Applications of artificial intelligence in drug discovery for neurological diseases

Abstract

Neurological disease encompasses over 1000 disorders, exacts a massive human health and financial toll as well as being a story of extremes. At one end are diseases that are complex and heterogeneous affecting millions, while at the other there are monogenic and rare diseases, with a handful of individuals. What are absent are drugs that can treat or cure the disease. Discovering these is challenging, held back by extreme costs to develop them or in some cases by the limited understanding of the diseases. After decades of drug discovery research there is now considerable data available which can be used to help develop novel compounds more strategically. This includes high throughput screening data with targets, crystal structures of proteins implicated in neurological diseases and adjacent data such as properties of molecules like blood brain barrier permeability as well as an array of in vitro and in vivo toxicity endpoints valuable for any drug targeting the central nervous system. While computational tools have been developing and applied to neurological diseases for decades, we are now in the age of machine learning and artificial intelligence (AI). This promises the potential to expedite the identification and discovery of new molecules. Whether by using individual computational techniques or complex end-to-end approaches, scientists can narrow the molecules they make and test as well as study more targets or diseases which might have been out of reach previously. This review highlights the many different applications of AI potentially enabling new discoveries and treatments for neurological diseases.

Graphical abstract

Introduction

Over the last 30 years, we have seen big pharmaceutical companies go through cycles in which they were productive in discovering new classes of drugs targeting the brain (1990s), and then not long after a period in which productivity declined (2000s) and many exited neurological disease research. Presently we are seeing a renewed interest. Why is this? Have there been any new breakthroughs in the science, are they chasing new financial opportunities or is it a symptom of “what is old is new again”? With each passing decade developing new pharmaceuticals has become increasingly expensive with staggering numbers in recent years (from $1bn [[1], [2], [3]] to ​> ​$4bn [4]). This cost can only be partially attributed to the high attrition rate with only a 14 ​% success rate of moving compounds from the clinic to the market between 2006 and 2022 [5]. Of particular note central nervous system (CNS) drugs suffer from the highest attrition rate, with only ∼7 ​% ultimately reaching the marketplace after an average of 12.6 years in development [6]. Developing drugs that target the brain for neurological indications is therefore certainly a difficult challenge and not one for the faint of heart. One example of a complex neurodegenerative disease is Alzheimer's disease (AD), which affects 10's of millions of individuals, with the underlying mechanisms of its pathophysiology still unclear. There are also rare diseases that affect a handful of patients caused by a mutation in a single (likely essential) gene which may have a devastating effect on development and shorten the life of the patient (e.g. Batten disease CLN1 as an example [7]). In both examples, to get to a reliable treatment, one must get the drug into the brain, then deliver the small molecule or biologic to the desired target receptor or organelle in the right region/s. This is akin to addressing a letter with a full address and zip code to ensure the exact delivery to the desired location. Optimal drug discovery for both common and rare neurological diseases would involve the ability to predict key factors across the development pipeline, from designing molecules with optimal target engagement and physicochemical properties for oral bioavailability and blood–brain barrier (BBB) penetration, to anticipating clinical performance in humans [8]. This serves as a prelude to a review of artificial intelligence (AI) applications in neurological drug discovery, drawing on both our work and that of others. Firstly, we present a brief rationale for the need of such approaches.

The need for treatments, the cost of neurological diseases and the opportunity

There is no doubt that neurological diseases inflict an incredible cost on 100's millions of patients in society. Over the years we have seen some quite staggering statistics and numbers which were likely already outdated by the time they were published. For example, AD is an irreversible neurodegenerative disease widely known as the most common cause of dementia, affects over 5.1 million adults over age 65, costs over $150 billion per year [9] and is the 6th leading cause of death in USA. AD is a typical age-dependent neurodegenerative disease that affects 5 ​% of individuals >65 years, 20 ​% of those >85 years, and more than one-third of those >90 years [10]. However, approximately 5 ​% of the 6.5 million people in the United States have younger-onset AD diagnosed under age 65. The mechanisms underlying AD pathophysiology are still unclear. Aggregation of tau and amyloid beta (Aβ) proteins as well as decreased acetylcholine (ACh) are the focus of many studies [11]. To date several in vivo models of AD have shown some improvement in cognition as well as tau and Aβ pathology [11]. While there are recently approved biologics (e.g. aducanumab and lecanemab-irmb), the only small molecule therapies that are currently on the market for treatment of AD are four acetylcholinesterase (AChE) inhibitors (tacrine, rivastigmine, donepezil and galantamine) and one N-methyl-d-aspartate (NMDA) receptor inhibitor (memantine) which are only symptomatic treatments and do not affect the underlying disease mechanisms or alter the disease course [9]. So, there is a huge opportunity to address this unmet need for AD.

Another important neurological disease is epilepsy, which affects around 50 million people worldwide, accounting for one third of the world's neurological disease burden [12,13]. The World Health Organization (WHO) estimated that 5 million people are diagnosed with epilepsy each year globally. The cause of epilepsy is still unknown in about 50 ​% of cases, although several disease mechanisms may lead to epilepsy. Critically, toxicity is one of the most important reasons for the failure of most clinical studies, with CNS toxicity linked to a 30 ​% clinical trial failure rate [14]. One of the more common drug-induced CNS issues is seizure liability, which leads to many challenges including reduced competitiveness, project delays, and increased costs of developing these drugs [15,16].

A major focus of many CNS targeting drugs is treating acute and or chronic pain. A study in 2011 by the Institute of Medicine found that chronic pain affects at least 100 million adults and costs society > $560 billion annually in the USA alone. Decades of overreliance on opioids such as fentanyl for treating chronic pain (despite their poor ability to improve function) has contributed to an epidemic of opioid overdose deaths and addictions in the US. The total number of synthetic opioid-related deaths in recent years has hovered between ∼74,000–76,000 according to the National Center for Health Statistics at the CDC [17]. One study recently estimated the annual financial cost to be around $504 billion [18]. Despite the enormous cost of opioid use disorder (OUD) in the United States and world-wide, there has been limited success in the creation of treatments. These examples may just be the tip of the iceberg in terms of a small selection of neurological diseases, but cumulatively the numbers of patients and costs to society of are collectively immense. If we remember that there are over 1000 neurological disorders (common and rare) [19], hence finding ways to develop treatments for all of these is a herculean challenge for any technology, but in recent years machine learning and AI approaches have appeared from disparate areas that may provide some hope or at least a vision of what may be possible in future.

Introduction to machine learning and AI

One way to differentiate machine learning from artificial intelligence (Fig. 1) is to remember the former learns from data and makes decisions on that within a defined set of rules and does not create anything new. AI however can learn from the data and make decisions based on it that to go beyond the training data so that it can create something new or optimize an outcome. When we build a machine learning model, we are generally using one of many potential algorithms to pick up the pattern in the properties in the dataset that relates to a measured output of interest. For example, with a set of small molecules we can generate potentially thousands of different types of molecule descriptors (features that describe the 1D, 2D, 3D properties or features of molecules) which we can then use with an algorithm (e.g. support vector machine, random forest, deep learning or another) to either classify the measured activity related to the data or to generate a regression model (to predict a continuous value). There are many examples in the following sections on different types of data that are modeled relating to neurological diseases, although this is dominated by in vitro data due to the increasing quantity available. There are far fewer examples of true AI applications were something new is created like a molecule for a specific protein involved in a neurological disease, likely because these techniques are at an earlier stage of technological development and validation.

Fig. 1.

Differentiating traditional AI (machine learning) from Generative AI.

Finding molecules for neurological diseases using AI

An often-used analogy in drug discovery, is that it is like finding a needle in a haystack. If you are looking for a small molecule there are billions of possibilities [20], so narrowing it down to a manageable number to make and test would be a viable starting point that would be financially feasible without breaking the bank. There are several ways to accelerate this search that are commonly used. One would be to take a target implicated in a neurological disease and dock millions of molecules into the binding site to filter them down to a set to ultimately test in vitro before in turn selecting the best to test in vivo. Of course, until recently if there was no crystal structure of a target you would need to generate a homology model. AlphaFold [21] has changed this using AI technology literally overnight, and this has greatly improved the numbers and quality of such models of human proteins available (in most cases). As an example, a recent study docked more than 16 million compounds in models derived from AlphaFold and other homology models for the trace amine-associated receptor 1 (TAAR1) [22]. From 62 molecules that were purchased and tested, 25 were agonists in in vitro [22]. In the absence of the target structure, one could also generate a machine learning model that learns from experimental data (such as K_i, IC₅₀, EC₅₀). In ChEMBL, there is data for over 600 agonists which after curation and cleaning can be used to generate a machine learning models for TAAR1 as just one example (Fig. 2). While a comparison of docking to ligand-based machine learning and in turn HTS approaches was not undertaken, this kind of prospective approach needs to be performed to provide some confidence in what these technologies can do in isolation or when combined. What follows are several examples of using ligand-based machine learning models to find new molecules for various targets (Table 1). It should however be noted that as with other areas there is likely substantial publication bias of positive examples and demonstrations of ‘failures’ with these methods are rarely if ever published.

Fig. 2.

5-fold cross validation metrics for classification (A) and regression (B) models for human TAAR1 agonism using the support vector machine classification (svc) and regression (svr) algorithms. Training data was obtained from ChEMBL and filtered to only include those with endpoints associated with agonism. (A) Classification models include a histogram of the probability-like score distributions (x-axis) with red and blue colors representing the active and inactive ground truth, respectively. (B) Regression model cross validation shows the actual (x-axis) versus the predicted (y-axis) activity (-logM).

Table 1.

Machine learning model statistics for neurological targets. B, Bayesian; DL, deep learning; KNN, K-nearest neighbors; LREG, logistic regression; RF, random forest; SVC, support vector classification, SVR, support vector regression; XGB, XGBoost.

Target	N	5- fold AUC	Model type	External validation	Reference
ACHE	Eel 4545 Human 4601 Eel 5459 Human 4075 Rat 1406 Mouse 368 Cow 457 Ray 307 Mosquito 72	0.92 0.91 0.94 0.94 0.95 0.95 0.94 0.95 1	B B RF RF, SVC DL DL, RF DL, LREG, RF DL, SVC DL	66 molecules tested N ​= ​203 - AUC 0.90 N ​= ​208 - AUC 0.82	[46] [46] [49] [49] [49] [49] [49] [49] [49]
BCHE	2496	0.94	Bayesian		[46]
CCR3	699	0.94	RF, SVC, XGB	17 molecules tested	[82]
CCR4	1273	0.98	KNN, LREG, DL, RF, SVC, XGB	17 molecules tested	[82]
CCR5	1948	0.94	SVC		[82]
P2Y6R	244	0.79	RF, XGB	19 compounds tested	[104]
A1AR	132	0.87	B	29 compounds tested	[112]
GSK-3β	2368 2618	0.90 0.85	B SVR	5 compounds tested Tested with an external dataset	[147] [148]
Nav 1.8	1276	0.84	B		[164]
Kv7.1	1276	0.77	B		[164]

Enzyme machine learning models

AChE has long been considered a key therapeutic target in the symptomatic treatment of AD to treat severe cognitive deficiency [23]. AChE inhibitors have also been implicated in the treatment of many other neurological diseases including Parkinson's disease, other form of dementias and autoimmune disorders like Myasthenia Gravis [24,25]. Inhibitors of this enzyme work by reversibly blocking binding of the substrate to AChE or by hydrolytic inactivation of AChE, resulting in an increasing concentration of the neurotransmitter acetylcholine (ACh) at the synapse. Traditional synthetic chemistry efforts employing medicinal chemistry have been a fertile source of multiple classes of AChE inhibitors over the years [26], which in turn have been employed to build a wide array of quantitative structure activity relationship (QSAR) models, machine learning models [[27], [28], [29], [30], [31], [32], [33], [34]] or protein based-docking models [35]. Many computational structure-based pharmacophore modeling [36], machine learning [37], molecular docking [[38], [39], [40]], 2D and 3D similarity searches [41], MIA-QSAR modeling [42], or combinations of these different strategies [[43], [44], [45]] have also been used. This also highlights that there is considerable in vitro data to model for generally narrow chemical series, but cumulatively there is also sufficient data for machine learning.

We have demonstrated this by developing Bayesian and other machine learning models for eel and human AChE using data from ChEMBL for over 4000 molecules [46]. We have observed very similar 5-fold cross validation model statistics across machine learning models for both species. These models were also used to search the MicroSource Spectrum collection (2431 compounds) to identify novel AChE inhibitors. 66 compounds were predicted as “active” against eel AChE (inhibiting enzyme activity ≥50 ​% at 20 ​μM). Seven of these compounds were identified as novel eel AChE inhibitors including tilorone (IC₅₀ of 14.4 ​nM) for the eel acetylcholinesterase. Tilorone also showed an IC₅₀ of 73.3 ​nM against the human AChE and was verified by a secondary assay (IC₅₀ of 56 ​nM) [46].

Following this study we further curated AChE inhibition data from public sources for many other species using data from ChEMBL [47] and BindingDB [48]. For each species we built 8 machine learning models using our Assay Central software and validated them by identifying published molecules that inhibit human AChE [49]. As there are many other different machine learning approaches available, we also performed a comparison of our descriptor-based models against a more recent machine learning algorithm called AttentiveFP [50] using the larger eel and human datasets. The internal cross-validation statistics of these models was performed using nested, five-fold cross validation (apart from DL, which used a 20 ​% leave out set), and for AttentiveFP which used five-fold cross validation. These 5-fold cross validation statistics for the different AChE models had AUC and F1 scores >0.9. To further test the utility of the models developed we used external validation for the human and eel models performed using test sets that had been manually curated from the literature for additional inhibitors of human and eel AChE. As we had generated many different types of models we predicted these test sets against consensus models for each species and ultimately demonstrated that the human and eel consensus models could predict activity in the test sets with 81 ​% and 82 ​% accuracy, respectively. We also illustrated that although the models had good predictions for general AChE inhibition, there was no species specificity. Besides classification models, regression models were also developed for the human and eel datasets which were of a smaller size. We found that the SVR algorithm outperformed all the other algorithms for both human and eel with R² values of 0.81 and 0.75, respectively. These regression models were subsequently used to predict the potency of the compounds in the human and eel external test sets and the human test set predicted with the human regression SVR model performed the best (Pearson's r ​= ​0.76, mean absolute percentage error ​= ​9.73 ​%). In order to perform prospective prediction these models were used to score >195,000 compounds using a larger human AChE model machine learning model consisting of over 10,000 molecules. This resulted in 111 compounds predicted to be active with an applicability domain score ≥0.8. From these we tested ten compounds for potential AChE inhibitory activity and eight of the ten selected compounds displayed modest inhibition of AChE at 100 ​μM, (with six of the ten displaying >10 ​% activity at this concentration and 2 of the that IC_50's of between 7 and 14 ​μM). This demonstrated that the machine learning model virtual screening could suggest new human AChE inhibitors [49]. In contrast to the dogma around quality of literature data derived from different laboratories, we have demonstrated how careful curation of such data can be used to develop machine learning models with value in finding new molecules with biological activity.

A closely related enzyme is butyrylcholinesterase (BChE) which is a hydrolase responsible for hydrolyzing esters of choline including ACh [51,52], but with less efficiency than AChE [53]. Accumulating evidence indicates that BChE may have neuronal and non-neuronal roles in the CNS due to the wide distribution of neuronal BChE in certain brain regions [54,55]. It is widely known that degeneration of the basal forebrain cholinergic system is an indication of AD [56]. Studies have found that BChE biochemical properties were changed in neurodegenerative diseases like in AD. For example, in the cortical region the BChE level is increased during AD which is responsible for plaques and neurofibrillary tangle formation [56]. Additionally the inhibition of BChE results in an increase in the ACh levels in the brain, which further suggests a regulatory function in the hydrolysis of ACh [57]. BChE is also involved in the regulation and development of the nervous system in several species [[58], [59], [60], [61]]. BChE inhibitors have therefore been suggested as a treatment for AD in which the AChE-to-BChE ratio changes [62,63] and would additionally counter the side effects of AChE inhibitors such as their gastointestinal and hepatotoxic effects [[64], [65], [66]] ¹⁸. The identification of selective BChE inhibitors has lagged behind AChE research but has become attractive, and remains challenging due to the structural similarities between AChE and BChE, (65 ​% amino acid sequence homology [53,54,[67], [68], [69]] and similar catalytic active site, gorge structure and peripheral anionic site [67,69,70]). We have therefore investigated the use of several machine learning algorithms and different model training procedures in an effort to identify selective BChE inhibitors in a large library of compounds. As described previously human AChE and additionally BChE inhibition data (IC₅₀) were acquired from CHEMBL [49] with only compounds with inhibition data for both targets included in the training dataset [71]. Several machine learning algorithms, including the state-of-the-art contrastive learning (CL) were used to identify selective BChE inhibitors. CL is a machine learning algorithm that was originally developed to label data without supervision [72,73], followed by a supervised version [74] of the same learning schema to generate a more accurate representation of data points and hence provide improved classification accuracy [71]. In addition to the supervised CL, we tested a deep learning and random forest model for comparison. The 3 algorithms were used to generate many models and predict a vendor library consisting of ∼5 million compounds. After filtering with absorption, distribution, metabolism, excretion and toxicity (ADME/Tox) models we shortlisted 30 compounds in total (12 from the DL model, 11 from the CL model and 7 from the RF model). 20 compounds were purchased and tested in vitro using both human BChE and AChE assays. Ultimately we identified 7 promising selective BChE inhibitors which did not inhibit AChE appreciably [71].

GPCR machine learning models

G protein-coupled receptors (GPCRs) are essentially sensors that enable the nervous system to respond to external stimuli [75]. They have very similar 7 transmembrane alpha helices and can engage different types of molecules of varying sizes. 35 ​% of FDA approved drugs act via these GPCR proteins and are linked with many major neurodegenerative diseases (such as AD, Parkinson's disease (PD), Huntington's (HD), depression, attention deficit hyperactivity disorder (ADHD), schizophrenia etc.). An excellent recent 2023 review by Wong et al. [75] describes the extensive role of GPCRs in these neurodegenerative diseases, although it did not delve into how computational, or AI approaches could be used to identify ligands for these receptors for neurodegenerative diseases. An earlier review on computational approaches in GPCR ligand discovery had highlighted structure-based design and ligand-based approaches and workflows with some examples related to CNS targets such as the dopamine D2 receptor, D4 receptor, mu-opioid receptor 5-HT_1A, 5-HT_1A 5-HT_2A, 5-HT_2B, and adenosine A1A, A2A, and purinergic P2Y receptors [76] with only limited examples of molecules that were described as “in clinical trials”. What follows in this section are several examples of applications of machine learning and AI to GPCRs (Table 1).

Chemokine receptors are G protein-coupled receptors implicated in numerous diseases [[77], [78], [79]]. Recently, CCR3 and CCR4 have emerged as potential stroke targets as these genes are predictors of both infarct volume and edema in stroke patients undergoing thrombectomy [80]. In addition, blocking multiple chemokine receptors was suggested as a potential treatment for neuropathic pain alongside opioids [81]. We have curated data for CCR3, CCR4, and CCR5 antagonists from the ChEMBL database and used various machine learning models for virtual screening, followed by testing 17 compounds in vitro [82]. In the process we demonstrated the repurposing of the FDA approved dual tyrosine kinase inhibitor lapatinib which had in vitro activity against all three CCRs [82]. CCR5 has also recently been implicated as a target in stroke and traumatic brain injury by others [83].

Modulators of purinergic signaling including 12 GPCRs (8 P2Y and 4 adenosine receptors) are relevant to many pathological conditions and therapeutic development [84], with the G_q protein-coupled P2Y₆ receptor (P2Y₆R) being one of the most important targets involved in cell proliferation, inflammation, cerebroprotection and immune responses [85,86]. The activation of the P2Y₆R has been linked to AD, PD, epilepsy, chronic neuropathic pain and other diseases [85,[87], [88], [89], [90], [91], [92], [93], [94]] yet there are few selective P2Y₆R antagonists although there has been considerable research [[95], [96], [97], [98]]. This research has focused on different classes of compounds such as diisothiocyanate derivatives [99,100], chromene derivatives [101,102] and 2-(1-(tert-butyl)-5-(furan-2-yl)-1H-pyrazol-3-yl)-1H-benzo [d]imidazole derivatives [103]. We have used public data for P2Y₆R agonists to develop machine learning models and validated these classification models through in vitro testing (human P2Y₆R and hP2Y₁₄R) [104]. Compounds identified included ABBV-744, an experimental anticancer drug which inhibited hP2Y₆R and showed hP2Y₁R enhancement. Other compounds identified included the experimental anti-asthmatic drug AZD5423, as well as TAK-593 and GSK1070916 [104].

Adenosine is an another extracellular signaling molecule which is involved in numerous physiological functions throughout the brain and periphery and activates membrane-bound GPCRs [105] including adenosine receptors (A₁AR, A_2AAR, A_2BAR, and A₃AR) [[105], [106], [107], [108]]. A₁AR in the brain slows metabolic activity for example at it reduces synaptic vesicle release at neuronal synapses and is implicated in sleep promotion [109], antiseizure as well as neuroprotective activity [108]. A₁AR activation can also inhibit presynaptic Ca²⁺ influx-related release of transmitters to provide neuroprotection [110]. A₃AR activation reduces inflammation and chronic neuropathic pain [107]. Selective compounds for the AR subtypes can potentially be used to control addictive psychostimulant consumption along with many other health problems [111]. We used machine learning models to find novel modulators with public data for A₁AR agonists to generate and validate a Bayesian machine learning model and to identify molecules for testing [112]. Febuxostat and paroxetine were two FDA approved drugs which we identified as inhibitors of orthosteric radioligand binding for A_2AAR and A₃AR. For human A_2AAR, stimulation of cAMP was observed for crisaborole (EC₅₀ 2.8 ​μM) and paroxetine (EC₅₀ 14 ​μM). Crisaborole also increased cAMP accumulation in A_2BAR-expressing HEK293 ​cells. Paroxetine was found to bind human A₃AR (K_i value of 14.5 ​μM), indicating antagonist activity [112]. These represent a few examples of where machine learning could be applied to GPCRs.

Kinase machine learning models

Glycogen Synthase Kinase 3 (GSK-3) is a serine-threonine kinase with broad specificity involved in inflammation, cell growth, cell differentiation, and energy metabolism [113,114]. It is also associated with age-related diseases like diabetes, AD, cancer [114,115] as well as rare diseases such Rett syndrome [116] and cyclin-dependent kinase-like 5 (CDKL5) disorder [117]. GSK-3 exists as two isoforms, GSK-3α and GSK-3β, with the latter found in higher abundance in the CNS and increases with age [118]. GSK-3β is also key in both peripheral and central inflammation, where it mediates inflammatory responses in astrocytes and microglia [[119], [120], [121]]. Inhibition of GSK-3β activates the Wnt pathway to promote neurogenesis [122] and therefore could be used to treat CNS diseases [123]. The brains of AD patients demonstrate increased levels of activated GSK3β [124,125], which phosphorylates the microtubule-stabilizing protein tau [[126], [127], [128]]. GSK3β inhibition also downregulates expression of soluble paired helical filaments, which are the major component of neurofibrillary tangles (NFTs) which are intracellular plaques [127,129,130]. GSK3β activity is further linked to the improper processing of the amyloid-precursor protein APP [131]. GSK3β inhibition downregulates expression of the β-site APP-cleaving enzyme 1 (BACE1) and the BACE1 gene reducing aβ production [132]. Tau hyperphosphorylation can hence be reduced by inhibiting GSK-3 [133] e.g. by PKC, which in turn is stimulated by M1 receptor agonists, including ACh, and a number of recent studies have proven that such products can reduce tau phosphorylation and aggregation [134]. While there are dozens of GSK3 inhibitors, most of these inhibitors are not suitable as therapeutics due to their toxicity [135], although some have displayed anti-AD activity in rodent models including lithium chloride [136], kenpaullones, tideglusib [137], 2-(alkylmorpholin-4-yl)-6-(3-fluoropyridin-4-yl)- pyrimidin-4(3H)-ones [138], isonicotinamides [139], pyrazine analogs [140], andrographolide [141], SAR502250 [142], AM404 [143], PF-367 [144]. Few GSK3β inhibitors have made it to clinical trials for AD or other diseases and none have shown promise in treating the disease [145,146]. We initially developed a GSK3β Bayesian machine learning model with 2368 molecules from ChEMBL (five-fold cross validation ROC of 0.905) which was used for virtual screening libraries of FDA approved drugs and clinical candidates [147]. In vitro testing of 5 selected compounds identified ruboxistaurin (GSK3β avg IC₅₀ ​= ​97.3 ​nM) and 2 other diverse inhibitors as hits. Ruboxistaurin (LY-333531) is a protein kinase C (PKC) inhibitor. Our more recent machine learning studies have used 2618 GSK3β inhibitors to build regression models with 7 different algorithms with the SVR model performing the best with 5-fold cross validation (MAE 0.52) [148]. These models could be used to screen much larger datasets to identify novel scaffolds that could overcome limitations in the molecules that have reached the clinic.

Ion channel machine learning models

Voltage-gated sodium channels (VGSCs) such as Na_v1.1, Na_v1.6, Na_v1.7, Na_v1.8, and Na_v1.9 which are expressed by adult sensory neurons [149] are known to be key determinants of sensory neuron excitability essential for: initial transduction of sensory stimuli, electrogenesis of the action potential, and neurotransmitter release. Therefore these ion channels are involved in chronic pain due to injury and disease caused by sensitization of the somatosensory nervous system involving these primary sensory neurons. Changes in expression of VGSCs, as well as post-translational modifications, can contribute to the sensitization of sensory neurons in these chronic pain states. Additionally, the gene variants of Na_v1.7, Na_v1.8, and Na_v1.9 have been linked to human Mendelian pain disorders [150,151]. For example, Na_v1.8 is a sodium ion channel voltage-gated channel subtype that is tetrodotoxin (TTX)-resistant and encoded by the SCN10A gene in humans [[152], [153], [154], [155]]. Na_v1.8 has a role in nociception and is also expressed in the dorsal root ganglion (DRG), in unmyelinated, small-diameter sensory neurons called C-fibers [156,157]. As Na_v1.8 is located in sensory neurons of the DRG this makes it a potential target for the development of new analgesics [158] for chronic pain [159]. Several drug companies have described Na_v1.8 inhibitors [160] including the suzetrigine from Vertex [161,162] which was recently approved by the FDA. Na_v1.8 was of initial interest to us as a validated target for the rare disease Pitt Hopkins Syndrome [163]. We initially screened the 1280 compound Prestwick chemical library using a Na_v1.8 FLIPR high throughput screen and identified 93 hits (7.2 ​% hit rate) at >2SD from the mean response (Z’ ​= ​0.6). This screen also identified numerous dihydropyridine calcium channel inhibitors such as nicardipine which were further tested using automated patch clamping (nicardipine IC₅₀ of 0.6 ​μM). At the same time, we used our Assay Central software to develop Bayesian machine learning models using ECFP6 descriptors in order to develop a model with this Na_v1.8 inhibition data (5-fold cross validation showed an ROC of 0.84 [164]). Outside of this study we further curated literature data from ChEMBL on several additional sodium channels in order to build machine learning models (Na_v 1.4 (ROC 0.80), Na_v 1.5 (0.87) and Na_v 1.7 (0.84)). In total, these models could also be used for virtual screening to identify potential selective inhibitors for each of these sodium ion channels.

Predicting molecule liabilities

Many drug candidate failures can be attributed to inadequate efficacy in animal models or in late stage clinical trials, indicating that improving the understanding of the candidates action in more complex systems could increase overall success [165]. In vitro or computational screening could be critical in doing this, while machine learning methods are being increasingly utilized there is still a shortage of datasets for in vivo endpoints. It is subsequently impossible to computationally determine these complex biological phenomena [166]. Increasingly, drug screening efforts skip to animal experiments which does not help when there is pressure to reduce animal use (e.g. more recently from the FDA and other regulators), and these efforts also require milligrams to grams of precious active compound to be synthesized, which may be far out of reach for those small companies that are resource constrained and likely working on rare and neglected diseases. It is at this juncture that computational approaches as an alternative may have their moment to impress. The following represent several examples of where computational approaches can be used to predict these molecule liabilities relevant for CNS drug discovery (Table 2).

Table 2.

Machine learning model statistics for CNS properties. B, Bayesian; BBB, blood brain barrier; DL, deep learning; RF, random forest; SVC, support vector classification, XGB, XGBoost.

Model	N	5- fold AUC	Model type	External validation	Reference
Hallucinogenic potential	Psychlight B 84 PIKHAL and TIKHAL 221	0.76 0.67	SVC B	22 molecules 22 molecules	[174] [174]
Drug induced seizures	67 targets with different numbers of molecules – led to ∼2000 individual models - 43−5470 molecules per target representing 59,419 molecules	Average 0.88	RF, SVC, XGBoost	25 ​% leave-out set, one hot encoding and ProtBERT were also used and did not improve predictions.	[176]
BBB	2296	0.94	B	Tested on 40 external molecules known to cross the BBB – 92.5 ​% correctly predicted.	[200]

Predicting hallucinogenic potential

Psychiatric medicine is undergoing increased interest with psychedelics effectively treating numerous hard to treat diseases like substance abuse disorders, depression and previously treatment-resistant depression [[167], [168], [169]]. Psychedelics have a checkered history, are legally controlled in most countries and may act as agonists of serotonin 5-HT_2A, as well as other receptors (e.g. dopaminergic and glutamatergic systems), to induce neuronal structure changes [169,170]. Development of psychoplastogens that avoid the psychedelic effects might benefit from efforts to model and predict this potential. Recent efforts to develop biosensor assays for 5-HT_2A such as PsychLight [171] was used in vivo and in vitro. A human in vivo hallucinogenic dataset [172,173] covers phenyethylamines and tryptamines. These datasets were used with various machine learning approaches including pretrained large language models (MolBART), conformal predictors and with various molecular fingerprints to predict potential psychedelics [174]. Conformal predictors generally improved the predictive accuracy of our machine learning models over support vector classification (SVC) and removed test set predictions with high uncertainty. We also curated molecules that were tested in mouse head-twitch assays and a set of 5HT_2A agonists [175] to test the models [174]. These models may offer an alternative approach to predict hallucinogenic potential prior to in vitro or in vivo testing.

Predicting Drug Induced Seizures

Drug Induced Seizures (DIS) are a component of CNS toxicity which leads to the high clinical failure rate observed of such drugs. Being able to predict this potential earlier in the drug discovery process would therefore enable us to likely avoid these failures and prioritize testing in experimental models of DIS. We have developed machine learning models for 67 target proteins implicated in seizures to predict this liability [176]. We also compiled a list of seizure liability compounds from various public references including a machine learning paper from 2011 that focused on predicting seizure liability (771 compounds) [177] as well as additional papers assessing the predictivity of their in vitro seizure liability models [[178], [179], [180], [181], [182], [183], [184]]. We also included the dataset (described earlier) which we compiled [185]. Once combined, this training dataset comprised 966 unique compounds (453 active, 543 inactive). Models were built using both ECFP6 and MACCS keys. Overall, these models had reasonable 5-fold cross validation scores with both ECFP6 descriptors (Fig. 3A) and MACCS keys (Fig. 3C), with the latter performing better overall. Example ROC plots are also shown for the best algorithms, specifically for K-nearest neighbors (knn) and random forest (rf) for ECFP6 and MACCS keys, respectively. These models represent a generalizable way to mitigate the potential of seizure liability early in the process in drug development.

Fig. 3.

Summary of the 5-fold cross validation metrics of models predicting general seizure liability using either (A,B) ECFP6 or (C,D) MACCS keys molecular descriptors. ROC curves are also shown for the best performing algorithms K- nearest neighbors (knn) (B) and random forest (rf) (D).

Blood brain barrier models

Considering the efficiency of the BBB, it is difficult to design pharmaceutical products that can effectively be delivered to the CNS. The BBB changes in response to each pathological condition, requiring extensive knowledge of how to protect the CNS from possible challenges or inflammatory immune cells while providing the necessary therapeutics. Various biological, chemical, and physical methods exist to open the BBB, including the use of cyclodextrin, viruses, and microwaves [186]. Some approaches for traversing the BBB include opening tight junctions, transendothelial passage, efflux pump disruption, and targeted receptor mediated transport. Our first published efforts modeling BBB were from 2008 [187] when we and others had published coincidently and reached the conclusion that such models were probably at their peak [187]. While available datasets are now many times larger, their relevance to the human BBB is unchanged. There are now many machine learning models or rules developed to predict the ideal molecular properties to cross the BBB in mouse [188,189] as well as quantitative structure property relationships (QSPR) [190] and various types of machine learning models [[191], [192], [193], [194], [195], [196], [197]]. Wager et al. published their analysis of 119 CNS drugs and developed a method to score 108 CNS candidates from Pfizer programs with a multiparameter optimization (MPO) algorithm [198] using molecular properties (ClogP, ClogD, MW, TPSA, HBD and pKa). This correctly scored 74 ​% of a set of 119 marketed drugs while 60 ​% of Pfizer's own drug candidates scored favorably with this MPO score. Another example tested an updated method on 21 CNS candidates with 77 ​% correct [199]. We recently reimplemented these methods [200] using ChemAxon descriptors to replace the ACD software used originally [198] as well as developed a Bayesian BBB machine learning model based on published data for over 2000 molecules. Evaluating the later model on 40 molecules from the MPO dataset demonstrated >90 ​% correct predictions and illustrated that such methods could have benefits over relatively simple rules like the MPO method [200]. These BBB models will likely not be perfect, as we have seen over the decades there are issues around the datasets used and the general lack of prospective validation. We have also seen pharmaceutical companies share results, but they generally do not share the underlying mouse BBB data relating to proprietary compounds.

AI and Machine Learning approaches for drug discovery of neurological disorders

The impact of AI and machine learning on drug discovery in the areas of neurodevelopmental disorders, depression, PD, AD, anesthesia and pain treatment has been reviewed, providing select examples [201]. There are numerous examples of the use of different AI and machine learning approaches integrating different types of data at various stages of the drug discovery continuum and what follows is only a small sampling in the interests of space. It should be noted that most of the CNS-related multimodal approaches are for clinical applications (MRI, CT data etc.) rather than for drug discovery. What follows are various applications of these AI and machine learning approaches applied to various neurological disorders.

Knowledge graphs and their applications

A large knowledge graph was described which used data from 29 public sources that captured data on 1552 small molecules and 137 complex diseases [202]. A machine learning approach then used this data to predict the probability of treatment (AUC ROC ​= ​97.4 ​%), leading to good predictions for DrugCentral and novel clinical trial indications resulted. Two case studies were provided, one being nicotine dependence and this provided molecules already used or in clinical trials for smoking cessation (e.g. galantamine). A second case study was epilepsy, and this provided some known anti-epileptics as well as novel compounds such as acamprosate (a GABA-A receptor modulator). In addition, 15 additional molecules that cause epilepsy were also provided [202]. The results of this approach would have been more impressive if prospective testing had been performed.

IBM Watson for Drug Discovery was used to build a knowledge network between psychiatric and neurological disorders, 1588 genes and 722 drugs from scientific publications over the past 50 years [203]. The network had 2 distinct groups including psychiatric diseases (anxiety, depression, schizophrenia, autism and bipolar disorder) and neurodegenerative diseases (AD, PD, HD and dementia). In the process, drugs targeting genes related to these diseases were uncovered that have yet to be used to treat them, including monoclonal antibodies (anti-CD33). The approach can also be validated by identifying drugs that are in clinical studies for treating the same diseases, but which lacked publications [203]. Again, a limitation of this study is that they did not however test any compounds to validate the predictions.

Using genome wide association studies, a network topology based deep learning framework NETTAG was used to identify AD disease associated genes [204]. Using this approach 156 AD-risk genes were found to be enriched in druggable targets. Network proximity approaches were used to computationally identify 118 high scoring candidates from nearly 3000 FDA approved drugs and gemfibrozil (a lipid regulator which is used in hypercholesterolemia) was top scoring. Using a database to search medical data on over 10 million patients, gemfibrozil was also identified as associated with decreased AD incidence as patients showed a 43 ​% decrease in disease risk [204].

Natural language processing has been used for mining over 30 million articles in PubMed to identify text relationships between diseases, genes and proteins to construct a knowledge graph using SemNet 2.0 and rank molecules [205]. This was then used to aid repurposing of drugs for PD starting from the known drug levodopa. Several antihistamines were identified (including ebastine, levocetirizine) the antibiotic vancomycin, the NMDA agonist neramexane and the angiotensin converting enzyme (ACE) inhibitor captopril. In several of these cases e.g. antihistamines, there was good evidence in animal models. These repurposed drugs have a role in oxidative stress, neurotransmitter imbalance, and inflammation. Limitations of the approach are the proposal of drugs that may not be available or approved in the US as well as having possible undesirable side effects. In addition, crossing the BBB is also important and some of those selected may not be able to [205].

Generative machine learning and transformer approaches

The evolution over 50 years from quantitative structure activity relationships (QSAR) to machine learning for neurological drug discovery has been briefly summarized [206]. It has evolved from limited data and small regression models to comparably very large datasets. Several examples were provided of molecules like the BACE-1 inhibitor elenbecestat, which was identified by machine learning, that ultimately failed in a phase III trial for AD. A second example was the AI designed molecule DSP-1181 for obsessive compulsive disorder which reached the clinic in 2019. It was suggested that the future of machine learning in neurodegenerative drug discovery is likely linked to integration with other approaches like personalized medicine and imaging [206].

Most of the machine learning examples described thus far use a model to score an array of molecules to predict a bioactivity or a property (e.g. BBB). Ideally there are multiple properties alongside bioactivity that need optimization, such that we could also start with a desired property value and design molecules that fulfill these requirements. A generative model is one way to achieve this as it learns the probability of a chemical with the desired properties given the structure [207]. There have been numerous discoveries in the field of de novo drug design [208,209] and several different algorithms such as Generative Adversarial Networks (GANs) [210], Variational Autoencoder (VAE) [211,212] and Recurrent Neural Networks (RNN) [209] have been used to design molecules for various targets [[213], [214], [215], [216]]. These methods can use different molecular representations including 2D strings [217,218] or 3D graphs [219]. Increasingly, examples include training with molecules from ChEMBL or other large sources of molecules (e.g. USPTO), using such methods as a two-layer long short term memory (LSTM)-based RNN that was used to generate libraries of molecules as SMILES (one dimensional strings) followed by transfer learning to tune the model and then generate a library of leads from a starting known active fragment [220]. Another example generated molecules based on predicted molecular properties [221]. Most of the published generative examples have optimized against a single physicochemical activity or a bioactivity and clearly this is a limitation as this will not help if the molecule needs selectivity against closely related targets or to avoid off-target effects. Several groups have also focused on graph structures which had returned 100 ​% valid structures [219]. Other algorithms such as transformers and attention-based models have also been used. The biggest challenge for these generative approaches is the limited prospective experimental validation which involves synthesizing compounds and testing for activity [222] or finding structurally similar compounds to those designed from vendors [223]. Generative models can propose molecules like a starting structure, explore chemistry space or fit design criteria such that a single generative model may not be able to respond to all the desired molecule design needs.

De novo design approaches generally have been frequently used to design molecules for a single target, however, there have also been efforts to also do this for multiple targets simultaneously. One example is MTMol-GPT a multitarget generative pretrained transformer model which uses a dual discriminator model with inverse reinforcement learning [224]. Both the 5HT_1A and dopamine receptor D2 were targeted for neuropsychiatric applications. Docking (AutoDock Vina) and pharmacophore mapping were then used to evaluate 1000 designed ligands. In general, the molecules had improved docking scores compared to known ligands. There was also good mapping to pharmacophore features. It should be noted that none of these molecules were synthesized and tested, while there was also no comparison of how far the molecules are from known ligands for both receptors [224].

Neuropeptides are produced by neurons attached to GPCRs and they play important roles in modulation of brain function including sleep, food intake and have neuroprotective roles that may be important for PD and AD. A recent machine learning model trained on 6807 neuropeptides (5–100 amino acid residues) and 8647 non-neuropeptides collected from public databases used a two-phase approach in which a bidirectional encoder representation from transformers (BERT), capturing the functional and structural aspects of the neuropeptide then a non-dominated sorting genetic algorithm which explores and changes the population of neuropeptides [225]. 80 ​% of the data was used for training while 10 ​% was used for validation and a further 10 ​% for testing. The BERT-NeuroPred model had a 91 ​% accuracy, however again no external testing with prospective designs was performed. Secondary and tertiary characteristics of neuropeptides were also not incorporated into the objectives of the model [225].

Docking and other applications with machine learning

Docking of small molecules could likely populate many reviews on the topic. Challenges here include the scoring function used to rank molecules before selection. Applications of docking for neurological diseases combined with machine learning reveals few examples. For example, α-synuclein aggregation is linked to the neurodegeneration which occurs in PD, therefore drugs that can block this can be disease modifying [226]. Docking of 2 million molecules to α-synuclein fibrils led to 79 compounds of which 68 were tested and 4 were found to be active. Similarity searching resulted in additional hits and enabled a training set for machine learning of 161 molecules. These models iteratively enabled the selection of more molecules for testing and showed 20-fold higher hit rates than high throughput screening as well as increased novelty. Limitations of this approach included screening a library of existing molecules as well as focusing on just one parameter [226].

As the human interactome is massive with ∼650,000 protein-protein interactions (PPI) these have key roles that if dysregulated they can cause disease and hence represent drug targets. There are therefore few if any examples of PPI inhibitors developed by AI specifically for CNS targets. A recent review summarized the evolution of computational approaches in this area and was largely silent on CNS targets [227] so there is certainly a potential opportunity here.

Replacing animal models with AI

In the area of AI in neuroregeneration, only 12 ​% of the papers relate to drug discovery, with most relating to diagnosis of disease and rehabilitation, and the methods were used for analysis of imaging data as well as animal models [228]. Much of neurological research uses animal experiments, but new AI approaches may offer one way to replace these such that the models are more human centric. The scope of such applications can reduce the need for animals in screening, length of longitudinal studies to understand disease progression, diagnostic development, neurotoxicity testing, minimize studies for brain computer interface development and potentially replace some developmental testing. By accumulating published examples that illustrated this potential of AI in neurology research it set the foundation for exploring how AI might evolve in future and the potential for impact in different areas [229]. Some limitations were noted such as relying on data from certain populations for model building which could bias models, overfitting of data due to misleading predictions from messy data, and predictions that may be difficult to explain. While very futuristic this review made the point that AI has the potential to replace animal experiments in neurology [229]. One might argue also that this might be feasible in other disease areas too.

Behavioral screening for discovery of neuroactive drugs does not have to rely on rodents or other animals, such as the even the aquatic invertebrate the freshwater planarian, to generate phenotypic data [230]. This can be used alongside physicochemical property data of molecules to generate classifiers for prediction of phenotype. One proof of concept study used 19 neuroactive molecules spanning anxiolytics, antidepressants and antipsychotics. The classification accuracy using PCA and other projection methods was similar for physicochemical and phenotypic data [230]. While this was a very small dataset there was no cross validation or external predictions, it does leave out the possibility that a larger training set would be of more potential value. Insects such as locust have also been used to model brain uptake of drugs. A recent study generated data on 25 drugs and used this to build machine learning models with simple interpretable descriptors to predict brain uptake [231]. The small dataset was used with cross validation and a hold-out test set to produce acceptable results, again suggesting the potential of this system for screening compounds and building models with bigger datasets. As yet these few examples suggest too little data for building machine learning models that could be used for general screening for this ability to reach the brain.

Clinical data applications of AI

Compared to using AI in drug discovery, applications to clinical data are relatively rare. One study integrated clinical whole genome sequencing and genetic data from 235 patients to build various machine learning models with nested cross validation to predict the response for the antiepileptic brivaracetam [232]. 106 clinical and 40 genetic features (4695 genetic variants) were used to evaluate different model types. A gradient-boosted tree model performed the best on cross validation (AUC 0.76) and with an external test set of 47 patient samples (AUC 0.75) to validate the model. Interestingly the same study also assessed the placebo response using the clinical features only, but was unable to predict this response better than random [232].

Approaches identifying potential clinical candidates for diseases include searching for associations across medical records for large numbers of patients using machine learning approaches. When many of these studies are performed globally, they can be combined into a systematic review. Efforts to do this to find drugs that modify dementia risk or AD yielded several different drug classes such as vaccinations, antimicrobials and anti-inflammatories that decrease risk while diabetes drugs and antipsychotics increased risk. Some classes such as antidepressants and antihypertensives produce conflicting evidence [233]. Again, this type of approach could be readily applied to other neurological diseases outside of those related to dementia.

The application of AI to dementia drug discovery and clinical trial design was recently reviewed and one example focused on unsupervised approaches to identify subgroups or progression patterns within cross-sectional databases, while another approach described subtyping AD using machine learning from RNA data [234].

Continuous electroencephalography (cEEG) is used for detection of seizures as well as monitoring depth of sedation and ischemia. Increased use of the technique has also led to the detection of non-convulsive seizures [235]. Challenges with the approach include the review of the data. AI has been applied to seizure prediction using machine learning to produce risk scores, EEG screening, automated annotation and interpretation of the EEG, seizure management, the diagnosis of non-convulsive seizures and paroxysmal events, treatment of seizures and status epilepticus, detection of cerebral ischemia in subarachnoid hemorrhage, neuroprognostication in traumatic and anoxic brain injury, as well as anesthetic dosing and sedation management. Even in this area, there are barriers including computing resources, lack of prospective studies, generalizability of approaches due to biases in training data, the need for larger training sets as well as lack of validation that need to be considered [235].

Drug repurposing in AD (DRIAD) is a machine learning approach which quantifies associations in AD pathology severity and molecular mechanisms [236]. When combined with human neural cell cultures exposed to 80 FDA approved predominantly kinase inhibitor drugs and RNA-seq performed resulting in gene lists that can be ranked to prioritize the compounds. The JAK inhibitor ruxolitinib scored well and suggested the JAK-STAT pathway is important in AD and worth further exploration. The results of the study were not further validated, but this could be partially aided by searching electronic health records. The BBB penetration of the compounds was also not evaluated so it is also possible while the compounds may look promising in neural cells in vitro, the delivery in vivo may be challenging and limit the value [236].

Phenotypic screening and imaging applications with AI

One early example of predicting neuropharmacological potential uses a mouse screening system that leverages computer vision to track their behavior and extract different measures for an array of 14 different compound classes called SmartCube [237]. This enabled the creation of signatures for each drug class and prospective prediction was demonstrated with 2 compounds (TP-10 and PF-670462) not in the model, one was predicted to be anxiolytic and the other an antipsychotic. These predictions were followed up with further animal studies to confirm the behavior predicted from the screen [237].

Live cell imaging of neuronal cells and screening of molecules is important for identifying new therapeutics. Screening of neuronal cell cultures combined with machine learning offers a powerful approach for learning from the imaging data to identify patterns that may be missed by human observers [238]. The development of automated imaging systems to generate large-standardized datasets for such models will also be important.

Algorithms that work well with small datasets include few-shot learning. Few shot meta learning allows models to learn from prior experiences and improve the learning performance. By combining this approach with a high throughput automated whole brain activity mapping (BAM) screen in zebrafish it is possible to improve the prediction of molecules for CNS agents like those for PD [239].

Neuroactive drugs can be repurposed for other uses. For example, a set of 132 such neuroactive drugs were profiled in glioblastoma patient material from 27 patients using single cell transcriptomes [240]. The convergence of secondary drug targets was analyzed by regularized regression (COSTAR). Logistic LASSO regression has been used to train a model to identify the secondary target genes that discriminated hit drugs from non-hits to derive a signature predictive of the ex-vivo efficacy for glioblastoma. A further 1.1 million molecules were screened with this signature model of which 23 top scoring and 25 bottom scoring molecules were tested across 4 patient samples (predictions and scores AUC 0.94) and only the hits were linked to the target BTG [240].

Reinforcement learning (RL) can be used to tune a generative model for the properties of interest and deliver molecules with a different property distribution when compared to those of a training set [241]. This method is however inefficient when compared to other methods and will take a long time to train, hence such methods are usually combined with adversarial or pretraining. RL can bridge target-based and phenotype-based drug discovery. A recent review has provided an overview of RL and may be used for systems-pharmacology in neurological disorders as well as other applications [241]. As yet these various imaging application of machine learning are in their nascent stages.

Designing illicit substances targeting the CNS using AI

A further unusual application of computational approaches is in trying to get ahead of the problem of illicit substances and one such area is in novel psychoactive substances (NPS) which are a global issue and are generally heavily regulated [242]. However, there is a consistent approach from violators where they subtly modify the molecules to circumvent legislation. A scaffold and transformer based NPS generation and screening (STNS) has been used for chemical space analysis. A discriminator model was generated from 2154 NPS molecules and 2154 negative samples from PubChem. The NPS generator is a generative model composed of T-Scaffold and Mol-GPT which was used to design over 120,000 unique cannabinoid molecules. The model was then validated and 3 novel, synthetic cannabinoids were identified that docked well in the cannabinoid receptor and were tested in vitro with a genetically encoded endocannabinoid (eCB) sensor (GRABeCB2.5) [242]. While not addressed in the article, the likely dual use nature of the work is clearly apparent as is the potential to repurpose some of these illicit molecules as potential therapeutics for neurological diseases. To our knowledge we have not seen examples in the literature, were machine learning has been used by criminals to create illicit CNS targeted molecules that have then been distributed and used. We have highlighted this potential capability in the area of nerve agents [243] and the software and datasets are clearly available.

A generative approach to developing treatments for opioid use disorder

Opioid abuse leads to complex and profound changes in the brain. Opioid use disorder (OUD) induces changes in the mesolimbic reward circuit, and these changes are thought to underlie the pathology and cycles of addiction [244]. Abrupt disuse of opioids leads to opiate withdrawal, with common symptoms including aches and pains, muscle spasms, rapid sensation of temperature swings, tremors, nausea and vomiting, anxiety and irritability, and insomnia [245]. Often times, despite the initial reasoning for opioid use, OUD occurs due to the desire to avoid these withdrawal symptoms [246]. The 5-HT2R family members, namely 5-HT_2A and 5-HT_2C, are directly involved in mediating neurotransmission in substance use disorder [247]. 5-HT_2A and 5-HT_2C are widely expressed in the brain and bring about the abuse-related effects of drugs such as cocaine, morphine and heroin through their interactions with the dopamine system (DA) [[248], [249], [250], [251], [252], [253], [254]]. Recently, a new class of molecules derived from psychedelics have been introduced known as psychoplastogens [255,256]. These molecules are described as various serotonergic and entactogens that rapidly display plasticity-promoting properties by activation of mTOR. The ability to induce neuroplasticity is thought to be the end result of increased BDNF release, which stimulates spinogenesis and neuritogenesis [257]. Critically, the psychedelic experience and induced neuronal plasticity are not entirely linked, and some evidence suggests that the psychedelic experience is separate from the therapeutic role of psychedelics [258]. While the 5-HT_2A receptor is critical for both the psychedelic experience and neuroplasticity, some psychoplastogens do not induce a psychedelic experience [256]. Recently, an ibogaine analog was synthesized which decreased drug-seeking responses and induced neuritogenesis and spinogenesis in rats while not inducing a head-twitch response, a common hallmark of a psychedelic experience in rodents [256]. This ibogaine analog was specific to 5-HT_2A, unlike traditional psychedelics which broadly target other 5-HT receptors as well. Psychedelic response has been attributed to the activation of 5-HT_2A receptors [259], and 5HT_2C has been targeted by molecules as potential treatments for addiction [[260], [261], [262], [263], [264], [265], [266]], suggesting that an appropriately selective serotonin agonist for either could treat OUD, although to date clinical translation has been mixed likely due to the small studies [[267], [268], [269], [270]].

We have recently described our generative drug discovery software MegaSyn and applied it to generating ibogaine analogs [271] that could be used for OUD. This model was first trained on a subset of ChEMBL representing ∼2 million compounds. Next, we applied transfer-learning by training for an epoch on a natural products library [272] then refined the RNN using a hill-climb maximum-likelihood estimator algorithm. The top 10 ​% of ranked compounds were kept and fed back into the model for training. The model starts by generating drug-like compounds is re-trained on the top scoring compounds until it ultimately finds those molecules that have the desired chemical properties. As a scoring reward function the Tanimoto similarity to the CANVASS library [272] (>0.2) was used as well as an activity model against 5-HT_2A (predicted active), and avoiding off-target models (hERG, 5-HT_1A, 5-HT_1F, 5-HT_2C). We wanted to obtain molecules with Tanimoto similarity >0.6 to Ibogaine and a lower cLogP. After generating 100,000 virtual molecules, all of the top 50 generated molecules had higher scores versus ibogaine and we were able to demonstrate that MegaSyn produced the psychoplastogen tabernanthalog [256] suggesting that it could discover molecules with the properties desired in this case [271]. We have recently further updated this approach (Fig. 4) and used MegaSyn in house for the design of additional molecules targeting serotonin receptors with good BBB penetration as well as other receptors [271,273] which have been validated after synthesis and testing in vitro.

Fig. 4.

A schematic overview of the MegaSyn generative approach to molecule optimization training, which consists of nested training loops [271,273]. Targets A-C could be individual CNS receptors of interest for which one desires selectivity, or they could be multiple targets that need to be hit to obtain polypharmacology goals.

Summary, limitations and future perspective

As neurological disease encompasses over 1000 disorders it would be impossible to cover them all in this review! Clearly only a fraction of these diseases has been the focus of machine learning and AI to date, in general those with larger populations and for which there is more published data. The costs of research and development require a rethink in how we can address developing treatments for more diseases which impact human health and have a dramatic financial toll particularly those targeting the CNS. Most of the diseases that have achieved the greatest attention from these efforts are those neurological diseases that are highly complex and heterogeneous which also affect the largest numbers of patients. Yet the diseases that are likely to be the most straightforward which are monogenic (and rare) have not received much attention from such machine learning and AI approaches [[274], [275], [276]]. There are few if any drugs for neurological diseases that are derived from machine learning or AI yet, although that may change soon. It is important to note that these computational approaches have been developing for many decades and are only now achieving a high level of recognition likely due to increased attention. This may also be due to the intersection of availability of large amounts of public data such as structure activity data (e.g. ChEMBL, BindingDB etc.), genomics, proteomics, metabolomics, increased computational power available (e.g. GPUs), more advanced algorithms (transformers, deep learning etc.), etc.

Just having the data and the software/hardware tools available though does not mean that the models will be built or used for drug discovery. There must be the motivation, the funding and know how to do this work. Someone has to do it! As we have described, much of the computational work is done entirely with little to no experimental validation. This is a substantial limitation. In the published examples there is a clear bias towards success, failed efforts are not discussed and limitations in the underlying data or approaches are rarely considered by the authors. If this is to change in future to impact neurological disease research, we need to be training more multidisciplinary scientists who can handle not only the computational side but also the experimental piece as well. Certainly, collaborations between computational and experimental scientists could also achieve the same end but these will need to happen and bringing together the right team is as much an art as science.

Discovering drugs is extremely difficult and perhaps more so for neurological diseases where you need more data such as BBB permeability and an understanding of any off-target effects with an array of GPCRs, enzymes or ion channels where signal dysregulation could lead to hallucinogenic effects or seizures. An added part of the challenge is not just computational, but the animal models for these diseases poorly reflect the human neurological diseases, so it is more difficult to validate computational and in vitro findings prior to going into clinical trials in humans. We are clearly embarking on the age of AI for neurological diseases with great promise although it is unclear whether it will have any more impact than any other technology in drug discovery that has been heralded as transformational over the last 30 years. This is another significant question that remains unanswered. We have described some of the nascent machine learning and AI applications whether they are individual techniques or consist of end-to end applications. Some of the more basic applications we have predominantly used and highlighted, help to focus the universe of molecules for testing in a single assay whereas some of the more complex generative approaches have the promise to design molecules with ideal properties such that few must be made or tested against one or more targets or the disease of interest. But again, we need more examples to convince other scientists and the industry in general that this is the future. We hope this review serves to inspire others to consider applying the currently commercial or academic available machine learning and AI tools to their neurological disease of interest. Alternatively, there are likely also many new opportunities to develop completely new software or applications to make the discoveries necessary for identification of new treatments for neurological diseases where there is still considerable unmet need. There are millions of patients globally with neurological diseases whether common or rare, and they need treatments urgently. That should provide the motivation to use and improve the tools we have described in order to find these transformational drugs.

Funding

We kindly acknowledge NIH funding from R44GM122196-02A1 from NIGMS and 1R44ES031038-01 from NIEHS for our machine learning software development and applications. “Research reported in this publication was supported by the National Institute of Environmental Health Sciences of the National Institutes of Health under Award Number R44ES031038. We also acknowledge 1R43DA055419-01 from NIDA. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.” NINDS have supported our work on Batten disease (2R44NS107079-02A1, 1SB1NS135733-01 from NIH/NINDS) which is briefly mentioned herein.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Sean Ekins reports financial support was provided by NIH National Institute of General Medical Sciences. Sean Ekins reports financial support was provided by National Institute of Environmental Health Sciences. Sean Ekins reports financial support was provided by National Institute on Drug Abuse. Sean Ekins reports a relationship with Collaborations Pharmaceuticals, Inc. that includes: employment. Sean Ekins has patent pending to Collaborations Pharmaceuticals, Inc. None If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.