A New Era of Artificial Intelligence (AI): Transforming Drug Discovery and Development

Artificial Intelligence (AI) has emerged as a transformative technology in contemporary drug discovery and development, fundamentally reshaping the processes by which novel drugs are identified, optimized, and translated into clinical applications. The inclusion of AI into drug discovery and development thus holds tremendous potential to change the landscape of therapeutics for various human diseases, enabling unprecedented precision, efficiency, and innovation. Drug development is a complex process, which encompasses numerous critical stages such as target identification, drug discovery, preclinical development, clinical trials, drug approval, and post-market monitoring. Traditional drug discovery is an expensive, lengthy, and complex process with the low success rate.^{1, 2} It is estimated that the average cost of developing a novel therapeutic agent has risen to about US $2.6 billion, typically with a development timeline of 12 to 15 years.³ Moreover, the overall success rate of novel drugs remains below 10%, even at the stage of clinical trials. By leveraging advanced computational techniques, including machine learning (ML), deep learning (DL), and predictive analytics, AI accelerates drug discovery and development process through automated data analysis, molecular modeling, target identification, and virtual screening. The applications of AI-based methods across all stages of drug discovery and development pipeline, from target identification to post-market monitoring, are outlined in Figure 1.

Figure 1. Applications of AI in distinctive stages of drug discovery and development.

The drug development process comprises of many crucial stages such as target identification, drug discovery, preclinical development, clinical trials, drug approval by regulatory authorities, and post-market monitoring. AI-based approaches hold promise in facilitating all stages of drug development.

Conventional approaches, including whole-genome knockdown screening and affinity pull-down assays, are typically employed for target identification. However, these methods are often constrained by extended timelines, substantial labor requirements, and relatively high failure rates. AI has shown great potential in identifying disease-related molecular patterns and causal relationships by generating multi-omics data, thereby accelerating the drug target discovery.^4-8 Graph deep learning combines graph structures with DL to efficiently identify novel drug targets. Moreover, convergence of scientific and medical literature with multi-omics data into knowledge graphs help AI to predict relationship between disease pathways and genes.^9-11 Integrating biomedical large language models (LLMs) with knowledge graphs or biological network provides precise and efficient means to link biological processes, diseases, and genes.¹² For instance, the PandaOmics platform identified TRAG2- and NCK-interacting kinase as a promising target for the treatment of fibrosis through multi-omics data and biological networks analysis that culminated in the development of INS018-055 (Figure 2 and Table 1), a specific inhibitor of TRAG2- and NCK-interacting kinase.¹³

Figure 2.

Chemical structures of representative AI-discovered clinical drug candidates in distinct phases of human clinical trials.

Table 1. An Overview of Representative AI-Discovered Drugs in Clinical Trials^a.

Drug	Sponsor	Indication	Phase	NCT Identifier
PXT3003 (1)	Pharnext	Charcot-Marie-Tooth type 1A	III	NCT03023540
BTRX-335140 (2)	Neumora	Major depressive disorder	III	NCT06058013
NDI-034858 (3)	Takeda	Plaque psoriasis	III	NCT06973291
REC-2282 (4)	Recursion	Neurofibromatosis type 2	II/III	NCT05130866
HLX-0201 (5)	Healx	Male fragile X syndrome	II	NCT04823052
LAM-001 (6)	Steven Hays	Bronchiolitis obliterans syndrome	II	NCT06018766
LAM-002A (7)	OrphAI Therapeutics	Amyotrophic lateral sclerosis	II	NCT05163886
SOM0226 (8)	SOM Innovation Biotech	Transthyretin amyloidosis	II	NCT02191826
SOM3355 (9)	SOM Innovation Biotech	Chorea in Huntington's disease	II	NCT05475483
LP-100 (10)	Allarity	mCRPC	II	NCT03643107
LP-300 (11)	Lantern Pharma	Advanced lung adenocarcinoma	II	NCT05456256
REC-994 (12)	Recursion	Cerebral cavernous malformation	II	NCT05085561
BT-11 (13)	NImmune	Ulcerative colitis	II	NCT03861143
INS018-055 (14)	InSilico Medicine	Idiopathic pulmonary fibrosis	II	NCT05975983
NDI-1150-101 (15)	Nimbus	Solid tumor	I/II	NCT05128487
ERAS-007 (16)	Erasca	Advanced gastrointestinal malignancies	I/II	NCT05039177
REC-4881 (17)	Recursion	Familial adenomatous polyposis	I/II	NCT05552755
LP-184 (18)	Lantern	Solid tumor	I/II	NCT05933265
BXCL701 (19)	BioXcel Therapeutics	mCRPC	I/II	NCT03910660
RLY-4008 (20)	Elevar Therapeutics	Solid tumor	I/II	NCT04526106
SGR-1505 (21)	Schrödinger	Relapsed/refractory B-cell lymphomas	I	NCT05544019
NX-13 (22)	Landos Biopharma	Ulcerative colitis	I	NCT04862741
BXCL-501 (23)	BioXcel Therapeutics	Agitation associated with pediatric schizophrenia and bipolar disorder	I	NCT05025605
PHI-101 (24)	Seoul National University Hospital	Acute myelogenous leukemia	I	NCT04842370
BPM-31543 (25)	BPGbio	Alopecia	I	NCT01588522
RLY-1971 (26)	Hoffmann-La Roche	Metastatic solid tumor	I	NCT04252339
RLY-2608 (27)	Relay Therapeutics	Advanced solid tumor or breast cancer	I	NCT05216432

^aData collected from https://www.clinicaltrials.gov. (Accessed on October 25, 2025).

Virtual screening (VS) is a robust and promising computer-aided drug discovery (CADD) approach, which analyzes a vast chemical library to identify compounds with a high probability of interacting with a specific target. VS is broadly divided into two categories: (1) structural-based virtual screening (SBVS), which predicts the interaction between large libraries of compounds and a protein target using the 3D structure of the target along with detailed knowledge of its binding site, and (2) ligand-based virtual screening (LBVS), which is typically employed when information about the target structure is unavailable. Quantitative structure-activity relationship (QSAR), pharmacophore modeling and similarity matching are commonly employed in LBVS. AI-based ligand-receptor docking models are capable of predicting ligand spatial conformations and directly creating complex atomic coordinates through algorithms such as equivariant neural networks.^{14, 15} Furthermore, these powerful models can learn the probability density distributions of ligand-receptor distances, allowing them to form, rearrange, and refine binding poses with improved accuracy.^{16, 17} Recent DL-based ligand-receptor co-folding models, including AlphaFold3 and RosettaFold, have shown great potential for predicting complex structures directly from sequence information.^{18, 19}

AI-based models have tremendous potential in generating novel chemical architectures with optimally anticipated molecular properties, thereby expediting drug discovery and development. Traditional approaches to drug discovery such as structure-based, ligand-based and pharmacophore designs rely on existing chemical libraries, expert knowledge, and predefined rules. In contrast, AI, especially DL, enables the automated generation of new molecular entities optimized for specific biological activities.²⁰ For instance, reinforce learning (RL) is a powerful method that iteratively improves compounds design based on feedbacks from each iteration’s results in meeting specific criteria such as stability, binding affinity, and pose.^{21, 22} Generative models such as variational autoencoders (VAEs) and generative adversarial networks (GANs) are capable of generating new compounds based on existing data.²³ Although VAEs and GANs are valuable methods in creating novel chemical space, these are accompanied with several shortcomings, including that newly generated compounds may not be synthetically viable, stable or biologically active.

Absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties play a pivotal role in determining the safety and efficacy of a drug candidate. Inferior ADMET profiles are responsible for the failure of many clinical trials. Early-stage ADMET assessment could help alleviate the risk of failure due to poor properties, thereby enhancing the likelihood of successful drug development.²⁴ Traditional wet-lab methods for ADMET evaluation have several shortcomings such as extended timelines and high costs, which impede drug development. To circumvent these limitations and accelerate the process, AI holds great potential for predicting ADMET properties employing descriptors or molecular fingerprints. The in silico ADMET platform at Bayer employs ML tools, including random forest (RF) and support vector machines (SVM) which utilize descriptors such as extended connectivity fingerprints.²⁵ DL-driven ADMET prediction methods allow automatically extracting feature representations from simple input data. It is reported that deep neural networks (DNNs) demonstrate either equal or a little higher effectiveness than the traditional ML models when trained on ADMET datasets.^26-28 In addition, GNNs of DL model have emerged as the powerful tools for predicting ADMET properties. For instance, GeoGNN model has shown super performance in prediction of ADMET using geometric information.²⁹

AI is also transforming the landscape of clinical trials by comprehensively analyzing patient data such as clinical history, genetic profiles, and lifestyle factors. Analysis of such data using AI tools may lead to discovery of biomarkers and characteristics of patients, thereby rendering clinical trials more informative and efficient. AI can increase the success of clinical trials and expedite the translation of drugs into medical use by optimizing treatment regimens, patient selection, and result measurements.

AI models may facilitate the discovery of diagnostic biomarkers, providing predictive insights and valuable guidance to support clinical diagnosis.³⁰ For instance, the ‘nuclei.io’ digital pathology framework³¹ helps pathologists in diagnosing by integrating AI, active learning, and real-time feedback, significantly enhancing the efficiency and accuracy of diagnostics. Further, AI has demonstrated its potential in identifying prognostic biomarkers, opening avenues for personalized and targeted therapies. For example, the DL models can describe the morphology of CD8⁺T in blood samples as effective prognostic indicators of sepsis.³² Moreover, the DL models hold potential to identify proteomic biomarkers, predicting liver disease consequences accurately.³³

AI-based approaches play an instrumental role in refining therapeutic windows, improving the safety profiles of drugs, and adjusting dose-response relationships, thereby addressing key pharmacometrics challenges in precision treatments. For example, a study involving 442 small molecule kinases and 2,145 adverse events (AEs) used a ML-based model to identify novel kinase-AEs associations, providing valuable insights to reduce side effects and guide the development of small molecule kinase inhibitors with improved safety profiles.³⁴ The PharmBERT model can automatically extract the critical pharmacokinetic information from prescription labels, facilitating the identification of adverse drug reactions and drug-drug interactions.³⁵

AI technologies have expedited the development of effective medications for various human diseases by repurposing existing drugs through the analysis of large-scale medical datasets. For example, AI has been effectively harnessed to repurpose existing drugs for treating coronavirus infection, underscoring its critical role in discovering new therapeutic applications for approved treatments.³⁶ Additionally, AI-driven drug repurposing can be facilitated through real-world data such as electronic health record and insurance claims. For instance, a DL-based model has been utilized on a large cohort comprising millions of individuals with coronary artery disease to find drugs and drug combinations that significantly ameliorate the disease outcomes.³⁷

The post-market monitoring involves the reporting of AEs of drugs, with the aim of ensuring and supporting their continued efficacy, safety and quality. The regulatory bodies obtain, identify, assess, and process AEs from across the world.³⁸ Consequently, these agencies encounter a significant challenge in monitoring the safety of their medicines. Given the excellent ability of AI technology to analyze large datasets, it may help regulatory agencies streamline the processing of AEs. The Center for Drug Evaluation and Research (CDER) of the FDA is utilizing AI-based approaches in handling and assessing Individual Case Safety Reports (ICSRs) submitted to the FDA Adverse Event Reporting System (FAERS).³⁹ For instance, the ML models have been developed by office of surveillance and epidemiology (OSE) of CDER to automatically classify AE reports via supervised ML and text engineering approaches.⁴⁰ In addition, the Information Visualization Platform (InfoViP) has been created, featuring AL/ML (NLP) and advanced visualization capabilities. Its objective is to analyze reported AEs by identifying duplicate ICSRs, classifying ICSRs based on label information, and visualizing the timeline of clinical events.⁴¹ To date, a significant number of AI-discovered drugs are currently being evaluated in distinct phases of human clinical trials, highlighting the growing importance of AI in drug discovery and development. The chemical structures of some representative drugs with their development facilitated by AI technologies are shown in Figure 2, and an overview of their clinical trials is provided in Table 1. PXT3003 (1) was discovered by AI-assisted repurposing of three already approved drugs: baclofen, a muscle relaxant; naltrexone, a therapy for opioid dependence; and sorbitol, a glucose-lowering agent utilized as a laxative.⁴² PXT3003 (1) is being evaluated in phase III to assess its long-term tolerability and safety in patients with Charcot-Marie-Tooth type 1A (NCT03023540). BTRX-335140 (2), a.k.a. navacaprant, is a selective small molecule κ opioid receptor (KOR) antagonist (IC₅₀ = 0.8 nM).⁴³ It was discovered by BlackThorn Therapeutics in a collaboration with the Scripps Research Institute. The company has employed its computational psychiatry platform, which integrates bioinformatics, neuroinformatics, and AI/ML technologies, to determine each patient's distinct neuroprint, thereby enabling the selection of optimal, on-target treatment options for biologically defined patient subtypes.⁴⁴ BTRX-335140 (2) is under study in phase III to examine its effects against placebo in patients with major depressive disorder (NCT06058013). NDI-034858 (3), a.k.a. zasocitinib or TAK-279, is a potent and highly selective tyrosine kinase 2 (TYK2) inhibitor, displaying remarkable selectivity for the human TYK2 JH2 (K_d = 0.0038 nM) over TYK1 JH2 (K_d = 4,975 nM) and TYK3 JH2 (K_d = 23,000 nM).⁴⁵ An AI-assisted SBVS strategy has been utilized to design molecules for enhanced selectivity and efficacy, ultimately leading to the identification of NDI-034858 (3).⁴⁶ It is under investigation in phase III to test its efficacy against deucravacitinib in adult patients with plaque psoriasis (NCT06973291). REC-2282 (4) is pan-HDAC inhibitor, discovered by Recursion Pharmaceuticals using its AI/ML platform.⁴⁷ It is being evaluated in phase II/III to determine its safety and efficacy in patients with progressive neurofibromatosis type 2 (NCT05130866). HLX-0201 (5), previously known as sulindac, is an approved nonsteroidal anti-inflammatory drug (NSAID). HLX-0201 (5) has been identified through the repurposing of sulindac using a novel omic-based drug-matching approach developed by Healx. This method holds potential for discovering new therapeutic connections and disease pathways by comparing the gene expression profile of a disease with gene expression profiles from Healx’s curated drug database.⁴⁸ It is being evaluated in phase II to assess its safety, efficacy, and tolerability in patients with fragile X syndrome (NCT04823052). LAM-001 (6), a.k.a. sirolimus or rapamycin, was first approved in 1999 for the prevention of kidney transplant rejection. AI Therapeutics has repurposed sirolimus or rapamycin utilizing its AI-platform, which leverages DL tools, to identify its new therapeutic use for the treatment of pulmonary arterial hypertension (PAH).⁴⁹ LAM-001 (6) is designed for direct supply to the lungs, thereby mitigating systematic exposure and minimizing the toxicities typically associated with oral sirolimus. LAM-001 (6) is under investigation in phase II to test its safety and efficacy in patients with bronchiolitis obliterans syndrome (NCT06018766). AI Therapeutics has identified LAM-002A (7) as the potential treatment for amyotrophic lateral sclerosis (ALS) by the repurposing of apilimod using its AI-platform, Guardian Angel.⁵⁰ LAM-002A (7) is in phase II studies to evaluate its safety, efficacy, and tolerability in patients with C9ORF72-associted ALS (NCT05163886). SOM0226 (8) has been identified by SOM Biotech through the repurposing of an approved drug (tolcapone). Using its AI-based VS platform, the company has discovered a new therapeutic application of tolcapone as a promising treatment for the transthyretin amyloidosis.⁵¹ SOM0226 (8) is being tested in phase II to assess its efficacy for the treatment of transthyretin amyloidosis (NCT02191826). SOM3355 (9), a.k.a. bevantolol, is a potent vesicular monoamine transporter type 2 (VMAT2) inhibitor that has been used for the treatment of hypertension for years. Using its AI-based SOM^AIPRO technology, SOM Biotech has repurposed bevantolol and identified a new clinical use as a potential therapy for chorea in Huntington’s disease (HD).⁵² The phase II studies of SOM3355 (9) are being conducted to test its safety and efficacy of two doses in patients with chorea in HD (NCT05475483). LP-100 (10), a.k.a. irofulven, has been discovered by Lantern Pharma using its AL/ML-based platform, RADR, and the NCI’s CellMiner cross database.⁵³ Through this approach, the company has uncovered biological insights and potential new targets indications for irofulven. LP-100 (10) is being evaluated in a phase II study to assess its anti-tumor efficacy in combination with prednisolone in patients with metastatic castration-resistant prostate cancers (mCRPC) who have been previously treated with androgen receptor-targeted therapy, and docetaxel (NCT03643107). LP-300 (11) is a small molecule developed using the AI-driven platform, RADR, designed for the treatment of patients suffering from non-small cell lung cancer (NSCLC) who have never smoked, and their cancer has grown following treatment with tyrosine kinase inhibitors (TKIs).⁵⁴ The RADR is assisting the HARMONIC^™ trial, evaluating LP-300 (11) in combination with pemetrexed and carboplatin in non-smokers with advanced lung adenocarcinoma for whom treatment with TKIs is ineffective (NCT05456256).⁵⁴ REC-994 (12) is a small molecule, for the treatment of cerebral cavernous malformation (CCM), discovered employing Recursion’s AI-based drug discovery platform, which leverages basic ML tools.⁵⁵ A phase II study is being conducted to assess the safety, efficacy, and pharmacokinetics (PK) of REC-994 (12) in patients with symptomatic CCM, compared to placebo (NCT05085561). BT-11 (13), a.k.a. omilancor, is a first-in-class, orally available, gut-restricted activator of lanthionine synthetase C-like 2 (LANCL2) with therapeutic potential for treating inflammatory bowel disease (IBD).⁵⁶ It was discovered using Landos Biopharma’s advanced modeling and AI-based LANCE precision medicine platform, which helps in identifying immunometabolic targets and designing new molecules that modulate these conserved pathways, delivering therapeutic effects.⁵⁷ BT-11 (13) is under investigation in phase II to examine its safety and efficacy in patients with ulcerative colitis (NCT03861143). INS018-055 (Rentosertib) (14) is a first-in-class, small molecule TNIK (TRAF2 and NCK-interacting protein kinase) inhibitor for the treatment of idiopathic pulmonary fibrosis (IPF), identified by generative AI.⁵⁸ It is being investigated in phase II to test its safety and efficacy in patients with IPF compared to placebo (NCT05975983). NDI-1150-101 or NDI-101150 (15) is a potent and selective inhibitor of hematopoietic progenitor cell kinase 1 (HPK1). It was discovered by Nimbus Therapeutics using computational tools integrated with ML-based predictive modeling approaches.⁵⁹ A phase I/II study is being carried out to evaluate the safety, efficacy, PK and pharmacodynamic (PD) of NDI-101150 as monotherapy or in combination with pembrolizumab in patients with advanced solid tumors (NCT05128487). ERAS-007 (16), a.k.a. ASN007, is a potent and selective inhibitor of extracellular signal-regulated kinases ERK1 and ERK2 (ERK1/2), with an IC₅₀ value of 2 nM.⁶⁰ It was discovered by Erasca, a biotechnology company that employes AI-based platforms to facilitate drug discovery and development. ERAS-007 (16) is being investigated in phase I/II to examine its safety and tolerability at higher doses in combination with other cancer treatments in patients with advanced gastrointestinal malignancies (NCT05039177). REC-4881 (17) is a non-ATP-competitive small molecule mitogen-activated protein kinase 1 and 2 (MEK1 and MEK2) inhibitor, being developed for the treatment of familial adenomatous polyposis (FAP). It was identified by Recursion utilizing its AI/ML-based drug discovery platform, the Recursion Operating System, which maps relationship between biological targets and small molecules.⁶¹ REC-4881 (17) is being evaluated in phase I/II to test its safety, efficacy, PK and PD in patients with FAP (NCT05552755). LP-184 (18) is a synthetic small molecule, belonging to a class of naturally occurring anti-cancer compounds. It exerts its therapeutic effects by preferentially destroying DNA in cancer cells. Lantern is using its AI platform, RADR, to accelerate the development of LP-184 (18) by discovering its mechanisms of actions for the treatment of challenging cancers and providing valuable insights into specific populations of patients.⁶² LP-184 (18) is being examined in phase I//II to assess its safety, efficacy, and maximum tolerable dose in patients suffering from advanced solid tumors (NCT05933265). BXCL-701 (19) is a non-selective inhibitor targeting dipeptidyl peptidase 8 and 9 (DPP8/9) as well as fibroblast activation protein (FAP). It was earlier investigated in phase II and III clinical trials; however, it failed to demonstrate sufficient efficacy in its intended therapeutic indication. BXCL701 (19) has been repurposed by BioXcel Therapeutics using its AI-based platform. Through this approach, the company has identified that BXCL701 (19) may act synergistically with checkpoint inhibitors, potentially providing therapeutic advantages for patients who don’t respond to checkpoint inhibitor monotherapy.⁶³ A phase II study of BXCL701 (19) is being conducted to assess its safety and efficacy as monotherapy and in combination with pembrolizumab in patients with mCRPC either small cell neuroendocrine prostate cancer or adenocarcinoma phenotype (NCT03910660). RLY-4008 (20), a.k.a. lirafugratinib, is a highly selective small molecule fibroblast growth factor receptor 2 (FGFR2) inhibitor.⁶⁴ It was discovered by Relay Therapeutics employing its AI-based platform.⁶⁵ RLY-4008 (20) is being evaluated in phase I/II to assess its safety, tolerability, PK, PD, and antineoplastic activity in patients with metastatic cholangiocarcinoma (CCA) and other solid tumors (NCT04526106). SGR-1505 (21) is a potent and orally available small molecule inhibitor of mucosa-associated lymphoid tissue lymphoma translocation protein 1 (MALT1). It was identified by the Schrödinger using integrated advanced physics-based modeling methods, free energy calculations with ML tools and chemistry-aware compound enumeration workflow.⁶⁶ SGR-1505 (21) is under study in phase I to assess its safety and tolerability in patients with relapsed/refractory B-cell lymphomas. NX-13 (22) is an orally active, gut-restricted, novel small molecule inhibitor of nucleotide-binding oligomerization domain, leucine rich repeat containing X1 (NLRX1) for the treatment of ulcerative colitis.⁶⁷ Landos Biopharma identified NLRX1 using its advanced AI-based integrated computational and experimental precision medicine platform, led to the discovery of NX-13 (22).⁶⁸ A phase I study of NX-13 (22) is being carried out to assess its safety, tolerability, and PD in patients with active ulcerative colitis (NCT04862741). BXCL-501 (23) was discovered through the repurposing of dexmedetomidine, an approved drug, by BioXcel Therapeutics using AI-based technologies.⁶⁹ BXCL-501 (23) is being evaluated in phase I to assess its safety and efficacy in treating agitation associated with pediatric schizophrenia and bipolar disorder (NCT05025605). PHI-101 (24) is an orally available, next-generation FLT3 tyrosine kinase inhibitor, with a novel mechanism of action designed to circumvent resistance to acute myeloid leukemia (AML).⁷⁰ It was discovered and developed by the Pharos iBio, using its AI-based platform, Chemiverse, which is capable to deal with all aspects of the drug discovery process, from novel drug target identification to clinical candidate selection.⁷¹ PHI-101 (24) is being investigated in phase I to examine its safety, tolerability, PK, and PD in patients with relapsed or refractory AML (NCT04842370). BPM-31543 (25) was identified by BERG, a biopharmaceutical company, using its AI-powered, Interrogative Biology^® platform for the treatment of chemotherapy-induced alopecia (CIA).⁷² It is understudy in phase I to assess its safety in patients with CIA (NCT01588522). RLY-1971 (26) is a small molecule protein tyrosine phosphatase SHP2 inhibitor, being developed for the treatment of solid tumors.⁷³ It was discovered by Relay Therapeutics using its AI-driven platform, Dynamo, which analyzes protein dynamic to identify disease-relevant conformational states.⁷⁴ RLY-1971 (26) is under investigation in phase I to determine its safety, efficacy, tolerability, and PK in patients with advanced or metastatic solid tumors (NCT04252339). RLY-2608 (27) is first-in-class, mutant-selective, allosteric inhibitor of PI3Kα, that dissociates antitumor activity from hyperinsulinemia.⁷⁵ It was discovered by Relay Therapeutics using its AI-driven platform and targets a novel allosteric pocket distinct from the primary active site of the protein that is preferentially favored in certain mutant forms of PI3Kα.⁷⁶ RLY-2608 (27) is being evaluated in phase I to test its safety, tolerability, PK and PD as monotherapy and in combination with endocrine therapy with or without a CDK4/6 or CDK4 inhibitor in patients with advanced solid tumors or breast cancer (NCT05216432).

Despite the immense potential of AI technologies, as exemplified by the milestone successes of the aforementioned drug candidates in human clinical trials, it is accompanied by numerous significant limitations to fully revolutionize drug discovery and development. The dearth of high-quality data required for training renders AI less effective in discovery of novel targets, biomarkers, and other related applications. In addition, the presence of incomplete information, biases, and errors in available data or data with potential ethical concerns further attenuate the reliability of AI. The lack of published negative data, including failed experiments and trials with unfavorable results, impedes the insights into efficacy, drug–target–disease relationships, and broader clinical characteristics.^{77, 78} Current generative approaches hold considerable promise for generating novel chemical architectures; many of these structures are likely to be synthetically unfeasible or unstable. While molecular generation methods complemented by detailed reaction knowledge have demonstrated significant potential, they still require further improvement.^{79, 80} The traditionally undruggable targets such as transcription factors or scaffolding proteins with flexible, dynamic structures or lacking well-defined binding sites pose significant challenges for AI-based drug discovery methods. These hurdles may be addressed by integrating AI tools with high-content screening to search the conformational space and discover suitable binding sites. In addition, many AI platforms or systems, often referred to as "black boxes", lack transparency in their decision-making process, raising a potential barrier to regulatory and clinical trust. Moreover, the FDA is still developing clear guidelines for AI-driven drug development, creating uncertainty for AI-assisted pharmaceuticals. The unique characteristics of AI continue to pose challenges for both drug developers and regulators.

While AI is facilitating drug discovery and development process, future efforts are required to address existing challenges to fully realize the potential of AI approaches and become standard practice. Viable strategies should be devised to determine data standards, increase data sharing, and generate new AI algorithms capable of making accurate predictions from limited data. Multimodal utilizing chemical and textual information has shown potential in solving data lacking concern.⁸¹ AI can uncover the therapeutic potential of existing drugs for treating orphan diseases through repurposing by integrating data, including protein-protein interactions, multiomics profiles, disease-specific molecular pathways, and clinical records.⁸² Multimodal such as DL-based classification of drugs, can identify underlying mechanism of actions, predict efficacy, and evaluate toxicity, demonstrating the future potential of AI-based approaches in drug development.^{83, 84} AI-based models are anticipated to change the landscape of medical modeling and simulation. AI models will enable the generation of comprehensive virtual human simulations,⁸⁵ providing valuable insights into disease mechanisms,⁸⁶ drug actions, and biological diversity among individuals.⁸⁷ AI can improve clinical trial design by simulating multiple scenarios to optimize selection criteria,⁸⁸ expedite patient recruitment, and ensure more diverse, representative populations.⁸⁹

In conclusion, AI holds immense potential in addressing the challenges associated with traditional drug discovery and development by accelerating timelines, reducing costs, and increasing the probability of success. Advanced AI methods are capable of analyzing large datasets for novel target identification, augmenting the efficiency of VS, performing lead optimization, and facilitating preclinical development. Moreover, AI plays a pivotal role in early clinical trials by assisting in patient recruitment and predicting outcomes, thereby attenuating the likelihood of trial failure. Further, AI can distinguish between simple prognostic biomarkers and those predictive responses to medication, thereby advancing the development of personalized treatments. The power of AI can be harnessed, if applied judiciously, across all stages of the drug discovery and development pipeline, ultimately enabling the discovery of novel and precision medicines for treatment of various human diseases.