The use of artificial intelligence (AI) in neuropsychiatric drug discovery: current challenges and future directions

Abstract

Artificial intelligence (AI) has emerged as a powerful tool for solving real world problems across a wide range of industries and is increasingly being utilised by pharmaceutical companies to discover novel drug targets, biomarkers, and new drugs. Several AI-driven small molecules have entered clinical trials over the past few years, but their fate remains unknown. Currently, no commercially available compounds have been developed solely using AI approaches. In this perspective, we examine the current use of AI in drug discovery for neuropsychiatry. The pace of drug discovery in neuropsychiatric medicine has been generally sluggish, largely due to challenges such as poor pharmacological selectivity, the blood-brain barrier, and a limited understanding of disease mechanisms. AI may offer innovative solutions to these challenges. However, relative to fields such as oncology, the impact of AI on the discovery of neuropsychiatric drugs has been limited. Although novel AI-tools have been developed to overcome some of the challenges involved in neuropsychiatry drug discovery, their effectiveness has not been sufficiently evaluated. To date, innovative tools such as AlphaFold have been used to identify drug candidates for multiple neuropsychiatric conditions. AI-driven platforms have been used to study behavioural data from preclinical models to identify novel clinical candidates in clinical trials (e.g., ulotaront, phase III). It is anticipated that the availability of large-scale multi-omics data (‘big data’) will likely increase in the future, allowing us to gain a better understanding of gene-associated mechanisms in psychiatry. Using AI-based technologies such as AlphaFold, future pharmacological targets will be identified based on gene expression data, and large libraries of chemical compounds will be screened rapidly to identify novel drug candidates, resulting in shorter pre-clinical phase with lower costs.

Introduction

Research and development of new drugs in neuropsychiatry has been slower and less successful than in most other areas of medicine [1]. Historically, drug development for neuropsychiatric diseases has been associated with higher attrition rates, longer development times, and higher costs [2]. These hurdles have driven many pharmaceutical companies out of the field. As a result, psychiatry is one of the few clinical disciplines where the innovation index, defined as the number of mechanisms of action per approved drug, has declined since the turn of the millennium [3]. Many currently used psychiatric medications were discovered in the 1950s and 1960s, including antipsychotics, antidepressants, anxiolytics (benzodiazepines), and mood stabilisers (lithium) [4]. Rather than systematic, hypothesis driven investigations, drug development in neuropsychiatry has largely relied on serendipity [4]. As a result, many psychiatric treatments lack a clear biological basis and are frequently linked with side effects [5]. For example chlorpromazine was first synthesised in 1950 as an antihistamine, but was later found to possess antipsychotic and anxiolytic properties [4]. The synthesis of chlorpromazine heralded the beginning of modern-day psychopharmacology.

The specific dopamine hypothesis of schizophrenia is the result of a 1966 study which discovered that chlorpromazine counteracted amphetamines by blocking dopamine receptors [6]. Most subsequent treatments for schizophrenia have targeted dopamine (D2) receptors. However, dopamine targeting drugs are frequently associated with extra-pyramidal side effects and remain ineffective for a significant number of patients [5, 7]. Similarly, the monoamine hypothesis of depression was first put forward over 50 years ago after it was observed that patients treated for hypertension with reserpine developed depression [4]. The monoamine theory of major depressive disorder was supported by findings that tricyclic antidepressants and monoamine oxidase inhibitors (MAOIs) enhanced monoamine neurotransmission (serotonin, noradrenaline). However, many patients do not respond adequately to monoamine-based antidepressants [7].

In recent years, the rational design of neuropsychiatric drugs has seen some success. However, it remains a slow and failure-prone process. A notable recent example in the treatment of schizophrenia is the development of xanomeline, an agonist of muscarinic receptors (M1/M4) that was originally developed for Alzheimer’s disease with antipsychotic-like benefits [8]. Although early trials demonstrated some efficacy, the compound was discontinued due to intolerable side effects. Decades later, Karuna Therapeutics repurposed xanomeline by combining it with trospium chloride, a peripherally restricted muscarinic antagonist. In phase II and phase III clinical trials, KarXT significantly reduced schizophrenia symptoms (both positive and negative) compared to placebo [9]. In 2024, KarXT was approved by the US Food and Drug Administration (FDA) and is currently marketed as Cobenfy, for the treatment of schizophrenia [8]. As the first non-D2-targeting drug approved for schizophrenia in over 30 years, Cobenfy represents a major breakthrough in the field, providing new treatment options for schizophrenia. Ulotaront (SEP-363856), a TAAR 1 agonist, is another recently developed compound for the treatment of schizophrenia. Ulotaront was discovered with the help of an artificial intelligence (AI)-driven platform called SmartCube®. The drug candidate was granted a Breakthrough Therapy Designation by the US FDA in 2019 [10]. The compound is currently in phase III clinical trials for schizophrenia (NCT06894212), major depression (NCT05593029), and generalised anxiety disorder (NCT05729373).

AI is revolutionising many areas of medicine, and offers human equivalent or superhuman performance for specific clinical tasks [11]. AI methods also offer potential solutions for many of the wicked problems involved in neuropsychiatric drug discovery, although application of AI in this field is nascent.

In this review, we discuss the emerging role of AI in neuropsychiatric drug discovery. We also discuss various AI-enabled approaches that aim to overcome longstanding challenges in neuropsychiatry.

Artificial intelligence (AI) in preclinical drug discovery

Drug discovery has historically relied heavily on serendipity, with many important breakthroughs resulting from chance observations [12]. In the field of psychiatry, accidental discoveries have led to the development of prototype psychotropic drugs and modern pharmacological agents [13]. The modern-day process of drug discovery involves the identification of biologically active compounds that can be used to treat diseases. A typical drug development process begins with the identification of molecular targets, followed by the identification of potential drug candidates, and finally, the optimisation of lead compounds. After this, preclinical and clinical trials are conducted to obtain regulatory approval. The development of a new drug typically costs between $100 million and $2 billion, with the timeline stretching over 15 years [14, 15]. Drug discovery is not only a time-consuming process, but it is also highly risky and expensive with high failure rates [16].

Pharmaceutical companies are adopting AI-based approaches (Table 1) to speed up the drug discovery process and overcome the limitations of traditional approaches with long preclinical stages. Based on recent reports, the world’s top 20 pharmaceutical companies (e.g. AbbVie, Merck, GSK, and Johnson & Johnson) are utilising AI in their research and development efforts [17]. Moreover, a significant increase in the number of AI-based companies has also been noted in the past few years (Table 2, Supplementary Table S1) that are developing multiple small molecule candidates for a variety of disorders. The growth of such companies has been driven by two critical factors, significant financial investment and advances in computing infrastructure [18]. AI companies now have much easier access to ‘big data’ as well as other computing infrastructure like graphics processing units and cloud computing resources [18]. This has enabled AI companies to process various data forms such as texts, images and videos, significantly accelerating technological progress [11]. Moreover, the global market for AI in drug discovery has seen rapid growth. An analysis of 200 AI drug discovery companies revealed that private investors have contributed over $18 billion in funding over the past 10 years [19]. During this same period, these companies’ small molecule drug pipelines expanded at an annual growth rate of 27% [19].

Table 1.

Definitions of AI algorithms and their applications within drug discovery.

Term	Definition	Applications	AI Tools
Genetic algorithms	Genetic algorithms are defined as algorithms that mimic the process of natural selection to discover the optimal solution to a problem [142].	Molecular optimisation and virtual screening	GENERA, ligand GA, DGMM
Deep-q-networks	Deep-q-networks combine deep learning with q-learning to determine ideal actions in an environment and optimise cumulative rewards over time [143].	Molecular optimisation	MolDQN
Generative adversarial network (GAN)	A generative adversarial network (GAN) is an AI model that utilises two competing neural networks to generate data [11].	De novo drug design	Chemistry42, Banyan, DR-A
Recurrent neural network (RNN)	A recurrent neural network (RNN) is a deep learning model typically used for processing sequential or time-series data [22].	De novo drug design, natural language processing and molecular property prediction (e.g. toxicity and solubility)	Reinvent4, PharmaNet
Variational autoencoder (VAE)	A variational autoencoder (VAE) is a generative model that learns the underlying structure of a dataset and can generate outputs that look similar to the original data [22].	De novo drug design, chemical space mapping and molecular property prediction	NP-VAE, TGVAE, PCF-VAE
Convolutional Neural Network (CNN)	A convolutional neural network (CNN) is a deep learning model that utilises three-dimensional data for image classification and recognition [22].	Molecular property prediction and modelling	AtomNet, DeepDTA, DeepPurpose
Multi-task deep neural networks (MTDNN)	Multi-task deep neural networks (MTDNN) are a type of machine learning architecture that performs multiple tasks simultaneously [71].	Drug repurposing, protein-ligand binding affinity prediction, multi objective lead optimisation and de novo drug design	DeepDTAGen, GeneralizedDTA
Large language model (LLM)	Large language models (LLM) are AI systems that can understand and generate human language by processing vast amounts of text data [11].	Retrosynthetic prediction, target discovery, literature mining	Osmanthus, MegaMolBART
K-means	K-means is a machine learning algorithm capable of clustering unlabeled data [22].	Chemical space analysis, virtual screening, QSAR modelling	Scikit learn, KNIME
Principal Component Analysis (PCA)	Principal component analysis (PCA) is a machine learning technique designed to reduce the dimensionality of datasets [22].	Chemical space mapping, QSAR modelling	ChemGPS, ChemMaps
Linear discriminant analysis (LDA)	Linear discriminant analysis (LDA) is a machine learning technique that projects high dimensional data onto lower dimensional space for dimensionality reduction and classification [144].	Activity classification, ADMET profiling	Scikit learn, WEKA 3
Naive bayes	Naive bayes is a statistical classification algorithm that utilises Bayes' Theorem [145].	Activity classification, molecular property prediction and virtual screening	PASS, Pipeline Pilot
Support vector machines (SVM)	Support vector machines (SVM) are machine learning algorithms used for classification and regression tasks [22].	Activity classification, virtual screening and molecular property prediction	Sci-kit learn, LIBSVM
Random forests (RF)	Random forests (RF) are a type of machine learning algorithm that utilises decision trees to make a prediction [146].	Activity prediction, QSAR modelling and molecular property prediction	RF-Score, Sc-kit learn, DeepChem
Decision trees (DT)	Decision trees (DT) are machine learning algorithms that employ a hierarchical tree structure for regression and classification tasks [146].	Activity prediction, QSAR modelling and molecular property prediction	DTIGBDT, C5.0, WEKA 3
Linear regression	Linear regression is a supervised learning technique that models the relationships between two continuous variables with a straight line [147].	Activity prediction, QSAR modelling and molecular property prediction	modSAR, caret, sci-kit learn
Polynomial regression	Polynomial regression is a form of linear regression that models the relationship between two variables by expanding the input into polynomial terms [147].	Activity prediction, QSAR modelling and molecular property prediction	Caret, sci-kit learn, KNIME
Ridge regression	Ridge regression is an extension of linear regression that adds a penalty for large coefficients to reduce overfitting [147].	Activity prediction, QSAR modelling and molecular property prediction	Sci-kit learn, glmnet
Natural language processing	Natural language processing is a field of AI that focuses on enabling computers to understand, interpret, and generate human language [148].	Target discovery and drug repurposing	NovareAI, PandaOmics, Osmanthus
Graph deep learning	Graph deep learning is a subfield of deep learning that is designed to process data represented as graphs [11].	Target discovery, molecular property prediction, de novo drug design	PandaOmics, KarmaDock
Knowledge graph	A knowledge graph is a data framework used in machine learning which can organise information in a structured graph format and detail relationships between entities within a domain [149].	Target discovery, drug repurposing	NovareAI, TarKG, OREGANO

Table 2.

AI-derived drug candidates for neuropsychiatric indications in clinical trials and preclinical stage as of 2025.

Name	Company	Target Class	Drug Target/Mode of Action	Disease Indication	Preclinical/ Clinical Phase	Clinical Phase Outcome
Igalmi (Dexmedetomidine) (BXCL-501) [150–153]	BioXcel	GPCR	α2-adrenergic receptor	Acute treatment of agitation associated with schizophrenia or bipolar I or II	Phase IV	Successful
				Opioid use disorder	Phase I/II	Successful
				Alcohol use disorder with comorbid post-traumatic stress disorder	Phase I	Successful
				Acute stress disorder	Phase II	Ongoing
(SEP-363856) [154, 155]	Sumitomo Pharma / Otsuka Pharmaceuticals	GPCR	5-HT1A agonist and TAAR1 agonist	Schizophrenia	Phase III	Unsuccessful
				Generalised anxiety disorder	Phase II/III	Completed
				Major depressive disorder	Phase II	Completed
DSP-1181 [156]	Sumitomo Pharma / Exscientia	GPCR	5-HT1A agonist	Obsessive-compulsive disorder	Phase I	Successful (Discontinued)
DSP-0038 [157–159]	Sumitomo Pharma / Exscientia	GPCR	5-HT1A agonist and 5-HT2A antagonist	Alzheimer’s disease psychosis	Phase I	Ongoing
DSP-2342 [157, 160]	Sumitomo Pharma / Exscientia	GPCR	5-HT2A antagonist and 5-HT7 antagonist	Psychiatric disorders	Phase I	Ongoing
MSD-001 [161]	Mindstate Design Labs	GPCR	5-HT2A agonist	Mental disorders	Phase I	Successful
ALTO-100 [162]	Alto Neuroscience	n/a	Brain-derived neurotrophic factor modulator (precise target unknown)	Bipolar depression	Phase II	Ongoing
				Major depressive disorder	Phase II	Unsuccessful
				Post-traumatic stress disorder	Phase II	Ongoing
ALTO-300 [163]	Alto Neuroscience	GPCR	MT1 agonist, MT2 agonist and 5-HT2C antagonist	Major depressive disorder	Phase II	Ongoing
ALTO-207 [164]	Alto Neuroscience	GPCR	DRD2 agonist, DRD3 agonist, DRD4 agonist and 5-HT antagonist	Major depressive disorder	Phase II	Ongoing
ALTO-203 [165]	Alto Neuroscience	GPCR	HRH3 receptor agonist	Major depressive disorder	Phase II	Ongoing
ALTO-101 [166]	Alto Neuroscience	Non-kinase enzyme	PDE4 inhibitor	Schizophrenia	Phase II	Ongoing
ALTO-202 [167]	Alto Neuroscience	GPCR	NMDA receptor antagonist, NR2B NMDA antagonist	Major depressive disorder	Phase I	Ongoing
Bavisant (BEN-2001) [168]	BenevolentAI	GPCR	HRH3 antagonist	Excessive daytime sleepiness in patients with Parkinson’s disease	Phase II	Unsuccessful
EB-003 [169, 170]	Enveric Biosciences	GPCR	5-HT2A agonist, 5-HT1B agonist	Post-traumatic stress disorder, treatment-resistant depression, generalised anxiety disorder	Preclinical	n/a
EVM-401 Series [169, 171]	Enveric Biosciences	GPCR	5-HT2 receptor agonist	Opioid withdrawal, attention-deficit/hyperactivity disorder, anxiety, depression and substance use disorders	Preclinical	n/a
KCB-200 [172]	Kapoose Creek Bio /Adapsyn Biosciences	n/a	Undisclosed	Depressive disorders	Preclinical	n/a
CPI-CG-8 [173, 174]	Collaborations Pharmaceuticals	GPCR	5-HT2A agonist, 5-HT2C agonist	Opioid use disorder, fibromyalgia, rare epilepsies, obesity and depression	Preclinical	n/a
Compound 8 [71]	n/a	GPCR	DRD2 antagonist, 5-HT1A agonist and 5-HT2A agonist	Psychosis	Preclinical	n/a
FJRL-14, FJRL-20 [72]	n/a	GPCR	KOR antagonist	Opioid addiction	Preclinical	n/a
Compound 1, Compound 2 [73]	n/a	GPCR	TAAR5 antagonist	Anxiety and depression	Preclinical	n/a
Compound 65 [77]	n/a	GPCR	TAAR1 agonist	Psychosis	Preclinical	n/a
Z2504, Q2118 [76]	n/a	GPCR	5-HT2A agonist and antagonist, σ2 uncharacterised ligand	Unspecified neuropsychiatric disorders	Preclinical	n/a

The primary goal of AI is to develop computational systems capable of performing tasks typically associated with human intelligence such as learning, reasoning, pattern recognition, making predictions and problem solving [20]. Machine learning and deep learning are both types of AI that encompass a wide range of algorithms [21]. Machine learning algorithms are generally classified as supervised, unsupervised, or reinforcement learning (Fig. 1A, Table 1) [22]. They have various applications within the drug discovery pipeline [11]. For instance, several AI-based tools have been developed to address different aspects of preclinical drug discovery, including target identification, hit identification, lead optimisation, ADMET, and chemical synthesis planning (Fig. 2) [23–31]. Additionally, the use of multi-omics data, including genomic, transcriptomic, proteomic, and metabolomic information collected from patient-derived samples, is increasing in the context of pharmacogenomics and personalised medicine [32–35]. Combining this approach with AI has demonstrated the possibility of discovering biomarkers, diagnosing diseases, predicting drug sensitivity, and optimising treatment [36–39]. For instance, AI algorithms such as natural language processing techniques (using machine learning to analyse human language) and graph deep learning (which employ graphs as inputs for AI models) have been used to synthesise biomedical literature and multi-omics data to uncover new targets [40–43]. In one study, 8 unreported genes linked to amyotrophic lateral sclerosis (ALS) were discovered using a target discovery program called PandaOmics [44]. The program uncovered these novel targets by analysing gene expression profiles from healthy and ALS-affected brain tissue. The study subsequently validated these genes in a fruit fly model of ALS. It was discovered that 8 genes significantly reduced eye neurodegeneration. The growth of ultra-large chemical libraries with billions of compounds has rendered traditional physics-based virtual screening impractical [45]. Concomitantly, various machine learning methods have emerged to address this challenge [46–50]. An AI platform called RosettaVS was used to screen multi-billion compound libraries against 2 unrelated targets: a ubiquitin ligase target (KLHDC2) and the human voltage-gated sodium channel (NaV1.7) [51]. The AI-driven screening discovered 7 hits for KLHDC2 and 4 hits for NaV1.7 with low micromolar binding affinities. Deep learning-based methods for de novo compound design have emerged to automatically generate compounds with specific molecular properties [52–55]. Examples of deep learning models used for de novo design include generative adversarial networks, recurrent neural networks, and variational autoencoders (Table 1, Fig. 1A, Figs. 3A, 3B). A variational autoencoder was used to design potent kinase discoidin domain receptor 1 inhibitors in 21 days [56]. The AI model was pretrained on a filtered database of commercially available compounds and refined using reinforcement learning (which is where a model learns by maximising a reward). The AI model proposed 30,000 compounds, of which 6 were subsequently synthesised and tested. The best compound had an IC50 value close to 10 nM [56]. Likewise, a generative deep learning approach was employed for de novo antibiotic design, resulting in the development of compounds with selective antibacterial activity. These compounds displayed novel mechanisms of action in vivo against multiple multidrug-resistant pathogens such as Neisseria gonorrhoeae and Staphylococcus aureus [57]. Planning the synthesis of compounds is an important phase in preclinical drug discovery. In one study, researchers trained a natural language processing model for synthesis planning [58]. The approach demonstrated performance comparable to that of traditional synthesis planning algorithms based on expert rules. However, machine learning in synthesis planning has yet to have more than a limited impact in the field of neuropsychiatry

Fig. 1. AI algorithms and AI-enabled small molecules in clinical trials across drug target classes and disease areas.

A Classification of machine learning algorithms. B AI-driven small molecules in clinical trials from 2015 – 2025. For drug candidates with multiple indications, only advanced clinical phases were considered. C AI-enabled discovery of small molecules currently in clinical trials based on different target classes.

Fig. 2. Integration of AI in neuropsychiatric drug development.

Neuropsychiatric drug development pipeline highlighting key stages, critical decision checkpoints, and the integration of AI programs at each phase. Created with BioRender.com.

Fig. 3. AI-driven approaches in drug development within neuropsychiatry.

A Automated deep learning for dual-targeting antipsychotic compounds with efficacy in preclinical models [71]. B Generative deep-learning framework for de novo design and experimental validation of novel κ-opioid receptor antagonists for opioid addiction treatment [72]. C Comparison of AlphaFold2-predicted and Cryo-EM structures in prospective docking campaigns reveals similar hit rate and potency [76]. D Mindstate Design Labs’ AI-driven Osmanthus platform integrates human reports and biochemical data to map neurotransmitter receptor combinations to specific emotional states. E Alto Neuroscience uses AI-derived biomarkers from neuroimaging data to stratify patients in clinical trials. Created with BioRender.com.

The state of the art in AI-enabled drug candidates in clinical trials

One of the most recent examples of an AI discovered small molecule is Rentosertib (ISM001-055) [59]. This compound was developed by Insilico Medicine to treat idiopathic pulmonary fibrosis (IPF), a chronic and progressive lung disease. In a large-scale multi-omics analysis, TRAF2- and NCK-interacting kinase (TNIK) were identified as a novel target implicated in IPF. Subsequently, a deep learning-based generative chemistry model was employed to design compounds with high binding affinity to TNIK. Rentosertib is reported to be the first generative AI-discovered drug candidate to enter clinical trials (NCT05975983) [60]. The compound reached investigational new drug approval in 18 months, which is a markedly shorter timeline than traditional drug discovery (4–5 years) approaches [60, 61]. The same company (Insilico Medicine) also developed ISM8207 to treat B-cell lymphoma and solid tumours [62]. ISM8207 is an inhibitor of glutaminyl-peptide cyclotransferase and was designed using an AI-powered small molecule generation platform called Chemistry42. The drug candidate progressed from target discovery to preclinical nomination within 9 months (NCT06445517). As a result of these successful examples, more pharmaceutical companies are adopting AI to expedite their drug discovery timelines [63]. The number of AI-enabled drug candidates entering clinical trials has been increasing for the past decade [17]. As of 2025, 68 AI-enabled small molecules have entered clinical trials (Fig. 1B, Table 2, Supplementary Table S1). Most AI-enabled small molecules are in phase I and II trials. This trend is expected as using AI in drug development is still a relatively recent advancement [17]. AI-enabled drug candidates span a broad range of therapeutic areas and target classes, although drug candidates for oncology constitute half of all AI-generated compounds currently in phase I (Figs. 1B, 1C). The success of these approaches is largely due to the availability of vast, high-quality, and labelled datasets to train and validate AI models. Multi-omics resources such as The Cancer Genome Atlas (TCGA) [64], Catalogue of Somatic Mutations in Cancer (COSMIC) [65], and the Cancer Cell Line Encyclopedia (CCLE) [66] are examples of such datasets. Oncology is further supported by an abundance of translatable preclinical models including three-dimensional organoid cultures, patient-derived xenografts, and cancer cell lines [67]. These models closely mimic human biology and can enable the thorough validation of AI-developed drugs before they enter the clinical phase. In contrast, the impact of AI on the field of neuropsychiatry appears limited. There have only been a handful of compounds for neuropsychiatry that have entered clinical trials (Fig. 1B, Table 2). A significant part of this is due to the challenges faced by the field of neuropsychiatry, including but not limited to poorly defined disease mechanisms, clinical heterogeneity, and a lack of predictive preclinical models [13].

The promise of ai in neuropsychiatric drug discovery

AI is being increasingly implemented in the drug discovery process for neuropsychiatric diseases. In 2022, the FDA approved Igalmi (dexmedetomidine), an alpha-2 agonist, for the treatment of acute agitation associated with schizophrenia and bipolar disorder [68]. Igalmi was originally developed as a non-opioid analgesic but was repurposed using an AI platform called Novare AI by BioXcel Therapeutics [69]. NovareAI is a program that can mine scientific literature and connect information about drugs, neural pathways, and diseases by using knowledge graphs [69]. A knowledge graph is a network of entities and relationships used to organise information, and is typically built using natural language processing techniques [11]. A notable collaboration between Exscientia (an AI native biotech) and Sumitomo Pharma led to the discovery of DSP-0038 in 2021 [70]. DSP-0038 is a 5-hydroxytryptamine receptor 1A (5-HT1A) agonist and 5-hydroxytryptamine receptor 2A (5-HT2A) antagonist designed for the treatment of Alzheimer’s disease psychosis. This drug candidate was designed with Exscientia’s AI technology to selectively engage two receptors while reducing activity at off-target sites such as dopamine receptor D2 (DRD2). The collaboration between Exscientia and Sumitomo pharma have led to a handful of other AI-designed drug candidates entering phase I clinical trials including: DSP-2342, and DSP-1181. These drug candidates are designed to treat a broad range of neuropsychiatric indications such as psychosis, anxiety and OCD.

There are several preclinical drug candidates using AI-enabled approaches to develop novel therapeutics for neuropsychiatric disorders. In one study, a multi-target compound was designed with antagonist and agonist activity at the DRD2, 5-HT2A, and 5-HT1A receptors respectively [71]. The researchers trained a recurrent neural network, a type of deep learning algorithm effective at processing sequential data, on 3.1 million drug-like molecules from the ZINC database and 10,286 molecules that had reported affinity for the target receptors (Fig. 3A). A subsequent multi-task deep neural network was trained to evaluate the generated compounds and improve the recurrent neural network. A promising compound was synthesised and administered orally using a phencyclidine (PCP)-induced hyperactivity mouse model. The compound produced an 80% reduction in PCP induced hyperactivity at 3 mg/kg, which is superior to some atypical antipsychotics [71]. In another study, researchers used deep generative chemistry to design kappa opioid receptor (KOR) antagonists for the treatment of opioid addiction [72]. The researchers trained a conditional generative autoencoder, a deep learning based generative model, on 4.6 million molecules from the ZINC database and 2056 compounds with reported KOR antagonism from the ChEMBL database (Fig. 3B). The generated compounds were refined using a reinforcement learning strategy that learned by maximising a reward such as KOR antagonism. The researchers applied rewards such as blood–brain barrier permeability, ease of synthesis, and novelty. They discovered two compounds with Ki values of 6.46 µM and 7.59 µM, respectively [72]. AI has also been employed in structure based virtual screening [11]. One study utilised AtomNet®, a deep convolutional neural network developed by the company Atomwise [73]. This AI screening method was used to identify candidate antagonists for trace amine-associated receptor 5 (TAAR5), a G protein-coupled receptor implicated in mood regulation. The study performed a virtual screen of a TAAR5 homology structure against 2 million drug-like small molecules [73]. Two molecules were characterised by TAAR5 antagonist activity and their IC50 values were 2.8 μM and 1.1 μM. The researchers subsequently evaluated the antagonistic activity of these hits by measuring the inhibition of ERK and CREB phosphorylation. The results showed that both compounds, at 10 µM, effectively blocked phosphorylation of ERK and CREB [73].

The use of alphafold in drug discovery

In 2020, AlphaFold won the CASP14 competition for its accurate prediction of three-dimensional protein structures from amino acid sequences [20]. AlphaFold is built on deep neural networks which are trained on 170,000 protein structures from the Protein Data Bank [74]. In 2024 the developers of AlphaFold were awarded the Nobel Prize for Chemistry [74, 75]. This technology, developed by Google DeepMind and Isomorphic Labs, utilises machine learning to predict a protein’s three-dimensional structure based on its amino acid sequences. The availability of more than 200 million protein structures in AlphaFold, as well as the ability to predict new structures rapidly, make this a powerful tool for understanding therapeutic targets [74]. In one study, researchers used experimentally resolved and high-confidence AlphaFold structures of the σ2 and 5-HT2A receptors to prospectively dock more than a billion molecules from the ZINC database [76]. The study discovered that prospective docking using the AlphaFold models was as effective as docking against the experimental structures. For the σ2 receptor, 55% of docking hits from the AlphaFold model were active at 1 µM, compared with 51% for the crystal structure (Fig. 3C) [76]. For the 5-HT2A receptor, 26% of compounds from the AlphaFold model bound at 10 µM, compared to 23% for the cryo-EM structure. The affinity comparisons between both receptors showed similar trends. The top σ2 hits had Ki values between 1.6 and 84 nM which was comparable to the crystal structure results. For the 5-HT2A hits, AlphaFold-derived hits were generally more potent and selective [76]. In another study, researchers docked 16 million compounds from the ZINC database to the AlphaFold-predicted and homology modelled trace amine-associated receptor 1 (TAAR1); a GPCR implicated in neuropsychiatric diseases [77]. The AlphaFold screen discovered 30 hits and 18 were confirmed to be TAAR1 agonists. In the homology model screen, 32 hits were discovered and 7 were confirmed to be TAAR1 agonists. Some promising compounds were characterised in vitro, and exhibited a pEC₅₀ from 4.3–7.6. The researchers also conducted TAAR1 knockout studies in vivo. Experiments were conducted to evaluate core body temperature, pre-pulse inhibition, and locomotor activity. A TAAR1 agonist was identified with a promising selectivity profile and drug-like properties and exhibited physiological and antipsychotic-like effects in wild-type but not in TAAR1 knockout mice [77].

Designing the next generation of psychedelics with AI

Psychedelics function via the activation of 5-HT2A serotonin receptors and encompass natural and synthetic compounds such as psilocybin, mescaline, and various phenethylamines [78]. There has been a resurgence of scientific and clinical interest in psychedelics in recent years [78]. Approximately 60 new companies have been established to explore opportunities for psychedelics in treating neuropsychiatric diseases [79]. This interest has been driven by a growing body of clinical evidence, including several phase II trials demonstrating that psilocybin can produce sustained reductions in symptoms of major depression [80]. AI is emerging as a promising tool to accelerate the discovery and development of psychedelic compounds for neuropsychiatric indications. A handful of companies are actively exploring AI’s potential to tackle the inherent challenges associated with psychoactive substances. Traditional psychedelics often elicit variable therapeutic outcomes which pose significant hurdles for clinical translation [79]. In response, AI-driven companies such as Mindstate Design Labs are attempting to rationally design compounds that can induce consistent and targeted psychedelic experiences to treat neuropsychiatric conditions [81]. The company employs natural language processing techniques and large language models, which are deep learning algorithms designed to understand and generate human-like text. They have used these algorithms to analyse approximately 70,000 ‘trip’ reports sourced from drug forums, clinical trial data, and online communities (Fig. 3D) [81]. These reports are integrated with pharmacological datasets to model relationships between drug structure, receptor activity, and subjective outcomes. Currently, Mindstate’s lead candidate, MSD-001, a non-selective serotonin receptor agonist, is undergoing phase I clinical trials for mental disorders (NCT06702332). Companies such as Enveric Biosciences have used AI-driven approaches to design and optimise neuroplastogens that do not possess hallucinogenic properties [82]. Neuroplastogens are small molecules that can induce sustained changes in neuronal structure and function, and virtually all psychedelics can trigger this process [83]. Enveric has synthesised over 1000 psychedelic inspired therapeutic candidates for a range of neuropsychiatric conditions such as anxiety, depression, ADHD, and substance use disorders [82].

The challenges of drug discovery in neuropsychiatry

As noted earlier, the development of new drugs in psychiatry has faced higher attrition rates, longer development times and higher costs than in other clinical areas [3]. This situation has occurred despite significant advances in neuroscience and billions of dollars of research expenditure in industry and academia [84]. The fundamental challenges in psychiatric drug discovery include: absence of causal mechanism for most psychiatric disorders, poor pharmacological selectivity, low translational relevance of preclinical animal models, significant clinical heterogeneity, psychosocial context and the presence of the blood brain barrier (BBB) (Fig. 4) [3, 85, 86]^. These challenges manifest in several ways. Many psychiatric drugs exhibit poor pharmacological selectivity and thus can produce significant off-target side effects [87]. The BBB is a highly selective and dynamic interface that protects the brain from potentially harmful substances [86, 88]. The BBB significantly limits the ability of many therapeutic compounds to reach their intended targets in the brain, often reducing efficacy and increasing the risk of peripheral side effects [88]. Another limitation is the low translational relevance of preclinical research [89]. Psychiatric diagnoses are based on complex human experiences. Subjective emotional states, cognitive assessments, and patient-reported symptoms are inherently difficult to model in animals [89]. For example, the forced swim test is often used as a proxy for behavioural despair in depression, yet compounds that show efficacy in animal models have been unsuccessful in humantrials. There are no established biomarkers within the field of neuropsychiatry. In general, biomarkers can be classified into several different categories such as diagnostic, prognostic, and disease monitoring [90]. A diagnostic biomarker is used to confirm the presence of a disease or to subtype patients, whilst predictive biomarkers help predict a patient’s response to a specific treatment and support personalised therapeutic strategies. Prognostic biomarkers provide a prediction of the course or progression of a disorder independent of therapeutic interventions and response biomarkers are used to indicate whether a biological response has occurred after treatment [90]. Disease monitoring biomarkers allow for monitoring disease status over time and safety biomarkers are used to predict adverse or toxic drug effects. Moreover, susceptibility/risk biomarkers are used to indicate whether a patient has an elevated likelihood of developing a disease [90]. In the last 50 years, the field of psychiatry has devoted a great deal of resources to the discovery of biological markers [91]. Yet none of these markers have successfully passed the necessary validation stages to be available for routine clinical use. Numerous biological markers have been proposed for psychiatric disorders: including genetic, epigenetic, proteomic, metabolomic, blood-based, and neuroimaging markers [91]. The lack of biomarkers in psychiatry is in sharp contrast to other areas of medicine which have an abundance of precise biomarkers that are used to support treatment strategies [92]. For instance, in oncology, the presence of estrogen receptor (ER), progesterone receptor (PR) and human epidermal growth factor (HER2), are used as markers to stratify patients with breast cancer [93]. Hormones appear to drive the growth of cancer in patients with ER/PR, which are often treated with hormone therapy, whereas HER2 positive patients are treated with targeted therapies. Likewise, tumour size (indicating extent and staging) and composition (e.g. genetics, molecular markers, and microenvironment) are fundamental biomarkers for tailoring therapeutic strategies for solid tumours toward personalised, precision medicine [93]. As another example, in neurodegenerative diseases, the Lumipulse G blood test is a diagnostic tool which can detect the presence of amyloid plaques and thus identify patients with Alzheimer’s disease [94]. Moreover, there are several neuroimaging techniques (PET, MRI, fMRI) which can be used for early diagnosis, subtype differentiation and treatment monitoring in neurodegenerative disorders. Although these tools exist in psychiatry, they are much less validated [95]. According to many studies, several fundamental obstacles have hindered the development of biomarkers in psychiatry [90–92, 96]. One of the primary limitations of the existing literature on psychiatric biomarker development is the excessive reliance on case-control studies that compare the behavioural and biological profiles of psychiatric patients and healthy controls [91]. While this approach may be useful for uncovering novel insights into neuropsychiatric disease pathology, it does not provide clinically actionable biomarkers. Another fundamental barrier to biomarker development in psychiatry is the inherent heterogeneity of neuropsychiatric diseases. For instance, there are approximately 1500 combinations of symptoms that can result in a diagnosis of MDD [91]. Thus, the patient samples used for biomarker discovery are unlikely to be representative of the entire population of patients. Furthermore, brain tissue is not available for routine sampling, and although cerebrospinal fluid can be collected, the procedure is invasive and unsuitable for widespread clinical application [92]. Lastly, our limited understanding of the pathogenesis of most neuropsychiatric illnesses further compounds these hurdles.

Fig. 4. AI approaches to address drug discovery challenges in neuropsychiatry.

Each panel display known challenges faced in neuropsychiatric drug discovery. The coloured boxes provide AI solutions to those challenges. There was no AI programs found for the patient and disease heterogeneity challenge. Created with BioRender.com.

Could AI address the challenges associated with drug discovery in neuropsychiatry?

AI is opening new avenues to address many of the challenges that have historically slowed drug discovery in neuropsychiatry. One of the major challenges of developing drugs for neuropsychiatry is the limited objective metrics available for assessing drug efficacy. A majority of neuropsychiatric tests are behavioural in nature and are based on subjective metrics and clinician ratings, which are not always accurate [13, 97]. However, AI holds the potential to overcome this barrier. In one study, researchers developed a convolutional neural network-based AI model (AttnKare-D) that can process behavioural data through virtual reality to diagnose ADHD in children [98]. AttnKare-D demonstrated diagnostic performance with an AUC of 0.893 when compared to diagnoses made by clinicians [98]. In another study, researchers developed an AI model (Whisper M) which can diagnose neuropsychiatric diseases based on speech analysis [99]. Whisper M was trained on the speech of 865 healthy adults which were questioned on their perceived mental states. The AI model demonstrated moderate-good detection performance for depression (AUC = 0.76), anxiety (AUC = 0.77) and insomnia (AUC = 0.73) [99]. Another promising application of AI in neuropsychiatric drug discovery is targeted polypharmacology, which involves the rational design of drugs that can modulate multiple targets [100, 101]. Targeted polypharmacology is valuable because neuropsychiatric diseases are complex and involve the interplay of multiple molecular pathways [1]. Modulating several synergistic targets at once can achieve enhanced therapeutic effects [101]. The pharmaceutical industry is embracing AI-enabled targeted polypharmacology. Harmonic Discovery has developed a platform called Banyan that uses deep generative chemistry to design compounds that selectively engage intended targets while reducing off-target interactions [102]. Cyclica, which was recently acquired by Recursion, has developed MatchMaker, an AI-enabled deep learning engine that facilitates proteome-wide evaluation to determine complex drug polypharmacology [103]. The MatchMaker program is used to drive the company’s Ligand Design platform which can design molecules with specific polypharmacological properties. Machine learning algorithms can also be applied to predict whether compounds are permeable to the BBB [104]. Common machine learning methods for BBB prediction include supervised learning algorithms such as support vector machines, naive bayes, and random forests, as well as deep learning techniques [104]. One study used MegaMolBART, a language model that can understand SMILES strings, to generate numerical representations of molecules [105]. These representations were then fed into an XGBoost classifier, which is a type of supervised learning classification algorithm, to predict BBB permeability, achieving an AUC of 0.88. The study also validated the results with 3-dimensional BBB spheroid models to confirm the AI model’s predictions. AI also has the capacity to address the challenges presented by preclinical models. A successful example of AI implementation in preclinical modelling comes from PsychoGenics in the development of the drug candidate ulotaront. PsychoGenics built a proprietary AI platform known as SmartCube®, designed to enhance their phenotypic screening process. Rather than relying on subjective markers, SmartCube® was trained on high-resolution video data that captures the nuanced behaviours of mice in response to behavioural challenges after treatment with well-characterised drugs [106]. Over time, the system learnt to recognise the subtle movement and activity patterns associated with effective antipsychotic therapies. When a novel compound is tested, SmartCube® analyses whether the behavioural signature of the mouse resembles that of a known antipsychotic [106].

AI in precision psychiatry

Precision psychiatry is an emerging paradigm that aims to deliver individualised treatment to psychiatric patients [107]. A number of precision psychiatry initiatives have been established such as the Innovative Medicine Initiative PRISM project [108], as well as efforts by big pharmaceutical companies such as Johnson & Johnson [109] and Boehringer Ingelheim [110]. The integration of precision psychiatry with AI can address significant challenges in neuropsychiatric drug discovery. There are a growing number of collaborative research initiatives aimed at integrating various omics datasets with AI to improve patient stratification, discover objective biomarkers and predict treatment response [111]. For instance, The Optimise and Predict Antidepressant Efficacy (OPADE) project is a European Union funded initiative to develop a predictive machine learning model for major depressive disorders [112]. The goal of OPADE AI is to integrate multi-omics data (such as genetic, epigenetic and transcriptomic data) to discover biomarkers, stratify patients and improve diagnostic accuracy in patients with major depression. Nevertheless, the OPADE research endeavour is still ongoing and has not yet generated any tangible results [112]. Moreover, AI models are increasingly being incorporated with multi-omics datasets to unravel the complexity of neuropsychiatric diseases [111]. In one study, researchers incorporated proteomic, metabolomic and post translational modification data with numerous machine learning models to discover two key molecules (phosphatidylcholine and sphingomyelin) that were correlated with better treatment response in patients with schizophrenia [113]. Machine learning has also been applied to process large neuroimaging, digital and clinical data to identify biomarkers [114]. In one study, four subtypes of depression that correlated with unique symptom profiles were discovered. Functional MRI data from patients with depression (n = 1188) were processed using a hierarchical clustering algorithm, which is an unsupervised learning algorithm that groups data into a tree of nested clusters [115]. In a separate study, the extent to which military veterans with PTSD would respond to psychotherapy was evaluated [116]. Resting state electroencephalogram (EEG) data was used to train a relevance vector machine, which is type of regression model. The study identified a moderate correlation (r = 0.53) between the machine learning model’s predictions and the clinician rated scores. Biotech companies have applied AI and precision psychiatry approaches to improve clinical trials. Alto Neuroscience is using machine learning to identify biomarkers that can group patients in clinical trials based on how likely they are to respond to treatment (Fig. 3E) [117, 118]. The company uses a machine learning regression algorithm on resting-state EEG data to identify a biomarker associated with favourable responses to antidepressant treatment [117]. Using this approach, the company has developed ALTO-300, a phase II candidate for major depressive disorder (NCT05922878) [117]. Support vector machines, supervised learning algorithms that can be used for classification tasks, have been used to analyse cognitive test results and find objective biomarkers that can predict how well patients respond to antidepressants [118]. This approach was used in developing ALTO-100, a drug currently being tested in a phase II clinical trials for bipolar depression (NCT06656416) [118]. Neumora therapeutics have used a proprietary AI platform, Data Biopsy Signatures™, to identify distinct disease subtypes for neuropsychiatric diseases [119]. The platform uses deep learning to process multi-modal data sets such as genomic, neuroimaging, and clinical records. The company is currently developing Navacaprant, a kappa opioid receptor antagonist in phase III clinical trials for major depressive disorder (NCT06058039).

Regulatory and ethical dilemmas with AI systems

The FDA’s vision of using AI, human cell models and computational models as the new standard is evident in its plan to exclude animal studies from preclinical safety and toxicity testing within the next three to five years [120, 121]. It is estimated that using AI and reducing the use of animal testing could reduce timelines and costs significantly in the near future [122]. Currently, government agencies are developing policies to ensure the safety, efficacy, and reliability of AI-enabled tools [123]. For instance, the FDA published a draft guidance in 2025 which provides recommendations on the use of AI to produce information intended to support regulatory decision-making regarding safety, effectiveness, or quality for drugs [124]. In addition, the US Centre for Drug Evaluation and Research (CDER) established an AI council in 2024 to ‘provide oversight, coordination, and consolidation of its activities around AI use’ [125]. Despite these efforts, current regulations are not adequate to address ethical concerns related to AI governance, including ethical use of AI systems, trust among stakeholders, and continuous monitoring and evaluation to ensure safety and compliance [123].

The use of AI in the clinical setting will play an increasingly important role when it comes to taking informed decisions, providing treatment plans, and progressing clinical trials during the drug development process [126]. Although AI-based approaches can provide benefits in terms of accuracy, speed, and analyses of big data, as well as making better decisions than humans, there are also potential threats when using the technology in a healthcare or drug development setting. It is well known that several AI methods are based on black box algorithms, which can lead to a lack of trust, bias and security flaws [127]. Large-scale, high-quality data is essential for the development of AI models that are capable of making meaningful and reliable predictions. An inaccurate decision can have a negative impact on the masses and may also adversely affect data protection, safety, and transparency. An example of this is IBM Watson Health’s cancer AI algorithm [128]. The algorithm was based on a small sample of synthetic, non-real cases with very limited input (real data) from oncologists around the world and was used by hundreds of hospitals worldwide for recommending treatments to cancer patients. A number of the actual recommendations for treatment included in the output were found to be incorrect, for example, suggesting the use of bevacizumab when a patient has severe bleeding, which represents a contraindication and a black box warning for the drug [128]. Similarly, a flawed algorithm can cause serious harm to patients, which demonstrates the scope of medical malpractice involving iatrogenic risk. Consequently, when an AI algorithm is released for use in clinical practice or in drug discovery-based approaches, systematic debugging, auditing, extensive simulations, and validation are required, as well as prospective assessment.

Challenges and future directions

AI applications in drug discovery have generated excitement and optimism due to its potential to automate workflows, accelerate timelines, reduce costs, and enable de novo drug design. The current landscape, however, is possibly over-hyped as no approved drugs have yet been developed solely through AI-driven platforms, and many AI-discovered targets and molecules are still under scrutiny. Although billions of dollars have been invested across the pharmaceutical industry, only a small number of AI-generated compounds have progressed to clinical trials. AI should not be viewed as a replacement for conventional drug discovery methods, but rather as a complementary tool within a highly interdisciplinary research process. The impact of AI in neuropsychiatric drug discovery has been limited compared to fields like oncology. This could be due to the unique challenges in neuropsychiatry such as a limited understanding of neuropsychiatric diseases, low translational value of preclinical models, lack of biomarkers and more. Despite this, the application of AI in precision psychiatry to discover biomarkers and subtype patients in clinical trials has shown promise.

AI algorithms often operate as ‘black boxes’, implying that their processes are generally beyond the interpretation of humans. There is, however, still a concern regarding the reproducibility, transparency, and transferability of AI models [11, 129]. A primary bottleneck in the development of AI models is the quality of the data. The adage “garbage in, garbage out” is aptly applied here, AI models are only as good as the data they are trained on [130, 131]. Currently, datasets are plagued by noise, inconsistencies, ambiguity, and are often derived from various sources with varying standards. AI models require large amounts of high-quality training data to make accurate predictions [131]. For refining predictions and driving meaningful outcomes in complex disease areas such as neuropsychiatry, where biomarker scarcity and disease heterogeneity pose significant challenges, high-quality, standardised datasets generated from reproducible, biologically relevant platforms (e.g., iPSC-derived organoids) will be crucial [131]. Using a multi-omics approach has been successful in various areas [40, 41, 132]. Particularly with the emergence of high-throughput sequencing, generating large-scale omics data has been made easier. Numerous research consortia have been established to collate multi omics data from psychiatric patients such as PsychENCODE [133] and the GTEx Project [134].

Quantum computing is a technology that uses the principles of quantum mechanics such as entanglement and superposition to process information [135]. It is proposed that quantum computing can solve complex drug discovery problems that are challenging using current supercomputers [136]. The discovery of KRAS inhibitors using quantum computation-enhanced algorithms shows great promise for future applications in the field of drug discovery, including neuropsychiatry. Quantum computing and quantum machine learning represents a promising frontier within the intersection of AI and drug discovery [137]. Quantum computation is better suited to model the advanced complexity of an interaction with higher precision as it can better simulate molecular behaviour at the atomic level [138, 139]. This opens doors for accurately predicting drug–target binding affinities, reaction mechanisms, and pharmacokinetic attributes. The development of quantum-based machine learning approaches has led to the identification of small molecules that exhibit improved drug-like properties, with some of these approaches proving to be more effective than classical ML approaches [140, 141]. Quantum machine learning represents a promising frontier in drug discovery although it is still in its early stages [138]. As quantum hardware continues to develop, it’s expected to play a more critical role in AI and drug discovery.

The application of AI for protein structure prediction, virtual screening, retrosynthesis planning, and ADME propertyprediction has demonstrated significant progress. Some predicted structures, particularly in dynamic or flexible proteins, such as GPCRs in neuropsychiatry, may not be representative of true active conformations, limiting their immediate utility in the discovery of ligands, requiring further verification and refinement.

As neuropsychiatric drug discovery is intrinsically complex, iterative, and unpredictable, AI alone is not a ‘silver bullet’, but when used effectively can be an important tool in the quest to develop faster, more effective therapies at a lower cost.

Author contributions

PCN conceptualised and wrote the first draft of the manuscript along with SJ. TB and LJP contributed to writing the manuscript.

Funding

This work received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Competing interests

The authors have no conflicts of interest to declare.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41398-026-04153-9.