How AI and QSAR Modeling Accelerate Ligand-Based Drug Design

By Hugo Francisco de SouzaReviewed by Lauren Hardaker

Why Ligand-Based Drug Design Remains Important
How Does QSAR Modeling Help Predict Drug Activity?
What Is Pharmacophore Mapping, and How Is It Used?
How Do Similarity Searches and Predictive Modeling Support Drug Design?
What Are the Strengths and Limitations of Ligand-Based Drug Design?
References
Further Reading

For decades, pharmaceutical development has predominantly relied on structural "locks" to design molecular "keys." However, as the modern medical community increasingly investigates the pharmacological potential of the "dark proteome", specifically, unknown or poorly studied proteins with highly flexible and sometimes transient structures, reviews highlight that traditional mapping has reached its limits.

Ligand-based drug design (LBDD) offers a possible solution. The method uses the known chemical properties and physiological effects of previously studied active compounds as "memory," which is then used to predict novel therapeutics without requiring an atomic-level map of the target itself.

LBDD has increasingly combined the advantages of classical Quantitative Structure-Activity Relationship (QSAR) modeling with generative artificial intelligence (AI), including deep learning and reinforcement learning frameworks, enabling researchers to compress decade-long discovery timelines into mere months.^1,6

This article reviews recent advances in ligand-based drug design, focusing on computational frameworks such as Activity Cliff-Aware Reinforcement Learning (ACARL) and deep graph neural networks that help researchers address data bias and identify subtle structure-activity patterns.

It emphasizes that as AI-discovered candidates like Rentosertib advance through clinical trials with positive efficacy and safety results, the emergence of ligand-centric AI suggests a future in which the absence of structural data no longer impedes the discovery of life-saving medicines.

Why Ligand-Based Drug Design Remains Important

LBDD is a machine (usually artificial intelligence [AI])-aided approach to novel drug discovery and development. The method works by analyzing molecules known to interact with a biological target, rather than by experimentally identifying and validating the target’s specific three-dimensional (3D) structure, a prolonged, highly resource-intensive process.

While a relatively novel approach to drug development, LBDD remains a fundamental and growing pillar of pharmaceutical research, particularly in scenarios where the 3D structure of a potential biological target is unavailable or insufficiently resolved for structure-based approaches.¹

While advances in contemporary technologies, such as cryo-electron microscopy, have expanded the library of known protein structures, reviews in the field highlight that a substantial portion of the human proteome remains "undruggable" by traditional methods due to inherent protein flexibility or in vitro instability. In these contexts, LBDD effectively shifts the analytical focus from the structure of the biological receptor to the small molecules (“ligands”) previously identified as interacting with it.²

The logical framework of LBDD approaches is rooted in the "similarity principle," which suggests that molecules with similar physicochemical properties are likely to exhibit similar biological activities. The benefits of this principle mean that LBDD technologies are not just secondary fallbacks in events of limited structural data, but also a complementary strategy that provides unique insights into interaction entropy.¹

By establishing quantitative correlations between structure and activity, LBDD enables the prioritization of molecules for synthesis, extensively reducing the time and financial burden associated with traditional high-throughput screening.¹

Reviews of contemporary literature frequently reveal that historically, ligand-centric design was the primary driver of rational drug design before the widespread availability of protein structures. The advent of the AI age has revitalized this field, and it has now evolved into a high-throughput computational discipline integrated with deep learning.³

Recent reviews further emphasize that AI-enhanced QSAR platforms are increasingly integrated with omics datasets, cloud computing infrastructure, and real-world biomedical evidence to support scalable virtual screening and personalized therapeutic design.^1,7

Video credit: TEDx/Youtube.com

How Does QSAR Modeling Help Predict Drug Activity?

The core of LBDD involves correlating molecular structure with biological potency through Quantitative Structure-Activity Relationship (QSAR) modeling. This process converts chemical structures into numerical representations known as molecular descriptors.¹

These descriptors are primarily categorized by dimensionality: 0D descriptors include basic atom counts; 1D descriptors capture functional groups; 2D descriptors utilize topological indices to reflect connectivity; 3D descriptors focus on molecular volume and electrostatic potential; and 4D descriptors incorporate conformational ensembles.¹

Modern AI-integrated QSAR workflows additionally employ learned molecular representations ("deep descriptors") derived directly from molecular graphs or SMILES strings using graph neural networks and transformer architectures.⁷

Modern QSAR workflows involve standardizing datasets by transforming half-maximal inhibitory concentration (IC50) values into the negative logarithm pIC50 scale (pIC₅₀ = -log₁₀(IC₅₀ x 10^-9)) to ensure a linear distribution suitable for regression.⁴

Subsequently, statistical algorithms (primarily Multiple Linear Regression [MLR] and Support Vector Machines [SVM]) establish the mathematical relationship between the candidate ligand and the target substrate: Activity = ƒ(Physicochemical + Structural Descriptors).¹

Recent research has described this process in studies targeting Kirsten Rat Sarcoma Viral Oncogene Homolog (KRAS) inhibitors, wherein “Genetic Algorithms (GA)” were used to select optimal descriptor (compound) subsets, maximizing adjusted R² values while penalizing model complexity.⁴

Comparative studies of KRAS inhibitor prediction models have further shown that Partial Least Squares (PLS) and Random Forest (RF) algorithms frequently outperform simpler linear approaches in predictive robustness, especially when handling nonlinear descriptor interactions.⁴

What Is Pharmacophore Mapping, and How Is It Used?

Pharmacophore mapping is the process that identifies the spatial arrangement of biological activity-defining “pharmacological” features: hydrogen-bond donors/acceptors, hydrophobic centers, and ionizable groups. A holistic pharmacophore model serves as a 3D blueprint, allowing researchers to screen extensive chemical libraries for ligand molecules that match specific biological functionality-defined spatial constraints.³

For example, in Dong & Hao’s (2025) study targeting VEGFR-2 and c-Met, the researchers screened more than 1.28 million compounds from the ChemDiv database. This screening enabled the research to identify pharmacophores with enrichment factors (EF) exceeding 10.0, thereby isolating 18 high-potential dual inhibitors in a fraction of the time and cost required by conventional approaches.⁵

Modern pharmacophore mapping protocols often integrate drug-likeness filters (e.g., Lipinski and Veber rules) into the screening process, which, when combined with subsequent molecular docking to determine binding modes, has been found to significantly improve algorithm accuracy.⁵

Reza and colleagues (2026) recently demonstrated that pharmacophore mapping approaches enabled the team to map more than 29,000 phytochemicals from the Natural Product Activity and Species Source (NPASS) database against the LKB1 protein (here, metformin and phenformin served as control ligands to filter out drug-like leads). Recent research in the field has established this integration of ligand-based queries with structural refinement as the current state-of-the-art in virtual screening.^3,5

How Do Similarity Searches and Predictive Modeling Support Drug Design?

The relatively recent integration of artificial intelligence (AI) in traditional LBDD approaches has transformed the field from a predominantly descriptive discipline into a generative framework. Modern deep learning models, particularly Graph Neural Networks (GNNs), enable networks to identify and "learn" relevant features directly from raw chemical graphs, where atoms are represented as nodes and bonds as edges.¹

Transformer-based architectures and reinforcement learning systems are increasingly being incorporated into de novo molecular design workflows because they can simultaneously optimize compounds across multiple biological and pharmacokinetic endpoints.^6,7

This integration, combined with the unprecedented computational prowess of modern datacentres, has already facilitated clinical success. Insilico Medicine’s Rentosertib (ISM001–055), a first-in-class TRAF2- and NCK-interacting kinase (TNIK) inhibitor discovered via the Pharma.AI platform, progressed from target discovery to Phase IIa trials in less than 30 months. This process is one of the fastest drug discovery pipelines ever reported, and it is estimated to have taken decades without AI assistance.²

Other AI-designed candidates, such as Exscientia's DSP-1181 for obsessive-compulsive disorder (OCD), entered Phase I trials following less than 12 months of discovery, highlighting that Rentosertib was not an exception, but likely represents the new norm in AI-assisted 'compressed' drug development timelines.²

Furthermore, the efficacy of AI-discovered pharmacological candidates matches or exceeds that of traditionally discovered agents. In clinical evaluations, Rentosertib has demonstrated a dose-dependent improvement in forced vital capacity (FVC) of 98.4 mL at 12 weeks, compared to a 62.3 mL decrease in the placebo group.²

One of the most significant recent methodological advances is the emergence of Activity Cliff-Aware Reinforcement Learning (ACARL), which explicitly models "activity cliffs" in which minor structural changes can produce major shifts in biological activity. By integrating a contrastive reinforcement learning loss function and an activity cliff index, ACARL frameworks improve the generation of high-affinity molecules while more accurately capturing discontinuous structure-activity relationships often missed by traditional machine learning systems.⁶

Discover How Proteomics and AI Are Unlocking Previously “Undruggable” Drug Targets

What Are the Strengths and Limitations of Ligand-Based Drug Design?

Reviews on ligand-based drug design emphasize that the platform’s primary strength lies in its exceptional speed and lower data requirements. Unlike structure-based methods, LBDD does not require an atomic-level map of the biological target, making it a critical tool for the 'dark proteome,' where experimental target data is costly.¹

Recent reviews additionally emphasize that AI-assisted computational pipelines substantially reduce preclinical failure rates by prioritizing promising compounds before chemical synthesis and biological testing begin.⁷

Modern LBDD approaches have further been shown to strengthen early-stage discovery by enabling the prioritization of compounds from databases of billions of molecules before laboratory synthesis begins. In contrast, traditional early-stage discovery often required the synthesis of each molecule under investigation, an extremely expensive process that took years of trial and error.¹

Despite these strengths, reviews caution that significant challenges persist, especially regarding prediction accuracy and data quality. Scientists have identified the activity cliff, where minor structural modifications can yield dramatic shifts in physiological activity that standard machine learning models often fail to predict, thereby substantially limiting their accuracy.⁶

Furthermore, "garbage in, garbage out" constraints remain structural limitations. A review exploring the limitations of QSAR approaches found that duplicated entries or conflicting measurements in public databases substantially distort similarity relationships and inflate performance metrics, leading to erroneous outcomes.⁷

Sampling bias, especially in AI model training data, also limits generalizability, as models trained on narrow chemical classes (e.g., pesticides) often fail when applied to structurally unrelated therapeutic compounds.⁷

Additional regulatory concerns include model interpretability, explainability, dataset governance, and reproducibility, all of which are increasingly emphasized by agencies such as the European Medicines Agency (EMA) and the World Health Organization (WHO).^8,9

While public health agencies and regulatory bodies, e.g., the US Food and Drug Administration [FDA] and EMA, aim to address these limitations through risk-tiered frameworks and credibility assessments, the rapidly evolving nature of the field means that product and methodological standardization remain lacking at the global scale.^8,9

WHO and EMA guidance documents further note that minimizing bias in AI training datasets, ensuring transparency in model validation, and maintaining human oversight will be essential prerequisites for trustworthy AI-assisted pharmaceutical development.^8,9

References

1. Koirala, M., Yan, L., Mohamed, Z., & DiPaola, M. (2025). AI-Integrated QSAR Modeling for Enhanced Drug Discovery: From Classical Approaches to Deep Learning and Structural Insight. International Journal of Molecular Sciences, 26(19), 9384. DOI:10.3390/ijms26199384, https://www.mdpi.com/1422-0067/26/19/9384
2. Verma, V., & Kumar, D. (2026). Artificial intelligence and machine learning in drug discovery: From lead discovery to clinical validation (2020–2025). Letters in Drug Design & Discovery, 22(12), 100341. DOI:10.1016/j.lddd.2026.100341, https://doi.org/10.1016/j.lddd.2026.100341
3. Reza, R., et al. (2026). Ligand‐Based Pharmacophore Mapping and Virtual Screening for the Search of Biguanide‐Like Molecules With Antidiabetic Potentials Targeting Liver Kinase B1. Biochemistry Research International, 2026(1). DOI:10.1155/2026/8369459, https://onlinelibrary.wiley.com/doi/10.1155/2026/8369459
4. Adebayo, O. S., Ambrose, G. O., Olusola, D., Oluwafemi, A., Alzahrani, H. A., & Hasan, A. (2025). QSAR-guided discovery of novel KRAS inhibitors for lung cancer therapy. Frontiers in Bioinformatics, 5. DOI:10.3389/fbinf.2025.1663846, https://www.frontiersin.org/journals/bioinformatics/articles/10.3389/fbinf.2025.1663846/full
5. Dong, J., & Hao, X. (2025). Pharmacophore screening, molecular docking, and MD simulations for identification of VEGFR-2 and c-Met potential dual inhibitors. Frontiers in Pharmacology, 16. DOI:10.3389/fphar.2025.1534707, https://www.frontiersin.org/journals/pharmacology/articles/10.3389/fphar.2025.1534707/full
6. Hu, X., Liu, G., Zhao, Y., & Zhang, H. (2025). Activity cliff-aware reinforcement learning for de novo drug design. Journal of Cheminformatics, 17(1). DOI:10.1186/s13321-025-01006-3, https://link.springer.com/article/10.1186/s13321-025-01006-3
7. Manica, A. K. (2023). Limitations of QSAR Modeling: Data Bias, Curation, and Predictive Reliability in Computational Drug Discovery. Bioinfo Chem, 5(1), 1-13, 10719. DOI:10.25163/bioinformatics.5110719, https://www.publishing.emanresearch.org/Journal/Abstract/bioinformatics-5110719
8. European Medicines Agency (EMA). (2024). Reflection paper on the use of artificial intelligence (AI) in the medicinal product lifecycle. EMA/CHMP/CVMP/83833/2023. https://www.ema.europa.eu/en/use-artificial-intelligence-ai-medicinal-product-lifecycle-scientific-guideline. Accessed 07^th May 2026
9. World Health Organization (WHO). (2024). Benefits and risks of using artificial intelligence for pharmaceutical development and delivery. ISBN: 978-92-4-008810-8. https://www.who.int/publications/i/item/9789240088108. Accessed 07^th May 2026

Further Reading

• All Drug Discovery Content
• Why Drug Discovery Needs Robots and Artificial Intelligence
• Hit to Lead (H2L) Process in Drug Discovery
• Understanding Lead Optimization
• How is AI Transforming Drug Discovery?

Last Updated: May 11, 2026