Computer-Aided Design and Artificial Intelligence Reshape Modern

A detailed scientific review published in Discover Pharmaceutical Sciences examines how computer-based methods and artificial intelligence (AI) are changing modern drug research and development. Bringing a new prescription drug to market currently costs an average of about 2.6 billion US dollars before taxes and takes 10 to 15 years. More than 96 percent of drug discovery projects fail, and about 90 percent of drugs fail during clinical development. Only about 13 percent of new small-molecule drug candidates are expected to receive clinical approval after going through the entire development process. To reduce these high failure rates and financial risks, researchers are increasingly using computer-aided drug design, computational chemistry, and automated analysis tools. These tools allow scientists to study a molecule's physical and chemical properties, selectivity, possible side effects, biological activity, and how the body absorbs and processes it before investing large amounts of time and money in laboratory experiments.

The traditional drug development process includes identifying a biological target, checking that it is suitable, finding promising compounds, improving those compounds, selecting a lead candidate, testing it before human studies, and then carrying out clinical trials. In the past, researchers mainly focused on targets that could be reached by traditional drugs. These targets were usually proteins with known three-dimensional shapes and pockets where small molecules taken by mouth could attach strongly. More than 90 percent of approved medicines target proteins. Drug design has traditionally been guided by Lipinski's Rule of Five, introduced in 1997. This rule identifies four chemical features linked to poor absorption or movement through biological membranes: a molecular weight above 500 Daltons, more than ten hydrogen-bond acceptors, more than five hydrogen-bond donors, and a calculated logarithm of the partition coefficient above five. Modern drug research is moving beyond these traditional limits. It now also studies other types of treatments, including peptides, peptidomimetics, oligonucleotides, two-part molecules such as proteolysis targeting chimeras, biologically active carbohydrates, and lipid signaling pathways involving enzymes and G protein-coupled receptors.

Artificial intelligence, especially machine learning and deep learning, has created automated computer systems that can examine huge collections of chemical compounds and biological data. Deep learning uses complex artificial neural networks inspired by the way electrical signals move through biological nerve tissue. These networks contain connected computing units called nodes and perceptrons. Different types of networks perform different tasks. Multilayer perceptrons use supervised learning, where models learn from examples, to recognize patterns, support control tasks, and make decisions about processes. Recurrent neural networks can store and reuse information through feedback loops, including systems such as Hopfield networks and Boltzmann machines. Convolutional neural networks are useful for modeling biological structures, recognizing patterns, and classifying images. Other approaches include Kohonen networks, learning vector quantization, radial basis function networks, clinical phenotype networks, and adaptive linear element networks. Computer systems such as the IBM Watson supercomputer have shown that they can examine complex clinical information and reference databases. One demonstration identified breast cancer-related data in less than 60 seconds.

Deep learning is also helping researchers solve long-standing problems in the study of large biological molecules. Scientists have often been unable to determine the exact atomic structures of proteins through experiments. Tools such as AlphaFold and its extension, ColabFold, can predict how a protein folds into its three-dimensional shape using its amino acid sequence. Their predictions can reach accuracy levels comparable to experimental methods such as X-ray crystallography and cryo-electron microscopy. These predicted structures can then be used in computer-based screening. This process tests how well potential drug molecules might bind to a target protein. Researchers can screen virtual chemical collections containing trillions of small molecules. Other methods include fragment-based drug discovery and synthon-based screening. These approaches use smaller, modular chemical building blocks to make it easier to create compounds and improve promising drug candidates. Ligand-based methods provide another option. They use mathematical and statistical methods to connect a molecule's chemical structure with its biological effects when the exact structure of the target is not known. This allows researchers to improve drug candidates even when experimental three-dimensional structural data are unavailable.

Three-dimensional quantitative structure-activity relationship methods help researchers understand the physical forces involved when drugs interact with their targets. Comparative molecular field analysis calculates electrical forces using Coulomb's law and measures how molecular shapes interact using Lennard-Jones potential functions. It helps show how small changes in a molecule can affect how strongly it binds to a target. However, this method requires molecules to be carefully aligned and requires researchers to set distance limits close to van der Waals surfaces. Comparative molecular similarity indices analysis expands this approach by also examining hydrophobic properties and the areas that can donate or accept hydrogen bonds, along with molecular shape and electrical forces. It uses continuous Gaussian distance functions to measure how molecules relate to points on a computer grid. This makes the results less sensitive to molecular orientation and distance limits. The method has produced consistent links between molecular structure and biological activity in computer studies of the gamma-aminobutyric acid type A receptor and microsomal prostaglandin E synthase-1.

Molecular dynamics simulations help researchers study how stable large biological molecules and their complexes are, and how they move over time, under simulated conditions similar to those in the body. Biological molecules constantly change shape and move their loops. A single structure obtained through crystallography cannot show all these movements. However, the main structure of a protein often remains stable during a simulation. Standard molecular dynamics simulations follow repeated steps. First, the computer sets the starting positions and speeds of all atoms. Next, it calculates the forces between the atoms using special mathematical descriptions of their interactions. Finally, it calculates how the atoms move over small time steps, usually measured in picoseconds. These simulations have helped explain how small drug molecules enter binding pockets in target proteins and how the contact areas between molecules change. Examples include microsomal prostaglandin E synthase-1 and the influenza M2 proton channel. Researchers measure changes in molecular shapes over time using root-mean-square deviation values.

Researchers have also improved the design of completely new molecules, known as de novo molecular design, using generative neural networks. These systems can avoid some of the problems of older methods, which often produced molecules that were difficult or impossible to make or had unpredictable biological effects. Generative models create new molecular structures using Simplified Molecular Input Line Entry System (SMILES) strings taken from large chemical databases such as ChEMBL. Advanced systems such as the Reinforced Adversarial Neural Computer combine reinforcement learning with deep neural networks. They can create different chemical structures with selected molecular weights, fat-solubility properties, and topological polar surface areas. Compared with other systems, such as ORGANIC, this approach has shown greater structural novelty while maintaining the integrity of molecular frameworks. Another system, Reinforcement Learning for Structural Evolution, combines generative networks with stack memory and prediction networks. It can create possible drug molecules while also predicting their biochemical properties. At the same time, retrosynthesis tools such as Synthia, previously called Chematica, use rules from organic chemistry, symbolic artificial intelligence, and Monte Carlo tree searches to find practical and affordable ways to make newly designed drug molecules.

Computer models and machine learning are increasingly used to predict how a drug will behave in the body before it is tested in humans. Researchers use these methods to estimate absorption, distribution, metabolism, excretion, and toxicity, together known as ADMET. These predictions can help reduce failures in later clinical studies. Deep learning models demonstrated their usefulness by winning the Tox21 Data Challenge, achieving classification accuracy above 80 percent across different chemical datasets. Other specialized computer models predict where epoxidation may occur in a molecule to help identify the formation of reactive metabolites. Pharmacogenomic research also uses information about single-nucleotide polymorphisms, which are small genetic differences between people, to explain why individuals process medicines differently. One example involves opioid pain medicines. The enzyme cytochrome P450 2D6 carries out O-dealkylation, while cytochrome P450 3A4 carries out N-dealkylation. Codeine and tramadol are opioid medicines that need functional CYP2D6 activity to be converted into forms that expose a phenolic hydroxyl group. This group is important for binding to histidine residues in the opioid receptor. These findings show how models of genetic differences can help doctors choose treatments and adjust doses for individual patients.

The real-world importance of computer-based drug design can be seen in AI-designed medicines that have reached controlled clinical trials in humans. Insilico Medicine used its generative AI platform to identify a biological target called TRAF2- and NCK-interacting kinase and designed a new small-molecule inhibitor named ISM001-055 to treat idiopathic pulmonary fibrosis. The drug was tested in a randomized, double-blind, placebo-controlled Phase IIa clinical trial at 21 sites in China. The study included 71 patients. ISM001-055 was taken by mouth and showed dose-dependent improvements in breathing measurements, along with a favorable safety profile. After 12 weeks, patients taking 60 milligrams once daily had an average increase of 98.4 milliliters in forced vital capacity. This measure shows how much air a person can force out of their lungs after taking a deep breath. By comparison, patients receiving the placebo had an average decrease of 62.3 milliliters. Mild to moderate diarrhea and abnormal liver function were reported in 14.8 percent of patients receiving the treatment. Another example is DSP-1181, a selective serotonin 5-HT1A receptor agonist developed by Exscientia and Sumitomo Dainippon Pharma for obsessive-compulsive disorder. It moved from the first stage of target identification to Phase I testing in humans in Japan within 12 months.

Despite these advances, the review explains that computer-aided drug design still has important limitations. Its main challenges include dependence on data quality, difficulty understanding how some algorithms reach their predictions, and the need to confirm all important findings through experiments. Deep learning models often work like black boxes. They can predict how a chemical compound may behave without explaining the biological or physical processes behind its interactions with other molecules. The accuracy of computer screening also depends heavily on the quality of the training data and experimental information used to build the models. Reliable data are still limited for complicated disease pathways and proteins whose structures have not been studied. Delivering medicines across barriers in the body is another major challenge. For example, getting drugs through the blood-brain barrier to treat cancers of the central nervous system remains difficult. Researchers are exploring special delivery methods, including convection-enhanced delivery catheters, as well as changes to drug molecules. Looking ahead, combining advanced computer models, new quantum computing simulations, and repeated laboratory experiments could provide a strong foundation for understanding complex biological systems and making drug discovery faster and more reliable.

Subject of Research: Computer-aided drug design and artificial intelligence in modern pharmaceutical discovery

Article Title: Advancing drug discovery: the role of computer-aided design and development in modern pharmaceuticals

Article References: Neelam, A. K. N. (2025). Advancing drug discovery: the role of computer-aided design and development in modern pharmaceuticals. Discover Pharmaceutical Sciences, 1(1), Article 8. https://doi.org/10.1007/s44395-025-00008-2

Image Credits: AI Generated

DOI: 10.1007/s44395-025-00008-2

Keywords: Computer-Aided Drug Design, Artificial Intelligence, Machine Learning, Deep Learning, Pharmacokinetics, Molecular Dynamics, Quantitative Structure-Activity Relationship, Clinical Trials, Advancing, drug, discovery, role