Can computers discover drugs? The promise and limitations of computer-aided drug discovery
Last updated:
13/11/26, 15:15
Published:
19/11/26, 08:00
The most effective drug discovery combines computational and experimental approaches, using computers to narrow the search and the laboratory to confirm the results
Bringing a new drug to market is an expensive, arduous process. The estimated cost is around $2.6 billion, having increased by nearly 150% in the last decade, and around 90% of compounds that enter clinical trials never receive regulatory approval. These figures have pushed researchers to look for ways to circumvent such large failure costs, and one of the most significant answers has been computer-aided drug discovery (CADD). This article looks at what CADD is, why it has become so prominent, and where its limitations lie.
Why drug discovery needed a new approach
The shift in drug discovery methods has been driven largely by economics, and by the increasing difficulty of treating complex diseases such as cancer. Traditional drug discovery relied mostly on time-consuming, resource-intensive trial and error experiments. This proved costly and slow, especially for more challenging disease targets.
Computational methods offer an alternative. They allow quick screening of vast compound databases and the identification of potential binders through modelling and simulation. With roughly 90% of clinical trial compounds failing, a thorough filtering process early in the pipeline is essential and CADD provides this. It is therefore not just a scientific enhancement but also a more economically viable way of progressing pharmaceutical research.
What is CADD?
CADD is rooted in several sciences, including chemistry, biology and computer science. Its aim is to predict how potential drug molecules may interact with targets such as proteins, enzymes and DNA. It does this using molecular modelling techniques including structure-based drug design, ligand-based modelling, molecular docking, molecular dynamics simulations and free energy calculations, which together help identify and optimise therapeutic candidates.
Effective drug activity relies on the interaction between a protein and a ligand (the small molecule that binds to it). Identifying and characterising a binding site is therefore crucial to commence a drug discovery programme. Computational methods are used to extract the physicochemical properties of binding sites, such as shape, size, hydrophobicity and ability to form hydrogen bonds. This allows the design of ligands with the greatest potential for complementary interactions.
CADD employs two main techniques, structure-based drug design (SBDD) and ligand-based drug design (LBDD). Which one is used depends on whether the 3D structure of the protein or of known ligands is available, and these approaches form the theoretical and practical foundation of the field.
Structure-based drug design relies on the principle that the 3D structure of the therapeutic target protein and the binding site cavity’s characteristics are accessible. It provides information about the target of small molecules, genetic information and sequences, binding information, cytotoxicity, pharmacokinetic data and other biological information, all of which are relevant for drug development. The discovery of the 3D structures of many biological molecules has made SBDD increasingly significant. Its basic steps include target preparation, binding site identification, molecular docking, virtual screening and molecular dynamics.
What distinguishes SBDD is that the drug is designed around the target. Researchers can see three-dimensionally the pocket a molecule needs to fit into. They can then screen or design molecules for shape and chemical complementarity, rather than relying on chance.
Ligand-based drug design is used when there is no 3D information about the receptor. It relies instead on existing knowledge of molecules that bind to the desired biological target. LBDD techniques use known active compounds (ligands) to establish a structure-activity relationship between physicochemical properties and biological activity. This can refine existing drugs or guide the development of novel drugs with enhanced activity. Two key LBDD methods are quantitative structure-activity relationship (QSAR) modelling and pharmacophore modelling.
How AI has impacted the field
Large-scale advances in artificial intelligence (AI) and machine learning (ML) in recent history have expanded what computational drug discovery can do. Deep learning and generative models are now being combined with CADD, allowing the prediction of molecular properties and the exploration of vast areas of chemical space.
One of the most profound advancements has been the emergence of protein structure prediction software such as AlphaFold, which has led to a major expansion in the availability of high-quality 3D structures. Large-scale models trained on chemical and biological data can also learn complex structure-property relationships from molecular representations. They can then provide robust predictions for property prediction, reaction prediction and de novo molecular design.
Where CADD falls short
Despite this progress, CADD is not a magic bullet. Theoretical models require assumptions and approximations, so computational methods are not always accurate and bring their own limitations. Molecular docking or molecular dynamics simulations may not accurately represent a biological system, leading to incorrect predictions. The main challenges reside in data quality and availability, protein flexibility and dynamics, computational limitations, predicting pharmacokinetic (ADMET) properties, and regulatory acceptance and validation.
Data quality: CADD relies on accurate, precise, high-quality information about the structures of target proteins and ligands. If experimental structures are unavailable and homology models are used as a substitute, precision is reduced. Its accuracy depends on the abundance of accurate, up-to-date data, such as 3D structures of target proteins and ligand binding affinities expressed as the dissociation constant (Kd). If data are incorrect, partial or out of date, predictions about how candidates interact with targets can be wrong.
Protein flexibility: Most CADD models depict proteins as flat and rigid. In reality, proteins are dynamic and constantly in transition, and can undergo conformational change in response to stimuli. A drug’s binding affinity depends on this conformational plasticity, so ignoring it hampers prediction accuracy and wastes computational resources.
Computational cost: Complex simulations of drug molecules targeting and interacting with their targets need large-scale computation. These calculations are often expensive and resource-intensive, for example when analysing large systems or modelling protein flexibility, solvents and long timescales.
Predicting ADMET: Absorption, distribution, metabolism and excretion (ADME) properties, along with toxicity, determine drug efficacy and safety in humans, and current CADD models are not yet fully adept at predicting them. This is because they result from complex, elaborate biological systems made of many networks, molecules and pathways, and condensing these into a model is difficult and may not reflect real life. A drug might look like an excellent binder to its target in silico, yet in vivo have poor bioavailability or adverse effects that the models fail to detect.
Regulatory acceptance and validation: Predictions only matter if they can be trusted. Computational results must be validated experimentally, and regulators generally still require laboratory and clinical evidence. CADD can guide which compounds are worth testing, but it does not yet replace testing.
Conclusion
CADD has to a large extent revolutionised how drugs are discovered by making early-stage screening faster, cheaper and more targeted. With the rise of AI and tools such as AlphaFold, its capabilities will only continue to grow. However, predictions are only as good as the data and assumptions behind them, and the gap between in silico promise and in vivo reality, particularly for pharmacokinetics and toxicity, remains a major challenge. For now, the most effective drug discovery combines computational and experimental approaches, using computers to narrow the search and the laboratory to confirm the results.
Written by Emaan Alvi
Related articles: Beyond conventional small molecule drugs / AI in drug research and discovery / Not all chemists wear white coats / AI in medicinal chemistry
Project Gallery


