AI in drug discovery

I. Research Background and Overall Development Trends

Artificial intelligence (AI) is transforming traditional drug discovery, particularly in the areas of drug target identification, candidate compound screening, and novel molecular design. Based on the three research papers, the development of AI in this field has gradually progressed from early computer-aided drug design to deep learning-based prediction of protein–ligand interactions, and further toward the use of generative AI to design entirely new drug candidates.

The three papers illustrate different aspects of this development.

Paper 1: AI Accelerates the Identification of Druggable Targets Through 3D Structures of Proteins and Compounds

npj Precision Oncology, 2026

This review examines how AI integrates multi-omics data, predicts three-dimensional protein structures, identifies potential drug targets, and accelerates cancer drug discovery. It primarily summarizes existing methods, technologies, and research cases.

Paper 2: AI-Guided De Novo Design of Novel S1PR1-Targeting Ligands

Scientific Reports, 2026

This study uses siponimod as a molecular scaffold to generate new compounds through AI. The resulting candidates are evaluated through molecular docking, pharmacokinetic property prediction, and molecular dynamics simulations.

Paper 3: AI Is a Viable Alternative to High-Throughput Screening: A 318-Target Study

Scientific Reports, 2024

This study evaluates Atomwise's AIMS program across 318 research projects to investigate whether the AtomNet deep learning model can identify biologically active compounds across a wide range of protein targets.

Together, these papers reveal an important trend: AI is no longer limited to analyzing existing drugs. It is increasingly involved in the entire early-stage drug discovery process, from identifying promising research targets and generating candidate molecules to predicting molecular properties and prioritizing experimental testing.

However, an important distinction must be made: predicting that a molecule may be effective in a computer model is fundamentally different from demonstrating its effectiveness through laboratory experiments and clinical trials.

II. Why Does Traditional Drug Discovery Need Artificial Intelligence?

2.1 Limitations of Traditional Drug Discovery

Traditional small-molecule drug discovery generally involves the following steps:

Identifying a biological target: Finding proteins, genes, or biological pathways associated with a disease.

Screening candidate compounds: Searching for molecules that may interact with the target through experimental screening or computational methods.

Optimizing lead compounds: Improving biological activity, selectivity, stability, and pharmacokinetic properties.

Conducting preclinical and clinical studies: Evaluating safety, therapeutic effectiveness, and risks associated with human use.

Paper 1 notes that traditional drug development may take approximately 12–15 years and cost around $2.8 billion. This figure represents an industry estimate cited by the review, rather than the actual development cost of every drug.

Traditional high-throughput screening (HTS) uses automated equipment to test large numbers of existing compounds experimentally. However, it has an important limitation: a compound generally needs to exist or be synthesized before it can undergo physical testing.

Even when experiments can process large numbers of molecules simultaneously, researchers remain constrained by compound availability, laboratory equipment, protein preparation, reagent costs, and testing time.

Paper 3 explains that many chemically possible compounds available for on-demand synthesis have never actually been manufactured. The potential chemical space may contain trillions of molecules, making comprehensive experimental screening extremely difficult.

2.2 The Fundamental Change Introduced by AI

AI changes the order of drug discovery. Researchers can evaluate compounds computationally before synthesizing them and then select the most promising candidates for experimental testing.

AI has not eliminated the need for laboratories. Instead, it allows researchers to make more informed decisions about which compounds deserve the time and resources required for experimental testing.

III. Three Major Directions of AI Development in Drug Discovery

Based on the three papers, AI applications can be divided into three closely connected directions.

3.1 Direction One: Using AI to Identify New Drug Targets

Drug targets are often proteins, genes, or biological pathways. The first step in drug discovery is not necessarily to manufacture a drug, but to determine which biological process should be targeted.

This challenge is particularly complicated in cancer research. Tumors from different patients may have different genetic mutations, protein expression levels, and mechanisms of drug resistance. Consequently, relying on a single gene or experimental result may not fully explain the underlying disease mechanisms.

Paper 1 introduces several AI-based approaches to target identification.

Multi-omics data analysis: Integrating genomic, transcriptomic, proteomic, and epigenetic data to identify genes and proteins associated with diseases.

Protein structure prediction: Using models such as AlphaFold2, AlphaFold3, and ESMFold to predict three-dimensional protein structures and provide structural information for identifying potential binding sites.

Druggability prediction: Evaluating whether a protein could potentially be targeted effectively by small-molecule drugs.

Biological network analysis: Using graph neural networks and knowledge graphs to analyze relationships among genes, proteins, drugs, and diseases.

An important example is PandaOmics, a platform developed by Insilico Medicine.

PandaOmics combines multi-omics data with scientific literature, existing biological knowledge, and disease-related information to analyze and rank potential therapeutic targets.

According to Paper 1, the platform has been used to investigate potential targets across different cancers, including the following examples.

These examples demonstrate how AI can help researchers identify promising targets from large volumes of biological data.

However, a target ranked highly by AI is not necessarily an effective therapeutic target. Researchers still need to use cell experiments, animal models, and further studies to determine whether targeting these molecules can actually improve disease outcomes.

3.2 Direction Two: Using AI for Virtual Screening

Virtual screening is one of the most important technologies connecting all three papers.

The basic idea is to use computational models to predict interactions between large numbers of compounds and a target protein. Researchers can then prioritize compounds with promising predicted results for experimental testing.

Unlike traditional experimental screening, virtual screening does not require every candidate molecule to be synthesized in advance. Researchers can computationally evaluate large numbers of molecules before committing laboratory resources to a smaller selection.

Common methods include:

Convolutional neural networks (CNNs): Analyzing three-dimensional spatial features of proteins and compounds.

Graph neural networks (GNNs): Representing molecular structures as graphs composed of atoms and chemical bonds, allowing models to learn molecular features and interaction patterns.

Molecular docking: Predicting possible positions and orientations of small molecules within protein-binding pockets.

Diffusion models: Generating or predicting possible protein–ligand binding poses in certain modern structural modeling systems.

It is important to note that molecular docking is not inherently an AI technique. It is a computational chemistry method that can be combined with AI models or used independently.

These methods generally rank candidate compounds and may then be combined with additional computational tools to evaluate pharmacokinetic properties, potential toxicity, and structural stability.

3.3 Direction Three: Using Generative AI to Design New Molecules

Traditional virtual screening primarily answers the question, "Which existing or constructible compounds might be effective?"

Generative AI goes one step further by attempting to answer a different question: "Can we design a new molecule that may have the desired properties?"

Generative models learn patterns from existing molecular structures, chemical properties, and protein interactions. They can then generate new candidate structures based on the patterns they have learned.

Paper 1 discusses several types of generative models, including:

Variational autoencoders (VAEs)

Generative adversarial networks (GANs)

Deep learning models based on molecular sequences

Diffusion models

These methods can generate candidate molecules with particular structural or predicted properties. Their outputs can then be evaluated using virtual screening, molecular docking, and pharmacokinetic prediction.

For example, researchers can select a target protein and use a generative model to produce candidate molecules. The resulting compounds can be evaluated computationally for potential binding interactions, molecular stability, and drug-like properties. Only a small number of promising candidates then proceed to synthesis and experimental testing.

This creates a new workflow:

Identify a disease-related target.

Generate large numbers of candidate molecules.

Apply AI prediction, virtual screening, and molecular property evaluation.

Select candidates for actual synthesis.

Conduct experimental validation and iterative optimization.

The importance of this workflow is that AI can help researchers explore chemical space that traditional screening methods may rarely investigate, rather than merely searching for drugs among existing compounds.

IV. Case Study One: AtomNet Virtual Screening Across 318 Research Projects

Source: Paper 3, "AI Is a Viable Alternative to High Throughput Screening: A 318-Target Study" (2024).

This is one of the most valuable real-world validation cases among the three papers because it goes beyond computational predictions and includes experimental testing of selected compounds.

4.1 How Was the Study Conducted?

The researchers used Atomwise's AtomNet convolutional neural network to conduct virtual screening across different protein targets.

AtomNet analyzes three-dimensional structural information about proteins and candidate small molecules to predict which compounds may interact meaningfully with the target.

The study consisted of two groups:

Twenty-two internal drug discovery projects.

Two hundred ninety-six projects conducted in collaboration with academic laboratories.

The total was 318 projects. The academic projects covered multiple areas, including cancer, infectious diseases, and neurological disorders, and involved a variety of protein classes.

The study reported the following results.

The figures above come from the study's experimental results. The hit rate refers to the proportion of tested compounds determined to be active; it is not the probability that an experimental drug will eventually reach the market.

4.2 How Did the Researchers Validate AI Predictions?

Rather than declaring success immediately after the model generated predictions, the researchers performed experimental validation.

First, AtomNet scored and ranked compounds from large virtual chemical libraries. The research team then selected highly ranked molecules while attempting to maintain structural diversity among the candidates.

Next, selected compounds were synthesized or purchased. Techniques such as nuclear magnetic resonance (NMR) and liquid chromatography–mass spectrometry (LC–MS) were used for quality control. The compounds were subsequently tested in biochemical or cellular assays.

When preliminary testing indicated activity, researchers could conduct dose–response experiments to determine how the molecules behaved at different concentrations, rather than relying on a single measurement.

This process is essential because a computationally promising molecule may show no activity in a real experiment. It may also produce assay interference or other unexpected effects.

4.3 What Do the Results Demonstrate?

The paper reported that 215 of the 296 academic projects identified at least one compound with experimentally confirmed biochemical or cellular activity. This corresponds to a project-level success rate of approximately 73%.

Across these projects, researchers tested an average of approximately 85 compounds per project, with an average hit rate of 7.6%.

More importantly, 207 of the 296 academic projects involved targets for which the AtomNet training data contained no known active molecules for that target or proteins with greater than 70% sequence identity to it.

For this subset, the paper reported an average hit rate of 5.3%.

This suggests that the model could identify experimentally active candidate compounds in many cases even when direct target-specific activity data were unavailable.

4.4 Specific Examples: LATS1 and VCP

The paper also describes two internal research projects in greater detail.

Case A: LATS1

LATS1 is a kinase involved in the regulation of cell growth.

The research team used AtomNet to screen compounds against this target and identified active candidates despite the absence of a crystal structure for the target and known active compounds.

This case suggests that AI-based virtual screening can be useful even for targets with limited structural and activity data.

Case B: VCP

VCP is an ATPase involved in various cellular processes.

The research team identified candidate allosteric and orthosteric modulators.

Orthosteric modulators interact with the primary binding site of a protein, whereas allosteric modulators interact with other sites and influence protein activity by changing its conformation or function.

These cases demonstrate that AI can help researchers discover not only compounds resembling known drugs, but also novel chemical scaffolds and potentially different modes of protein interaction that traditional screening may struggle to identify.

4.5 Limitations of the Study

Although the results are encouraging, the 73% project-level success rate should not be interpreted as meaning that 73% of AI-discovered drugs will reach the market.

The study's success criterion was the identification of at least one biologically active compound, not the demonstration that the compound could treat patients.

The study also does not establish that AtomNet is always superior to traditional high-throughput screening across every target and experimental condition. Different screening methods may use different compound libraries and experimental conditions, so hit rates alone cannot establish that AI has completely replaced HTS.

A more accurate conclusion is that this study provides substantial experimental evidence supporting AI as a tool for early-stage drug discovery and screening.

V. Case Study Two: AI Design of Novel S1PR1-Targeting Candidate Molecules

Source: Paper 2, "In Silico AI-Guided De Novo Design of S1PR1-Targeting Ligands from a Siponimod Scaffold for Angiogenesis and Fat-Cartilage Graft Survival" (2026).

Unlike AtomNet, which primarily searches large numbers of candidate compounds for potential activity, this study takes a further step: it uses generative AI to design new molecular structures and then evaluates their potential through multiple computational methods.

It illustrates the transition from AI-assisted compound screening toward AI-assisted molecular design.

5.1 Why Did the Researchers Choose S1PR1?

S1PR1, or sphingosine-1-phosphate receptor 1, is a G protein-coupled receptor involved in vascular regulation and immune cell migration.

The specific research question was whether researchers could design candidate molecules that bind S1PR1 and potentially support vascular formation and tissue survival following fat-cartilage grafting.

If transplanted tissue does not receive sufficient blood supply during its early stages, it may not obtain adequate oxygen and nutrients, potentially affecting its long-term survival. Researchers therefore sought to investigate potential molecules that could regulate angiogenesis while maintaining vascular stability.

However, modulating S1PR1 does not automatically result in beneficial blood vessel formation. The receptor is also involved in immune regulation, and different forms of receptor activation or modulation may produce different biological effects.

5.2 How Did AI Design the Molecules?

The researchers selected siponimod, a molecule with an established pharmacological background, as the reference scaffold and used the WADDAICA platform to generate structural derivatives.

The platform combines a three-dimensional convolutional neural network-based conditional variational autoencoder with a molecular descriptor generation model based on long short-term memory networks.

In simple terms, the system attempts to learn structural features of the original molecule and modify its latent chemical representation to produce new candidate molecules.

The researchers then evaluated these candidates using several computational approaches.

Step 1: Generate new molecules

AI generated a series of candidate derivatives based on the structure of siponimod. Duplicate or chemically invalid outputs were excluded.

Step 2: Molecular docking

Two independent docking workflows, PyRx and AutoDock Vina Expanded, were used to predict candidate binding positions and interactions within the S1PR1 binding pocket.

Step 3: Predict drug properties

ADME and toxicity prediction tools were used to evaluate absorption, distribution, metabolism, excretion, and potential toxicity. Density functional theory (DFT) was also used to analyze molecular electronic properties.

Step 4: Molecular dynamics simulations

Long-timescale simulations were used to observe structural changes in the protein–ligand complexes. MM-GBSA and MM-PBSA calculations were used to estimate binding energetics.

The key point is that researchers did not evaluate the AI-generated structures in isolation. They combined generative modeling with conventional computational chemistry tools to progressively narrow the candidate pool.

5.3 Key Results: AI Derivative 1

One of the most notable candidates in the study was AI Derivative 1.

The study reported the following molecular docking results.

Both docking workflows identified the molecule as a candidate worthy of further investigation.

These values are computational scoring estimates, not experimentally measured binding free energies.

The study also reported several additional findings.

Drug-likeness: Computational predictions indicated that the molecule complied with Lipinski's Rule of Five and had favorable predicted gastrointestinal absorption.

Structural stability: During the study's 1-microsecond molecular dynamics simulation, the candidate remained accommodated within the S1PR1 binding pocket.

Electronic properties: Density functional theory analysis indicated that its electronic properties were broadly comparable to those of the reference molecule.

Estimated binding energetics: Both MM-GBSA and MM-PBSA produced favorable binding energy estimates, but both models favored siponimod overall.

The last point is particularly important. Although AI Derivative 1 achieved competitive docking scores, the results did not demonstrate that it had stronger actual binding affinity than siponimod, much less that it would produce better therapeutic outcomes.

5.4 The Real Significance of This Study

This study demonstrates how generative AI can contribute to a specific drug design problem: starting from an existing molecule, generating structurally different candidates, and using multiple computational methods to evaluate them.

However, the paper explicitly indicates that the findings remain computational evidence. The authors did not establish the candidate's actual biological activity through cell or animal experiments, nor did they demonstrate improved vascularization or survival of transplanted tissue.

Further studies would need to establish:

Whether the candidate molecule actually binds to S1PR1.

Whether it acts as an agonist, antagonist, or another type of receptor modulator.

Whether it produces the intended vascular effects under appropriate biological conditions.

Whether its toxicity, selectivity, and in vivo pharmacokinetic properties support further development.

Therefore, this case is best understood as an example of AI-assisted drug design, rather than proof that a new drug has already been successfully developed.

VI. How Do the Three Papers Collectively Demonstrate the Development of AI in Drug Discovery?

The three papers focus on target identification, generative molecular design, and large-scale virtual screening, respectively.

Their relationship can be summarized as follows.

These comparisons suggest that AI drug discovery does not rely on a single model capable of solving every problem.

Instead, practical drug discovery increasingly depends on combinations of tools: one model generates candidate molecules, another evaluates their potential binding interactions, other tools predict drug properties, and laboratory experiments ultimately validate the results.

VII. What Challenges Does AI Drug Discovery Still Face?

Although the three papers demonstrate substantial progress, they also reveal several important limitations.

7.1 Computational Predictions Do Not Equal Biological Activity

This is the most important limitation.

A favorable molecular docking score indicates that a molecule performs well under a particular computational scoring model. It does not independently prove that the molecule will bind to a protein in a real biological environment, much less demonstrate that it can alter protein function.

Paper 2 provides a clear example. AI Derivative 1 achieved competitive docking scores, but its actual receptor-modulating activity and effects on angiogenesis have not been experimentally established.

7.2 AI Performance Depends on Data and Models

AI models learn patterns from training data. If the data are incomplete, biased, or unrepresentative of real biological conditions, their predictions may also be affected.

Paper 1 highlights the complexity of biomedical data, insufficient model interpretability, and reproducibility challenges across different studies as important obstacles to practical AI applications.

Paper 3 demonstrates that AtomNet can identify candidate compounds across many projects with limited target-specific activity data. However, this does not mean that all AI models possess the same ability to generalize to unfamiliar targets.

7.3 Predicted Drug-Likeness Does Not Guarantee Safety

AI can estimate absorption, metabolism, and toxicity, but these results remain predictions.

Even if a molecule binds to its intended target, it may encounter several problems:

It may fail to reach an effective concentration in the human body.

It may interact with unintended proteins.

It may produce liver, cardiovascular, or other toxic effects.

It may be unstable under cellular or physiological conditions.

It may fail to achieve the intended tissue distribution or therapeutic effect.

These questions generally require further experimental investigation.

7.4 There Is Still a Long Path from a Candidate Molecule to an Approved Drug

AI can accelerate early-stage discovery, but drug development also involves safety assessment, pharmacokinetics, clinical trial design, patient variability, and regulatory approval.

Even a molecule that demonstrates biological activity in the laboratory may fail during subsequent preclinical studies or human clinical trials.

Therefore, the speed at which AI identifies candidate molecules cannot automatically be translated into an equivalent reduction in the total drug development timeline.

VIII. Future Development Directions

Based on the technological trends reflected in the three papers, AI drug discovery may continue to develop in several directions.

8.1 More Accurate Protein Structure and Interaction Prediction

Structural prediction models will continue to help researchers identify potential binding pockets, analyze protein conformational changes, and investigate targets that were previously difficult to study.

However, static structure predictions will still need to be combined with dynamic simulations and experimental validation.

8.2 Moving from Molecular Generation to Multi-Objective Optimization

Future models will need to do more than generate molecules that may bind to a target.

They will increasingly need to consider selectivity, toxicity, solubility, stability, and synthetic feasibility simultaneously, reducing the number of candidates that appear promising computationally but are difficult to develop in practice.

8.3 Combining AI with Automated Laboratories

AI may propose candidate molecules, automated laboratories may test them, and the experimental results may then be fed back into the models.

This closed-loop approach could make subsequent molecular design more targeted and help researchers identify incorrect predictions more efficiently.

8.4 Greater Emphasis on Personalized Medicine

Particularly in oncology, integrating genomic, proteomic, and clinical information may help researchers identify disease mechanisms and potential therapeutic targets specific to different patients.

However, implementing personalized treatment in clinical practice still requires rigorous clinical evidence.

The key future advance will not simply be the ability to generate more molecules. It will be the ability to improve the overall efficiency and reliability of the process from prediction and synthesis to experimental validation and clinical investigation.

IX. Conclusion

Taken together, the three papers support several important conclusions.

First, artificial intelligence is shifting drug discovery from a process centered primarily on experimental screening toward an integrated approach combining computational prediction with experimental validation. By analyzing protein structures, integrating multi-omics data, and predicting molecular interactions, AI can help researchers identify potential targets and candidate compounds more efficiently.

Second, large-scale virtual screening has received substantial experimental support. Atomwise's study involving 318 projects demonstrated that AtomNet could identify biologically active compounds across diverse targets, including targets for which direct target-specific activity data were unavailable in the model's training data. This suggests that AI can explore some areas of chemical space that have received limited prior investigation.

Third, generative AI is expanding the possibilities of molecular design. The S1PR1 study demonstrated how AI can use an existing drug molecule as a starting point to generate new candidate ligands and evaluate them through molecular docking, pharmacokinetic property prediction, and molecular dynamics simulations. However, the study has not yet demonstrated that its candidate molecule possesses actual receptor-modulating activity or therapeutic efficacy.

Finally, AI's development does not mean that experimental research has become unnecessary. On the contrary, experimental validation remains essential for distinguishing promising computational predictions from candidate molecules with genuine biological significance.

Overall, the most important contribution of AI is its ability to help researchers generate candidate solutions more quickly and systematically, while directing limited experimental resources toward the most promising research directions. AI has demonstrated considerable potential to transform early-stage drug discovery, but the path from a computationally generated molecule to a safe, effective, and ultimately approved drug still requires rigorous experimental and clinical research.

References

The following three papers form the primary basis of this analysis.

Li, D. et al. (2026). AI Accelerates the Identification of Druggable Targets Through 3D Structures of Proteins and Compounds. npj Precision Oncology, 10, 133. https://doi.org/10.1038/s41698-026-01310-7

Buhsem, O. & Ali, N. (2026). In Silico AI-Guided De Novo Design of S1PR1-Targeting Ligands from a Siponimod Scaffold for Angiogenesis and Fat-Cartilage Graft Survival. Scientific Reports, 16, 31320. https://doi.org/10.1038/s41598-026-73676-4

AI Is a Viable Alternative to High Throughput Screening: A 318-Target Study. (2024). Scientific Reports, 14, 7526. https://doi.org/10.1038/s41598-024-54655-z

Note: This analysis is based on the three papers and distinguishes between technologies summarized in the review, experimentally observed results, and candidate molecules that remain at the computational prediction stage.