven learning through reward and penalty mechanisms, allowing models to iteratively optimize actions toward desired outcomes (Figure 1).^(p16) Together, these AI-powered paradigms provide the computational foundation for modern data-driven approaches in drug discovery and are increasingly applied to address the complexity and high failure rates associated with CNS drug development. Moreover, graph neural networks (GNNs) and knowledge graph systems enable the identification of hidden disease networks and multitarget opportunities.^(p17),(p18) This is particularly important because network dysregulation drives pathology in many CNS diseases. AI also encompasses a range of computational methods that enable machines to learn from large datasets and make predictions. Within this framework, ML and DL underpin modern generative artificial intelligence (GenAI) approahes and foundation models, such as deep neural networks (DNNs) and large language models (LLMs) (Figure 1). These approaches are interconnected and collectively support data-driven prediction, molecular design and decision-making in CNS drug discovery. Addressing this core complexity is generating renewed investment confidence: according to global CNS therapeutics reports, the CNS drug market size was estimated at USD 130.1 billion in 2024 and is projected to reach USD 254.6 billion by 2030.^(p19)

FIGURE 1.

Relationship hierarchy of AI methodologies. AI encompasses ML approaches, including SL, SSL and RL. DL underlies these approaches and enables the development of several models, such as GenAI, DNNs, GNNs and LLMs (created using BioRender).

Beyond target discovery, AI has rapidly restructured the medicinal chemistry workflow. AI builds generative models and AI-assisted virtual screening tools that enable chemists to explore CNS-relevant chemical space and optimize potency, physicochemical properties and CNS drug criteria. AI-generated predictive tools can identify pharmacokinetic (PK) properties, metabolic stability and safety liabilities early in development, enabling chemists to eliminate weak candidates before synthesis or in vitro and in vivo testing.^(p20),(p21) These capabilities dramatically reduce the design–make–test–analyze (DMTA) cycle, decrease the number of required analogs and shift failures to earlier, less costly stages of drug development.^(p22)

Furthermore, the integration of AI-enhanced CADD has reshaped pharmaceutical research by accelerating drug discovery and improving predictive accuracy throughout the development pipeline.^(p23) Recent AI models have enhanced the accuracy and specificity of CADD predictions for CNS drug targets, especially for targets lacking X-ray crystallography or cryo-EM structures, such as orphan receptors.^(p24) Billions of compounds can be screened using ML-accelerated virtual screening, enabling the prediction of potential hits before computationally intensive docking and thereby reducing computational costs.^(p25) Beyond discovery, AI enhances translational and clinical decision-making by integrating neuroimaging, CNS patient data and digital phenotyping data. CNS disease-specific biomarker profiling further refines indication selection and patient stratification, which is particularly important in heterogeneous CNS disorders where treatment responses vary widely between individuals.^{(p26),(p27),(p28)}

The incorporation of AI-based models into established CADD frameworks is transforming modern drug discovery, enabling researchers to expedite the identification of novel drug candidates and optimize their properties with unprecedented speed and efficiency.^(p23) This approach directly addresses bottlenecks such as high preclinical attrition rates and limited interpretability of conventional models.^(p29) Given the renewed interest in applying AI to neuroscience and the growing need for faster, more cost-effective CNS drug development, this review examines how AI technologies are being deployed across the discovery pipeline. We focus specifically on how AI reshapes medicinal chemistry workflows, shortens development timelines, reduces clinical risk and enables novel strategies to maximize the potential for addressing complex and underserved neurological diseases.

Artificial intelligence in the central nervous system drug discovery pipeline

In CNS drug discovery, AI encompasses a broad class of computational methods that support human decision-making by learning patterns from complex biological data, as illustrated in Figure 2 and Table 1.^(p26) Within this framework, ML enables predictive modeling across multiple stages of CNS drug discovery.^(p5) Supervised learning (SL) is widely used to predict experimentally defined end points, including target bioactivity, BBB permeability, efflux transporter interactions, CNS safety liabilities and brain-to-plasma exposure ratios, enabling the early elimination of compounds unlikely to achieve sufficient CNS penetration.^(p30) Self-supervised learning (SSL), in turn, addresses the limited availability of labeled CNS datasets by pretraining molecular representations on large-scale, unlabeled chemical datasets and transferring them to downstream predictive tasks.^(p31) RL complements these paradigms by enabling CNS-constrained multiobjective optimization, particularly in molecular design problems that require balancing potency, BBB permeability, CNS multiparameter optimization (CNS-MPO) exposure and development ability within a unified reward framework.^(p32) DL is the subset of ML that underlies these approaches, using DNNs that support nonlinear and multitask learning.^(p15) Architecture-specific DL models further align with drug discovery data structures, with convolutional neural networks (CNNs) capturing spatial features from CNS-specific protein–ligand complexes for G protein-coupled receptors (GPCRs), ion channels and neuroinflammatory receptors and GNNs modeling molecular topology and biological interaction networks.^(p33)

FIGURE 2.

AI-assisted drug discovery and development pipeline for CNS disorders. Genetic (e.g. genome-wide association studies) and multiomics data (epigenomics, transcriptomics and proteomics) guide target prioritization and validation at molecular, cellular, chemical and in vivo levels. AI-supported hit and lead discovery (de novo drug design, virtual screening, drug–target interaction prediction and drug repurposing) and lead optimization for drug-like properties, pharmacokinetics/pharmacodynamics and safety. Optimized candidates progress to preclinical evaluation and clinical trials, where AI accelerates efficacy testing, safety assessment and patient stratification in CNS disorders (created using BioRender).

TABLE 1.

Overview of AI models discussed in this review and their primary roles in CNS drug discovery.

AI category	Model type	Key function	Example applications in CNS drug discovery
ML	SL	Learns relations from labeled datasets	Prediction of bioactivity, BBB permeability, ADMET properties and CNS toxicity
	SSL	Learns molecular representations from large unlabeled datasets	Pretraining molecular embeddings for downstream prediction tasks
	RL	Optimizes decisions using reward-based learning	Multiobjective optimization of potency, BBB permeability and CNS-MPO properties
DL	DNNs	Multilayer neural architectures capturing nonlinear relations	Prediction of molecular properties and multitask learning across CNS end points
	CNNs	Extract spatial features from structured biological data	Analysis of protein–ligand interactions for GPCRs, ion channels and neuroinflammatory targets
	GNNs	Model graph-based molecular and biological networks	Molecular topology analysis and disease–gene–drug network modeling
Generative AI	GANs	Generate novel molecular structures through adversarial learning	De novo molecular design for CNS-active compounds
	VAEs	Latent-space molecular generation and chemical space exploration	Generation of compounds with optimized CNS-MPO properties
Language-based AI	LLMs	Extract and synthesize biomedical knowledge from text data	Literature mining, hypothesis generation and target identification
	NLP models	Text mining and semantic analysis of biomedical documents	Extraction of gene–disease associations and clinical insights
Network-based AI	Knowledge graph models	Integrate heterogeneous biomedical data into relational networks	Disease–gene–drug relation discovery
Systems-level AI	Digital twin models	Simulate disease progression and treatment response	Virtual patient modeling and prediction of therapeutic outcomes

In the CNS domain, GenAI extends predictive modeling by enabling de novo molecular design and systematic exploration of chemical space.^(p34) Generative adversarial networks (GANs) and variational autoencoders (VAEs) represent complementary generative strategies and are typically integrated with predictive models and RL to prioritize compounds for CNS-MPO alignment, BBB permeability, efflux avoidance and reduced risk of central toxicity.^(p34),(p35) LLMs and natural language processing (NLP) methods facilitate large-scale extraction and synthesis of biomedical knowledge from scientific literature, patents, clinical narratives, CNS-specific trial data, neuroimaging biomarkers and gene–disease associations relevant to neurological disorders, thereby supporting target identification and evidence integration.^(p36) Knowledge graph-based models formalize these heterogeneous data into structured networks, enabling reasoning across disease–gene–drug relations in complex CNS disorders such as AD, PD and major depressive disorder.^(p18) At the system level, digital twin-based AI models integrate molecular predictions, biological networks and clinical data to simulate disease progression and therapeutic response. Together, these AI models serve as powerful components of a single pipeline that improves prediction, ranking and optimization in CNS drug discovery.^(p37) They can more efficiently support human intelligence by making fast, accurate predictions, accelerating early screening and helping researchers focus on the most promising compounds. Meanwhile, these models complement laboratory experiments and expert judgment, guiding better decisions for advancing compounds toward clinical development.

In the early stages of target identification, AI and ML models integrate genomic, transcriptomic, proteomic, epigenetic and brain imaging data to analyze molecular pathways underlying disease causation.^(p38) By integrating system-level analysis with literature mining, AI enables the discovery of novel targets. These novel targets feed into hit finding and lead optimization, where DL models rapidly evaluate large chemical libraries and prioritize compounds with CNS-relevant features.^(p39) Generative models can assist in predicting novel chemical structures optimized for CNS drug-like properties during de novo design.^(p40) The most impactful applications of AI in CNS discovery are the early prediction of PK and safety properties. AI-enabled models predict absorption, distribution, metabolism, excretion and toxicity (ADMET) properties, BBB permeability, glycoprotein efflux risk and potential CNS-specific toxicities before compounds enter experimental testing.^(p20) Beyond preclinical development, AI increasingly supports clinical trial design by analyzing AI-driven neuroimaging data, digital phenotyping and real-world patient information, as depicted in Figure 2.^(p41)

Representative AI-based tools for target identification, molecular design, ADMET prediction and synthesis planning in CNS drug discovery are summarized in Table 2. These resources span the full continuum of drug development, from early target identification to late-stage optimization and translational decision-making. Platforms such as Open Targets,^(p42) INDRA,^(p43) Hetionet^(p44) and PyKEEN^(p45) support AI-enabled integration of genomics, transcriptomics, proteomics and disease-network information to prioritize CNS-relevant targets and uncover gene–disease–drug relations in neurological disorders. Structure-focused open platforms, including Alpha-Fold Protein Structure Database,^(p46) ColabFold^(p47) and RoseTTAFold,^(p48) facilitate AI-driven protein structure prediction for CNS targets that lack high-resolution X-ray crystallographic or cryo-EM structures, while GNINA,^(p49) DiffDock^(p50) and AutoDock-GPU^(p51) enable structure-based hit discovery through DL-assisted docking and large-scale virtual screening.

TABLE 2.

Open-source AI tools and platforms support different stages of the CNS drug discovery pipeline.

Drug discovery stage	Tool/platform	Primary AI application in CNS drug discovery
Target identification and disease biology	INDRA	Automated literature mining and causal network modeling for disease pathways, including neurological disorders
	Hetionet	Biomedical knowledge graph integrating disease–gene–drug relations for target discovery and drug repurposing
	PyKEEN	Knowledge graph embedding framework for predicting drug–target–disease interactions
	Open Targets Platform	Integrative platform combining genomics, genetics and disease evidence for target identification
Protein structure prediction	AlphaFold Protein Structure Database	AI-predicted protein structures supporting structure-based drug discovery for CNS targets
	ColabFold	Accelerated AlphaFold pipeline enabling rapid protein structure prediction
	RoseTTAFold	DL-based protein structure prediction model for unresolved CNS targets
Structure-based hit discovery	GNINA	DL-based molecular docking and scoring for protein–ligand interactions
	DiffDock	Diffusion-based DL model predicting protein–ligand binding poses
	AutoDock-GPU	GPU-accelerated molecular docking enabling large-scale virtual screening
De novo molecular design	REINVENT4	RL-based generative molecular design optimized for multiparameter objectives such as CNS-MPO
	DrugEx	Deep RL framework for de novo drug design and lead optimization
	MOSES	Benchmarking platform for generative molecular models used in drug design
Molecular property prediction	DeepChem	Open-source ML toolkit for QSAR modeling, ADMET prediction and molecular property prediction
	Chemprop	GNN framework for molecular property prediction
BBB permeability prediction	DeepBBBP	DL model predicting BBB permeability from molecular descriptors
	GCN-BBB	GNN framework for predicting BBB permeability
CNS penetration modeling	CANDID-CNS	AI model predicting CNS penetration including stereochemistry and beyond R05 molecules
Synthesis planning/retrosynthesis	ASKCOS	AI-driven retrosynthetic planning and synthetic feasibility analysis for CNS-optimized candidates, enabling rapid evaluation of molecules designed for BBB penetration and CNS-MPO
	AiZynthFinder	DL-based retrosynthesis tool for predicting synthetic routes of CNS-active compounds, supporting efficient DMTA cycle progression and scalable synthesis of structurally complex brain-penetrant molecules

In the medicinal chemistry and hit-to-lead stages, open-source AI frameworks such as REINVENT4^(p52) and DrugEx^(p53) enable property prediction, CNS-specific optimization and de novo molecular design, allowing chemists to efficiently explore CNS-relevant chemical space while balancing potency, physicochemical properties and safety considerations. MOSES provides benchmarking support for evaluating generative molecular models.^(p54) DeepChem and Chemprop support graph-based Quantitative Structure-Activity Relationship (QSAR) modeling and prediction of CNS-relevant molecular properties, including ADMET and pharmacokinetic end points.^(p55),(p56) Specialized BBB prediction models, such as DeepBBBP and GCN-BBB, employ DL and GNN architectures to estimate BBB permeability from molecular descriptors and structural features.^(p57),(p58) In the case of CNS penetration modeling, tools such as CANDID-CNS extend beyond simple permeability prediction by incorporating stereochemical considerations and beyond the rule-of-five (R05) properties to estimate brain exposure more comprehensively.^(p59) Finally, AI-assisted retrosynthesis platforms, including ASKCOS and AiZynthFinder, support synthetic feasibility analysis, directly accelerating the DMTA cycle.^(p60),(p61) Collectively, the open-source tools listed in Table 2 illustrate how AI functions as an integrated, cross-cutting enabler across CNS drug discovery stages, from target identification to CNS penetration modeling, reducing uncertainty, shortening development timelines and shifting attrition toward earlier and less costly stages.

Artificial intelligence improving medicinal chemistry efficiency

In CNS drug discovery, medicinal chemists have traditionally operated within slow, sequential DMTA cycles characterized by empirical decision-making and the late recognition of failure. Optimization efforts often depended on trial-and-error adjustments of potency and physicochemical properties, combined with crucial liabilities such as limited BBB penetration, unfavorable PK or neurological safety concerns, which frequently emerged only after extensive synthesis and in vivo testing. This late-stage feedback contributed substantially to prolonged timelines, high costs and elevated attrition in CNS programs. These challenges are particularly pronounced in CNS drug discovery because compounds must simultaneously satisfy stringent constraints on BBB penetration, brain exposure and neurological safety, which are often difficult to predict using traditional medicinal chemistry approaches.^(p7)

The integration of AI has fundamentally altered this workflow by embedding predictive intelligence directly into daily medicinal chemistry decisions. Rather than generating molecules through largely heuristic reasoning, chemists now use AI-guided prioritization to rank candidate structures based on predicted CNS-relevant properties, including target engagement, brain exposure, safety liabilities and synthetic feasibility. CNS-MPO scores integrate weighted physicochemical parameters including ClogP, ClogD at physiological pH, molecular weight, topological polar surface area, pKa of the most basic center and hydrogen bond donor count to estimate BBB permeability and CNS exposure.^(p62) Recent AI-driven frameworks have extended this concept by learning nonlinear relations among these descriptors, enabling generative and predictive models to directly optimize CNS-MPO-aligned property spaces rather than relying solely on rule-based thresholds.^(p32),(p63) Additionally, generative models further accelerate discovery by proposing novel, CNS-optimized scaffolds under predefined constraints, allowing chemists to focus on evaluation, refinement and strategic selection rather than exhaustive manual ideation. At the same time, rapid in silico profiling of key ADME properties (BBB permeability and efflux risk) enables the early elimination of weak candidates, effectively shifting failure from costly experimental stages to low-cost computational filtering. Importantly, these AI-driven predictions allow medicinal chemists to simultaneously evaluate multiple CNS-relevant constraints, including potency, BBB permeability, pharmacokinetic stability and safety liabilities.

Therefore, AI compresses DMTA cycles from months to weeks, reduces the number of required analogs and enables the same medicinal chemistry teams to test significantly more hypotheses per year without increasing resources. Importantly, AI-powered analysis also provides earlier insight into probable causes of clinical failure, including insufficient efficacy, safety risks or suboptimal drug-like properties, allowing programs to be terminated earlier and more rationally. In this transformed paradigm, the medicinal chemist is no longer primarily a trial-and-error optimizer but a data-informed decision-maker whose expertise is amplified by AI-driven prediction, knowledge integration and prioritization. In practice, medicinal chemists increasingly rely on AI-enabled web servers that provide rapid predictions of CNS-relevant properties, safety risks and synthetic feasibility; representative examples of such decision-support tools are summarized in Table 3.

TABLE 3.

AI-enabled web servers for ADMET, BBB and decision-support platforms used in CNS medicinal chemistry and drug discovery.

Research group/organization	Type of AI/model used	Primary use for medicinal chemists (CNS context)
SIB Swiss Institute of Bioinformatics	Rule-based + statistical ML models	Rapid assessment of drug-likeness, physicochemical filters and CNS-relevant properties
BioSig Lab, University of Queensland	Graph-based ML	Prediction of PK, ADMET, BBB permeability and toxicity risk
LMMD, East China University of Science and Technology	ML classifiers/regressors	Early ADMET and toxicity screening of CNS candidates
CBDD Group, Central South University	ML + DL ensemble models	Comprehensive ADMET profiling and CNS safety evaluation
Charité – Universitätsmedizin Berlin (Preissner Lab)	ML and DL	Prediction of acute and organ-specific toxicity, including CNS risk
Integrative Biomedical Informatics Group (IMIM-UPF)	Data mining and scoring models	Gene–disease association discovery for CNS target validation
STRING Consortium	Probabilistic network inference	Protein–protein interaction networks for CNS pathway analysis
DrugBank Team, University of Alberta	Curated database + computational inference	Drug–target relations and CNS drug repurposing support
Frontiers in Genetics (reported tool)	Supervised ML models	Prediction of BBB permeability to filter CNS-active compounds early
BBBper developers (academic group)	ML classification models	Binary and probabilistic prediction of BBB penetration for CNS drug candidates
SSBio Group, Chung-Ang University	Regression-based ML models	Quantitative prediction of LogBB (brain-to-blood concentration ratio)
QCDevs (Hugging Face Space)	LightGBM/ML classifier	Rapid BBB permeability classification through a web interface
bio.tools community	LightGBM-based ML model	Screening of BBB penetration likelihood for CNS-focused compounds
Greenstone Bio	ML ADMET prediction models	CNS-relevant ADMET and toxicity prediction for medicinal chemistry triage
NAR Web Server (academic developers)	DL pharmacokinetic models	Prediction and analysis of PK parameters and toxicity risks
vNN Research Group	Variable nearest-neighbor (ML-based) models	Prediction of multiple ADMET and toxicity end points, including CNS safety

LightGBM: Light Gradient-Boosting Machine.

Platforms such as SwissADME,^(p64) pkCSM,^(p65) admetSAR 2.0^(p66) and ADMETlab 3.0^(p67) are frequently used to evaluate physicochemical properties, PK properties, ADMET profiles and CNS-relevant safety attributes that influence brain exposure, including lipophilicity, polarity and predicted BBB permeability. These tools help medicinal chemists rapidly assess whether candidate molecules satisfy key criteria for CNS-MPO and are therefore more likely to achieve adequate brain penetration. Toxicity-focused prediction tools, including ProTox-II/ProTox 3.0,^(p68) Greenstone Bio and the vNN web server, assist in identifying potential organ-specific and neurological toxicity risks, which are particularly crucial in CNS drug development, where compounds interact directly with brain tissue.^(p20),(p69) For BBB assessment, dedicated predictors such as BBB prediction servers, BBPpredict,^(p70) BBBper,^(p71) LogBB tools^(p72) and ML-based classifiers hosted on platforms such as QCDevs are used to estimate brain penetration and efflux transporter interaction risks.^(p73) In addition, knowledge-based resources, including DisGeNET,^(p74) STRING^(p75) and DrugBank,^(p76) support gene–disease association analysis, pathway interpretation and drug–target relation assessment for neurological disorders, thereby reducing unnecessary synthesis and animal studies and supporting more informed decision-making in early CNS drug development. Together, these AI-enabled servers allow medicinal chemists to rapidly screen drug candidates against CNS-specific constraints, thereby reducing unnecessary synthesis and animal studies and supporting more informed decision-making in early CNS drug development.

Artificial intelligence drug discovery across neurological disorders

Alzheimer’s disease and dementia

Given its multifactorial nature, AD encompasses cholinergic pathway dysfunction, amyloid pathology, tau aggregation, synaptic dysfunction, neuroinflammation and genetic predisposition.^(p3) Another relevant neurological disorder is dementia, a clinical syndrome characterized by progressive cognitive decline. AD is the most common cause of dementia, accounting for the majority of cases.^(p77) AI is being used to integrate vast, heterogeneous data sources, including genomics, proteomics, neuroimaging and clinical records, to identify novel therapeutic targets and drug candidates.^(p37) AI models can mine large datasets to detect molecular signatures and stratify AD subtypes, improving patient-specific drug development.^(p37) Moreover, AI-based digital twins and simulation frameworks enable in silico testing of compounds and modeling of disease progression, significantly reducing the costs and duration of early-phase trials. These advancements are complemented by AI-driven phytochemical screening and multiomics integration, which facilitate the identification of new small molecules and repurposed drugs targeting neuroinflammation, a central contributor to AD pathogenesis.^(p78),(p79) A few research groups, pharmaceutical companies and biotech firms are harnessing AI models to discover small-molecule drugs by screening compounds and analyzing preclinical and clinical data before moving into clinical trials. Several representative examples of AI-developed drugs for AD are highlighted in Table 4. DSP-0038 is an AI-designed drug developed by the Japanese pharmaceutical company Sumitomo Dainippon Pharma using Exscientia’s (now Recursion’s) AI technologies.^(p80) This drug is currently in Phase I clinical trials and is designed as a dual-acting agent, an antagonist at the 5-HT_2A receptor and an agonist at the 5-HT_1A receptor. The company claims to have developed its own AI models capable of converting complex biological challenges into clinical candidates more quickly and with high accuracy.^(p81) In preclinical studies, DSP-0038 has demonstrated antipsychotic-like activity with reduced extrapyramidal side effects compared with traditional dopamine receptor antagonists, suggesting potential advantages for managing neuropsychiatric symptoms associated with AD. These findings support its development as a next-generation serotonergic antipsychotic candidate optimized using AI-driven design strategies.^(p82) Another AI-enabled drug candidate is NMRA-511, developed by Neumora Therapeutics, Inc., using a data science-based precision neuroscience platform that integrates genetic, clinical and biomarker datasets to identify novel CNS targets.^(p83) NMRA-511 acts as an antagonist of the vasopressin 1a receptor (V_1aR), a signaling pathway implicated in social behavior, stress responses and neuropsychiatric symptoms associated with AD.^(p84),(p85) In early-stage clinical investigations, the compound demonstrated favorable pharmacokinetic properties and tolerability, supporting further evaluation of modulation of vasopressin signaling as a therapeutic strategy for neuropsychiatric symptoms in neurodegenerative disorders.

TABLE 4.

Representative drug candidates developed using AI across the CNS drug design and discovery pipeline.

Drug name (company, country)	Chemical structure	Target indication/description	AI relevance	Clinical trial status
DSP-0038 (Recursion Pharmaceuticals, USA)	Undisclosed	5-HT2A receptor antagonist and 5-HT1A receptor agonist. Efficacy in the treatment of behavioral and psychological symptoms of dementia	The drug was designed using Exscientia’s AI technology	Phase I
NMRA-511 (Neumora Therapeutics, Inc., USA)	Undisclosed	Antagonist of the V1a receptor. The drug used for dementia due to AD	This drug was designed using Neumora’s AI technology	Phase Ib
MTS-004 (METiS Therapeutics, China)	Undisclosed	This drug is used to treat the pseudobulbar effect that commonly occurs in ALS and PD	METiS TechBio’s self-developed AI-enabled nanodelivery platform to design, optimize and formulate through AI algorithms	Phase III
AIT-101 (Orphai Therapeutics, Inc., USA)		Focuses on ALS and patients with PD with the inhibition of PIKfyve kinase	Used AI for the drug development process	Phase II
VRG50635 (Verge Genomics, Inc., USA)	Undisclosed	Inhibitor of PIKfyve, a therapeutic target of ALS	Verge Genomics used CONVERGE™, Verge’s all-in-human, AI-powered platform	Phase I failed
Dexmedetomidine (BioXcel Therapeutics, Inc., USA)		Used to treat agitation associated with schizophrenia or bipolar I or II disorder	Used AI approaches to identify and develop transformative medicines in neuroscience	FDA approval of IGALMI™
NMRA-898 (Neumora Therapeutics, Inc., USA)	Undisclosed	M4-positive allosteric modulator	Neumora used AI and ML algorithms to identify and quantify behavior	Phase I
NMRA-861 (Neumora Therapeutics, Inc., USA)	Undisclosed	M4-positive allosteric modulator	Neumora used AI and ML algorithms to identify and quantify behavior	Phase I
NMRA-140 or Navacaprant (Neumora Therapeutics, Inc., USA)		Novel KOR antagonist	Neumora used AI and ML algorithms to identify and quantify behavior in neuropsychiatric disease	Phase III
Fasoracetam (Nobias Therapeutics, Inc., USA)		Use for the neuropsychiatric symptoms associated with 22q11.2 deletion syndrome	Used AI-based deep phenotyping techniques	Phase II
SOM1311 (SOM Biotech, S.L., Spain)	Undisclosed	Small-molecule chaperone for phenylketonuria, behavioral problems and mental disorders	SOM Biotech used AI to screen new indications for existing drugs, clinical development and reluminate the cost and time	Phase II
DSP-2342 (Sumitomo Pharma Co., Ltd, Japan; Exscientia plc., UK)	Undisclosed	Serotonin 5-HT2A and 5-HT7 receptor antagonist	Sumitomo Pharma collaborated with Exscientia (Recursion) to use their AI platform to screen libraries	Phase I
Bevantolol hydrochloride or SOM 3355 (SOM Biotech, S.L., Spain)		β-Adrenergic blocker and a VMAT1 and VMAT2 inhibitor	Used AI drug screening platform operated by SOM	Phase II
Sulindac or HLX-0201 (Healx Ltd, UK)		Repurposed drugs used to treat FXS	Healx’s novel omics-based drug matching methods using an Al and ML model	Phase II
Gaboxadol or HLX-0206 (Healx Ltd, UK)		Repurposed drugs used to treat FXS	Healx’s Al platform that finds novel connections between drugs and diseases	Phase II
REC 994 (Recursion Pharmaceuticals, USA)		Reduces the ROS in CCM2-deficient endothelial cells	Recursion used ML algorithms to prepare a compound dataset	Phase II

CCM2: cerebral cavernous malformation 2; ROS: reactive oxygen species.

Amyotrophic lateral sclerosis

ALS is another multifactorial progressive neurological condition characterized by the loss of upper motor neurons in the motor cortex and lower motor neurons in the brainstem and spinal cord.^(p86) In 97% of ALS cases, TAR DNA-binding protein 43 was found to be misfolded and aggregated in neurons and glial cells.^(p87) There are a few approved drugs that are used in late-stage patients. These drugs slow down the progression of the disease, but nothing truly stops or reverses ALS. Owing to the complex and heterogeneous disease biology, researchers are struggling to discover new drugs for ALS.^(p88) In recent years, several innovative pharmaceutical companies have developed AI-based small-molecule drugs that have progressed through various stages of clinical trials, as shown in Table 4. MTS-004 is the first AI-enabled drug in China to undergo Phase III clinical trials. It was developed by METiS TechBio employing AiTEM, its proprietary AI-based platform for small-molecule design and optimization.^(p89) This drug is reported to treat the pseudobulbar affect that commonly occurs in ALS and Parkinson’s disease.^(p90) Early clinical studies indicated that the compound demonstrates favorable tolerability and symptomatic improvement in patients with pseudobulbar affect, supporting its progression into late-stage clinical evaluation. Another AI-discovered drug candidate, AIT-101 (1), was acquired by Orphai Therapeutics. It is an endosomal Phosphatidylinositol-3-Phosphate 5-Kinase (PIKfyve) inhibitor.^(p83) Mechanistic studies suggest that PIKfyve inhibition increases the rate of transcription factor EB (TFEB) activation, leading to its translocation from the cytoplasm to the nucleus.^(p91) Later, the lysosomal and autophagosomal activity of misfolded proteins is activated in ALS, AD, Huntington’s disease (HD), PD and dementia.^(p92) This mechanism highlights the therapeutic potential of targeting lysosomal–autophagy pathways to enhance the clearance of toxic protein aggregates that contribute to neurodegenerative disease pathology. This drug candidate shows the reverse pathological and functional consequences in animal models by depleting the nuclear TFEB in neural tissues.^(p91) A US-based pharmaceutical company, Verge Genomics, introduced its first AI drug candidate, VRG50635, in 2022. They used the CONVERGE™ AI model to investigate human data and human model systems throughout the development process. This AI model provides insights into the complex biological underpinnings of ALS. The candidate targets molecular pathways associated with neuronal degeneration and is currently undergoing clinical evaluation for ALS, demonstrating how AI-driven platforms can accelerate the translation of human genomic insights into therapeutic candidates.

Parkinson’s disease

AI is emerging as a transformative tool for addressing the complex, heterogeneous nature of PD, particularly in the search for disease-modifying therapeutics.^(p3) By leveraging multiomics data, electronic health records, neuroimaging and wearable biosensor outputs, AI enables early identification of PD biomarkers, classification of disease subtypes and acceleration of drug repurposing pipelines.^(p3),(p93) DL models trained on speech analysis, gait patterns and motor biomarkers have shown great promise for early PD diagnosis, allowing earlier therapeutic intervention and patient stratification in clinical trials.^{(p94),(p95),(p96)} AI has been employed to predict patient response in PD by integrating genomic, proteomic and biomarker data, enabling precision trial design and personalized therapeutic strategies.^{(p58),(p94),(p97)} Moreover, AI/ML-integrated nanocarrier systems are facilitating precise delivery of therapeutic agents across the BBB, an enduring challenge in PD pharmacotherapy.^(p98),(p99) Today’s medicinal chemists use different AI models to screen drug molecules for PD. Several drug candidates are at different stages of clinical development, as shown in Table 4. AI-derived candidates such as MTS-004 (METiS TechBio) and AIT-101 (Orphai/Verge Genomics), previously discussed in the section ‘Amyotrophic lateral sclerosis’, are also being evaluated in open-label clinical studies for neurodegenerative indications, including PD.^(p89),(p100)

Schizophrenia

AI addresses the long-standing challenges in schizophrenia drug discovery, a field hampered by biological complexity, diagnostic ambiguity and limited pharmacological innovation.^(p101) AI-driven approaches now enable the analysis of multimodal datasets, including genomics, neuroimaging, behavioral phenotypes and electronic health records, to unravel the heterogeneity of schizophrenia subtypes and predict individualized treatment responses.^(p102) DL and NLP are enhancing drug repurposing efforts and refining early-phase diagnostics by analyzing symptom patterns, polygenic risk scores and real-world clinical narratives.^(p103) Moreover, AI-powered modeling has contributed to dopaminergic target optimization and the identification of novel neurotransmitter pathways beyond traditional D₂-receptor antagonism, which is crucial for tackling treatment-resistant schizophrenia.^(p96),(p103) Dexmedetomidine (BXCL501, 2) is an FDA-approved drug marketed as IGALMI™.^(p104) BioXcel Therapeutics used AI technology to identify BXCL501 (2) as a selective a2-adrenoceptor agonist for the treatment of agitation associated with bipolar I or II disorder or schizophrenia.^(p104) Neumora, a pioneering neuroscience drug development company, has used AI-assisted computational tools and precision neuroscience platforms to develop two drugs (NMRA-898 and NMRA-861) that have progressed to Phase I clinical studies. Both drugs are M4 muscarinic receptor-positive allosteric modulators being investigated for schizophrenia. By selectively enhancing M4 receptor signaling, these candidates indirectly modulate dopamine D₂ receptor signaling, which is central to the pathophysiology of schizophrenia and may offer antipsychotic effects with a potentially lower risk of extrapyramidal side effects (Table 4).^(p105)

Psychiatric disorders

Psychiatric disorders are multifactorial diseases, including depression, schizophrenia, bipolar disorder, anxiety and post-traumatic stress disorder (PTSD).^(p106) These conditions arise from multiple factors, such as neurotransmitter imbalances, disrupted signaling pathways, altered receptor function, genetic vulnerability, environmental stress/trauma, neuroinflammation and neuroendocrine dysregulation.^(p107) Among the causes, the dopaminergic system (dopamine), the serotonergic system (serotonin; 5-HT) and the NMDA receptor in the glutamatergic system are the most prevalent.^{(p108),(p109),(p110)} According to the WHO’s report, nearly one in every seven people suffers from a mental disorder.^(p111) To date, more than 1.1 billion people are affected by mental depression or anxiety.^(p111) A list of drug candidates and their design and development processes is described in Table 4. Navacaprant (NMRA-140, 3), a highly selective kappa-opioid receptor (KOR) antagonist developed by Neumora Therapeutics, has shown promising results in a Phase III clinical trial as a monotherapy for major depressive disorder, demonstrating significant improvement in depressive symptoms and a favorable safety and tolerability profile.^(p112) Preclinical and emerging clinical evidence also suggests its potential utility in addressing mood-related and negative symptoms of schizophrenia, leveraging its nondopaminergic mechanism to target treatment-resistant subpopulations.^(p112) Neumora identified this candidate drug using AI- and ML-driven precision neuroscience platforms that integrate clinical, genomic and behavioral datasets to identify novel CNS targets.

Fasoracetam, another drug candidate developed by Nobias Therapeutics using AI-based deep phenotyping approaches, has completed Phase I clinical trials and is currently in Phase II evaluation for neuropsychiatric conditions.^(p113) SOM311 is a small-molecule chaperone for behavior-related mental disorders, discovered by SOM Biotech. Currently, this drug candidate is under investigation in a Phase II clinical trial, with preclinical studies suggesting potential benefits in modulating neurological dysfunction. In addition, DSP-2342, developed by Sumitomo Pharma in collaboration with Exscientia, was evaluated in a completed Phase I clinical trial as a serotonin 5-HT_2A and 5-HT₇ receptor antagonist.^(p114) This compound was identified by Sumitomo Pharma, in collaboration with Exscientia, using AI-driven medicinal chemistry and data science-based drug design approaches, highlighting the growing role of AI platforms in accelerating CNS drug discovery.^(p114)

Rare neurological diseases

HD, fragile X syndrome (FXS) and cerebral cavernous malformations (CCM) are rare neurological disorders driven by genetic mutations that lead to distinct abnormal protein functions and neurological pathologies.^{(p115),(p116),(p117)} HD is caused by a CAG trinucleotide repeat expansion in the huntingtin (HTT) gene, which produces misfolded and aggregated mutant HTT protein in neurons.^(p118) SOM Biotech introduced a new drug candidate, bevantolol (SOM-3355, 5), which is currently undergoing Phase II clinical evaluation.^(p119) This company employs an AI-based drug-screening platform to identify existing drugs for repurposing into new therapeutic indications. SOM-3355 (5) acts as a β-adrenoceptor blocker and inhibitor of vesicular monoamine transporter 2 (VMAT2), a mechanism that can reduce excessive dopaminergic signaling and alleviate chorea symptoms in HD patients.^(p119) Another neurodevelopmental disorder, FXS, is caused by a CGG repeat expansion in the fragile X messenger ribonucleoprotein 1 (FMR1) gene.^(p120) This expansion leads to epigenetic silencing of the FMR1 gene, resulting in the loss of the fragile X mental retardation protein (FMRP). FMRP is an important regulator of synaptic mRNA translation and neuronal plasticity.^(p121) To address this condition, Healx introduced two repurposed drug candidates, Sulindac (HLX-0201, 6) and Gaboxadol (HLX-0206, 7), which are currently in Phase II clinical studies.^{(p13),(p100),(p122)} These candidates were identified using AI/ML-driven omics-based drug-matching platforms that integrate genetic, molecular and clinical datasets to identify compounds capable of modulating disease-associated pathways. CCM represents a vascular neurological disorder. This condition results from the loss of function in any of the three CCM genes (CCM1, CCM2 or CCM3).^(p122) These three genes encode proteins that form the CCM signaling protein complex. These signaling proteins are essential for maintaining vascular integrity in endothelial cells. A mutation in the CCM gene disrupts the CCM complex protein in endothelial cells, ultimately leading to the formation of thin-walled, dilated blood vessels prone to leakage or hemorrhage.^(p123) Recursion has advanced a candidate drug, REC-994 (8), into Phase II clinical trials.^(p13) REC-994 (8) is a bioactive small molecule identified using Recursion’s ML-based discovery platform, which analyzes large-scale biological imaging data to uncover relations between chemical compounds and disease phenotypes.

Current challenges and benchmarking tools in artificial intelligence-driven drug discovery

Despite the growing adoption of AI in drug discovery and development, several challenges continue to limit its translational and regulatory impact, particularly for CNS therapeutics.^(p124) A major limitation is the insufficient availability of high-quality biological and clinical data, including useful negative data.^(p125) CNS disorders are biologically complex and heterogeneous, while clinical trial datasets are often small, incomplete and biased.^{(p126),(p127)} These limitations reduce the robustness and generalizability of AI models, making reliable prediction of therapeutic efficacy and safety difficult.^{(p127),(p128)}

Another significant challenge is the lack of interpretability of AI models. Many DL approaches operate as ‘black-box’ systems, providing predictions without clear mechanistic explanations.^(p129) This opacity hampers biological understanding and undermines confidence in AI-driven decisions, especially in CNS research, where mechanistic insights are crucial.^(p27) As a result, regulators such as the US Food and Drug Administration (FDA) remain cautious in relying on AI outputs, treating them primarily as supportive tools rather than definitive evidence in drug approval processes.^{(p130),(p131)}

Data transparency and validity, along with the FDA regulatory framework for AI, further complicate AI integration. Many models are trained on proprietary or inconsistently curated datasets, hindering reproducibility and independent validation.^(p132) Biases arising from data imbalance or poor annotation can lead to misleading predictions or new ‘garbage in, garbage out’ (GIGO) information available in public data domains, raising concerns about safety and reliability.^(p133) These biases are particularly concerning in CNS research, where clinical trial populations may underrepresent diverse ethnic and demographic groups, potentially limiting the external validity and equity of AI-driven therapeutic predictions. Such disparities raise ethical concerns regarding fairness, inclusivity and equitable access to AI-informed CNS therapies. In addition, the lack of standardized regulatory frameworks makes it difficult to govern adaptive AI systems, as their decision-making logic and performance may change over time.^{(p134),(p135)}

Alongside these methodological limitations, the rapid expansion of AI applications across scientific and industrial sectors has substantially increased the demand for high-performance computing infrastructure. In recent years, the cost of advanced graphical processing units (GPUs), specialized accelerators and large-scale computing clusters has risen markedly. Establishing and maintaining supercomputing facilities, therefore, require significant financial investment, technical expertise and long-term operational support. For smaller academic institutions, newly established laboratories and resource-limited settings, these infrastructural and economic constraints can represent a major barrier to adopting state-of-the-art AI methodologies. This disparity might contribute to unequal access to AI-driven research capabilities, potentially widening the gap between well-funded centers and emerging institutions.

Finally, although AI can accelerate target identification, molecular design and optimization, translating in silico predictions into in vivo and clinical success remains limited.^(p136) CNS drug development requires long-term safety and efficacy evaluation, which current AI models struggle to predict accurately owing to insufficient longitudinal data. Recent clinical failures, such as Exscientia’s DSP-1181 and Verge Genomics’ VRG50635, underscore that AI-accelerated molecular design cannot compensate for limitations in target novelty or underlying disease biology. DSP-1181 was discontinued after Phase I due to limited clinical promise because of an insufficiently differentiated mechanism that targets the well-established 5-HT_1A receptor and structural similarity to existing antipsychotic scaffolds.^(p137) Conversely, a no-go decision was made for the ALS candidate VRG50635 in late 2025 due to a lack of risk–benefit data to meet efficacy thresholds in Phase 1b trials (Verge drops sole clinical candidate in return to AI roots). Therefore, while AI is revolutionizing early-stage CNS drug discovery, its effective application depends on improved data quality, model interpretability and regulatory alignment, with continued human oversight remaining essential.

Concluding remarks and prospects

This review highlights how the incorporation of AI-driven methodologies within CADD and medicinal chemistry pipelines is reshaping drug discovery, particularly in the realm of CNS disorders. It is well known that traditional drug development processes often fail because neurological diseases are complex, challenging to study experimentally, costly to investigate and associated with very high failure rates in clinical testing. The convergence of AI-powered platforms with CADD now enables faster, more precise methods for identifying and developing drug candidates using a range of drug discovery tools. When combined with AI-accelerated technologies such as ML, DL and generative modeling, these approaches can analyze large biological datasets more efficiently and predict drug–receptor interactions with greater precision. Furthermore, AI-enabled tools also facilitate early ADMET evaluation, drug repurposing and patient-focused strategies. This helps develop new molecules with improved CNS properties, including greater BBB penetration, lower toxicity and better safety. Many AI/ML-focused emerging technologies, such as digital twins and biomedical knowledge graphs, are further improving clinical trial planning and patient selection. Despite these advances, challenges persist. AI models require high-quality data, and issues such as transparency, reproducibility, privacy and regulatory approval still limit the wide-spread use of AI. Continued efforts by scientific communities and global regulatory agencies are crucial to ensure the safe and reliable translation of AI discoveries into real therapies. Overall, the convergence of AI with drug discovery not only accelerates innovation but also promotes more ethical, efficient and data-driven research practices, with the potential to reduce experimental burden, lower failure rates and ultimately improve patient outcomes in complex neurological diseases.

Declaration of Generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used Grammarly and ChatGPT for language editing, including correcting spelling and grammar, as well as improving clarity and readability. In some instances, AI tools were used to assist with literature navigation; however, any information included in the manuscript was independently verified through primary sources and cross-checked for accuracy. No scientific content, data or interpretations were generated without author validation.

Data availability

No data were used for the research described in the article.