Artificial Intelligence in Drug Discovery: Applications, Challenges, and Future Perspectives

Drug discovery is one of the most expensive, time consuming and complex scientific activities in contemporary biomedical research. An average of over 10–15 years and investment exceeding USD 2.5 billion is required for a single new therapeutic agent to reach the market (Sarkar et al., 2024) [6]. Despite these impressive investments, the failure rate of drug candidates is still terribly high at less than 12% of compounds that enter clinical development gaining regulatory approval [4] (Blanco?González et al., 2023). Common causes for failure are insufficient target choice, suboptimal pharmacokinetic or toxicological properties and poor predictive value of preclinical models [3] (Askr et al., 2023).

Artificial Intelligence (AI) has been playing an increasingly more important role in pharmaceutical R&D and is being heralded as a transformational paradigm in this landscape. Utilizing computational algorithms, which can handle highly complex and nonlinear associations in large heterogeneous datasets to achieve predictive modeling, pattern recognition and decision-making at scales that traditional methods cannot [2] (Arabi, 2021). In this space, the emergence of Machine Learning (ML) and Deep Learning (DL) approaches has introduced such tools as critical for data-driven decision-making across the drug discovery process, including target identification, lead optimization, clinical trial design and post-marketing surveillance [3]  (Askr et al., 2023).

Drug discovery is a multistep process where all stages become interdependent, including identification of the molecular target; initial lead compound synthesis or discovery; optimization through improving both pharmacodynamic and pharmacokinetic parameters, as well as preclinical and clinical approval [4] (Blanco-González et al., 2023). Together these stages produce enormous quantities of biochemical, genetic, and clinical data - big biomedical data. The exponential increase in these types of data due to the advances in high-throughput screening, next-generation sequencing (NGS) and omics technologies has posed both an unprecedented opportunity as well as a challenge. Conventional statistical and mechanistic models are not designed for the high volume, velocity, and variety of these data streams [7] (Hughes et al., 2022). As a result, it has become imperative for AI to decode multi-dimensional datasets and discover novel patterns that explain drug efficacy and safety.

ROLE OF AI IN PHARMACEUTICAL R&D:

The adoption of artificial intelligence by the pharmaceutical industry is not just a technical revolution but a transformation in the way we think about great scientific challenges. Old drug development pathways were a fairly linear sequence from target discovery, hit finding, lead optimization and preclinical experimentation to clinical trials with large turnaround times between phases and relatively low success rates at each phase. With AI, this paradigm has been converted into a loop cycle and adaptive design–make-test–analyze (DMTA) strategy [6] (Sarkar et al., 2024).

Pre-discovery stage AI systems that learn from multi-omics (genome, transcriptome, proteome and metabolome) data could enable identification of disease-associated targets by revealing obscured relationships between molecular changes and phenotypic readouts [5] (Udegbe et al., 2024). Graph neural networks (GNNs) and probabilistic graphical models are among the most successful techniques for modeling biological networks and predicting new target--pathway associations. In medicinal chemistry, deep generative models like VAEs can provide the tool for designing new molecules with optimised multiple properties such as receptor affinity, solubility and synthetic accessibility [3,4,7] (Askr et al., 2023; Blanco-González et al., 2023; Hughes et al., 2022).

Further than compound design, AI enhances experimental pipeline state-ofthe-art with predictive toxicology, in-silico ADMET modeling and virtual screening by which researchers triaged massive compound libraries, steering the laboratory resources towards the most promising candidates [6] (Sarkar et al., 2024). For instance, models developed based on ensemble machine-learning algorithms could incorporate physicochemical descriptors, molecular fingerprints and docking score that is able to yield overall accuracy rate over90% in toxicity prediction using Tox21and ToxCast benchmark datasets [2] (Arabi 2021).

The scope of AI is not limited to compound generation, but it gains importance in process chemistry and synthesis planning where deep reinforcement-learning (DRL) algorithms predict synthetic routes with optimal performance by simulating retrosynthesis way based on millions of historical reactions simulated [3] (Askr et al., 2023). IBM’s RXN system is such an example, where synthetic routes of new compounds are automatically proposed in a data-driven manner and re-adjusted on-the-fly with the arrival of experimental results. These models allow a significant reduction in human labor for reaction route optimization and enable faster bench-to-prototype translation.

In the clinical space, AI helps to drive trial design and patient stratification as well as safety monitoring. Models of prediction based on electronic health records (EHRs) and real-world evidence (RWE) help to identify the subgroups of patients who are most likely to respond to a given treatment, leading to increased trial efficiency and lower drop-out rates [5] (Udegbe et al., 2024). Furthermore, NLP pipelines can interrogate unstructured clinical reports, AE databases and published work to identify early safety signals, informing pharmacovigilance and post-marketing surveillance [4] (Blanco-González et al., 2023).

ARTIFICIAL INTELLIGENCE: CONCEPTS AND TECHNOLOGIES

Artificial intelligence (AI) is a general term that refers to computerized systems designed to be able to carry out activities typically requiring human intelligence, for example learning, reasoning, perceiving and decision making. In the biomedical field, AI specifically describes the algorithmic machinery that manages complex biological data, recognizing patterns and producing predictive models with which to guide experimental or clinical choices. Unlike typical programming, following prespecified rules, AI solutions are developed by learning from data—adjusting their parameters to maximize a performance metric for the corresponding problem [3] (Askr et al., 2023).

In the field of pharmaceuticals, the real strength of AI is in integrating all these data modalities—chemical, biological, clinical and text based on one analytical framework. Unique datasets are inherent to drug discovery, such as: molecular structures, assay readouts, gene expression profiles and/or patient responses. Traditional statistical methods are not suitable to model the nonlinear relationships between these multi-scale variables; while AI models, including but not limited to machine learning (ML) and deep learning (DL) architectures can be used as a way to bridge this gap of complexity by means of learning on an iterative basis [4] (Blanco-González et al., 2023).

The application of AI in drug discovery spans the continuum from predictive to prescriptive:

1. Predictive models predict the activity, binding affinities or toxicity of a molecule.
2. Prescriptive models, including reinforcement learning agents, recommend next experimental actions and thus drive automated design–make–test cycles [4] (Sarkar et al., 2024).

A simplified schematization of the conceptual pillars supporting AI for biomedicine discovery is given in Figure 2, and illustrates how acquisition of data, training model and hypothesis generation interact in a continuous loop of learning and feedback.

GENERATIVE AI MODELS: FROM DE NOVO DESIGN TO PROBE OPTIMIZATION:

One of these disruptive innovations in AI-aided drug discovery is the entry into force of generative artificial intelligence (GenAI) algorithms that generate new data instances themselves that closely resemble their training sets [3] (Askr et al., prediction year, 2023). In chemistry this entails creating new molecular objects which, while of arbitrary structure, are nevertheless consistent with various pharmacological and physiochemical criteria without human guidance. The disconcerted approach what was considered, until recently, as the ‘state of art’ for de novo design of screening existing libraries is turned to the design of compounds tailored in a best-fit fashion to match the target profile.

Generative model families include:

1. Variational Autoencoders (VAEs): Project molecular descriptors to a latent space, sample new points and generate molecules with similar properties [68] (Sarkar et al., 2024).
2. Generative Adversarial Networks (GANs): Use a generator–discriminator system to continuously optimize chemical and biological compliance of the molecular designs [4] (Blanco-González et al., 2023).
3. RL-Guided Generators: Incorporate generative modeling with goal-orientated optimization to optimize predefined reward functions such as drug-likeness, Lipinski’s rule or binding affinity maximization [5] (Udegbe et al., 2024).
4. Diffusion Models: A class of aforementioned generative methods that can iteratively “denoise” random data to obtain valid molecular graphs, resulting in the most recent state-of-the-art 3D structure generation model [7] (Hughes et al., 2022).

Industrial uses of generative AI are already palpable. Insilico Medicine’s Chemistry42 platform leveraged a hybrid GAN–RL pipeline to predict inhibitors of fibrosis-related enzymes in several weeks, validating predictions with experimental validation in preclinical assays. In the same way, Ex Scientia’s AI-led discovery engine developed an oncology drug candidate (DSP-1181) that progressed to clinics fastest ever [4] (Blanco-González et a.l., 2023). These examples serve to highlight the generative prowess and potential scalability of such models in molecular innovation.

APPLICATION OF AI IN DRUG DISCOVERY PIPELINE:

Overview:

Drug discovery constitutes a series of activities from target identification and validation to molecular design, synthesis and preclinical optimization, preclinical and clinical testing phases, all the way through postmarketing surveillance. At every stage there are large, multi-dimensional non-linear interdependent datasets involved, making this an ideal space for AI approaches to not only speed up discovery but also increase the predictivity [4] (Blanco-González et al., 2023).

Today AI-driven methods are used in virtually every part of this pipeline, and it is now more like a stochastic learning loop rather than a linear hypothesis-driven process. The feedback loops in this unified framework iteratively refine hypotheses according to real-time experimental results and inferred predictive analytics, resulting in a closed-loop design–make–test–analyze (DMTA) system.

De Novo Molecular Design and Lead Optimization:

The generative abilities of AI have revolutionized de novo molecular design, enabling chemists to survey chemical spaces that are remote from those catalogued in combinatorial libraries. Generative Model autolearning organ sectional Learning environments are a part of this transformation [3] (Askr et al., 2023).

These models are trained on molecular data sets (e.g., chEMBL, ZINC and PubChem) to have learnt about the underlying distribution of “drug-like” chemical space from which they can propose novel compounds that optimize desirable characteristics—like potency, solubility and synthetic feasibility [3] (Sarkar et al., 2024).

One of the most recent industrial advances was made by Insilico Medicine, which used a GAN–RL hybrid model to discover inhibitors for enzymes related to fibrosis. Within 46 days the platform developed, synthesized, and experimentally validated a pre-clinical candidate for idiopathic pulmonary fibrosis – demonstrating the capability of AI in condensing multi-year timelines to weeks [5] (Udegbe et al., 2024).

There is a growing interest in the application of MOO methods in lead optimization. For example, Molecule. and Exscientia’s CentaurChemist™ employs AI systems to optimize the lead compounds based on feedback from the bioassay data, delicately balancing pharmacodynamic potency and ADMET properties [4] (Blanco-González et al., 2023).

 Virtual Screening and QSAR Modeling:

Virtual screening (VS) is a mainstay tool of contemporary drug discovery, which allows the estimation of millions of potential compounds in silico before investing time and effort on physical testing. Classical docking-based VS approaches are computationally expensive and require accurate 3D structures. These limitations are mitigated by AI-based approaches, particularly deep QSAR (Quantitative Structure–Activity Relationship) models that learn from data to predict the bioactivity of molecules [3] (Sarkar et al., 2024).

DeepChem, Mol2Vec and ChemProp are some of the most prominent libraries which use deep learning to predict bioactivity by directly inferring from molecular descriptors or SMILES strings. AI-based QSAR surpasses generalizations of traditional QSAR (PLS and MLR) methods even over novelty scaffolds [2] (Arabi, 2021).

Finally, a hybrid approach of molecular docking with DL-based QSAR (AI-assisted virtual screening, AIVS) is gaining momentum. For instance, Atomnet and Atomwise platforms use 3D CNNs to predict binding affinity of thousands of small molecules, with reported identification of several new hit compounds for oncology and infectious diseases [4] (Blanco-González et al., 2023).

The number of physical experiments that need to be performed is significantly reduced with the use of AI-based VS, resulting in increased hit rates and cost effectiveness. It democratizes the process of drug discovery, making it possible for smaller research groups to compete with big pharmaceutical companies using cloud computing as a computational environment [5] (Udegbe et al., 2024).

Toxicity, ADMET, and Safety Assessment:

The ability of AI to predict toxicity has also been significantly enhanced via multi-task learning in which representations of related endpoints are learnt jointly. For instance, DeepTox outperformed other methods on the NIH Tox21 Challenge by predicting 12 toxicity pathways simultaneously using an integrated deep neural network [4] (Blanco-González et al., 2023).

Artificial intelligence also moves into silico pharmacokinetic modeling to support, interactive with this physiologically based pharmacokinetics (PBPK) simulation. Models such as DeepADMET and Chemprop-PK predict oral and plasma exposure as well as clearance directly from molecular graphs [5] (Udegbe, et al., 2024)

Explainable AI for toxicology is another major development. Feature attribution methods, including SHAP and LIME, have been used to bring out substructures that contribute to adverse effects – in some cases identifying nitroaromatic or aniline groups often associated with mutagenicity [7] (Hughes et al., 2022). This transparency promotes not just trust, but also allows for rational molecular redesign to reduce risk.

Preclinical AI predicts off-target interactions as well as metabolic liabilities by analyzing enzyme–substrate relationships through deeper graph networks. Combining with omics-based toxicity biomarkers and organ-on-chip data may provide a more biologically accurate prediction that overlaps towards the regulatory acceptance framework of computational toxicology.

.Clinical trial optimization and patient stratification:

Clinical development is the costliest period in the drug pipeline and has success rates typically below 10% and average costs of more than USD 2 billion/drug approved [4] (Blanco-González et al., 2023). AI’s influence here is revolutionary, especially in terms of trial design, recruitment and monitoring, as well adaptive management.

Predictive analytics, powered by AI, simplify patient recruitment by automatically searching electronic health records (EHRs), genomics records and real-world evidence (RWE) for potential candidates. Unstructured text and structured field can be processed using machine learning algorithms to match eligibility criteria in real time [3] (Askr et al., 2023). This is a need that platforms such as Deep6AI and TriNetX are equipped to fill, potentially cutting recruitment times by half.

AI also aids in patient stratification using clustering and supervised learning on multi-omics data, which result in identification of subpopulations with varied responses to therapies. This form of precision medicine, wherein AI refers to models which predict the sets of patients that are most likely to derive benefits from given interventions [7] (Hughes et al., 2022).

In more advanced clinical environments, digital twin models—computer-based proxies of individual patients constructed from physiological and molecular data—facilitate in silico trial tests. These models predict treatment responses for different conditions and could eventually replace some of the animal and human studies in the future [4] (Blanco-González etal, 2023).

Table 1. Representative AI Applications in Drug Discovery.