What if researchers could pinpoint viable drug targets, simulate molecules and weed out crappy compounds months before they start the tedious lab work? That day is arriving, and pharmaceutical research is never going to be the same.
This remains a tough business equation to crack-growing R&D expenditure versus shrinking research productivity, also known as Eroom’s Law. This bottleneck can be alleviated by the application of AI in drug discovery, which can substitute some traditional trial-and-error research with predictive modelling and data-driven research.
This paper explores the impact of AI on target identification, molecular design, compound screening and experimental validation and discusses its limitations.
Target Identification and Validation
All drug discoveries start from defining the right biological target for the drug. Associations of genes with a disease are found through genome-wide association studies but this does not imply causality. Genes associated with a disease are not necessarily the cause of a disease. Difficult because a researcher will often use different isolated data sets.
With the integration of transcriptomic, single cell RNA sequencing and proteomics, AI could, if anything, have the potential of giving a broader view of the biology of the disease. Knowledge graphs and graph neural networks (GNNs) could then be used to connect genes, proteins, pathways and phenotypes. This helps researchers identify hidden biological connections and investigate less-explored proteins beyond familiar, heavily studied kinase targets.
A 2026 analysis using Google DeepMind’s AlphaGenome Atlas demonstrated the potential of this approach. The findings revealed 22% additional non-coding genetic associations when variants are grouped by their molecular effects predicted in genome data from the UK Biobank. This enhances the clinical interpretation of these genetic signals, but they still need biological validation before they can be used to inform drug development.
Also Read: M42 Launches Beyond Bounds, Bringing Global Voices Together on Healthcare’s Biggest Questions
Molecular Prediction and De Novo Compound Generation
Generative AI use cases in drug discovery: Searching commercial chemical libraries for hits, or designing new molecules and predicting their properties prior to synthesis.
SE(3)-equivariant diffusion models: they learn to model molecules in 3D, and to infer the existence and location of atoms and bonds, all the while preserving the shape of the molecule when it is translated and rotated during the procedure. For generation, these models iteratively remove ‘noise’ to generate a molecule from a blank slate. Chemical language models based on transformers take another approach: they learn the structure of a molecule by learning a language of chemistry.
With structure-based prediction, you can build further by creating models that suggest how compound interacts with a binding site of a protein target, including binding pockets that are not evident in the screen. If the compound is predicted to strongly bind to the target it is not necessarily a ‘drug candidate.’ Multi-parameter optimization takes into account solubility, metabolism stability, binding strength, and score of SAS (synthetic accessibility score, indicating the complexity of synthesis).
Google Deep Mind showed the pipeline in action via an AI agent that enumerated thousands of molecules, performed matched molecular pair analysis and generated rigidified macrocycles with synthetic routes in a little over an hour. Typical workflows can take days per cycle. These suggestions still need to be synthesized and tested, but they demonstrate how AI can accelerate the journey from concept to candidate.
High-Throughput Virtual Screening and ADMET Profiling
When AI produces candidates, researchers then need to reduce the list to compounds that are worthwhile to make and test. Virtual screening can screen out billions of compounds, including combinatorial libraries, based on their likelihood to bind to a target. This reduces the amount of compounds to test in the lab and helps researchers identify lead series.
But this is merely one small piece of the puzzle. A drug can bind a target avidly and still not make it to the clinic if it’s not absorbed into the body, metabolized effectively, or cleared sufficiently. However, with the help of in-silico ADMET profiling, machine learning models can circumvent the difficulties. They determine ADMET – absorption, distribution, metabolism, elimination, and toxicity, that is – before compounds are synthesized and tested, and identify potentially serious safety issues, such as hERG inhibition, hepatotoxicity, or cytochrome P450 inhibition.
Microsoft’s Project Quine. Thousands of compounds were tested for their ability to induce a therapeutic state shift in pancreatic cancer cells. The top ranked hits induced the largest definition change in wet-lab assays, with the prioritization done in a single weekend. AI can be used to accelerate the selection of candidates but not to determine ADMET safety or replace toxicology.
Experimental Design and the Closed-Loop Autonomous Lab
All these calculations must be validated in the lab-this is where the Design-Make-Test-Analyze (DMTA) cycle is so crucial. AI models suggest a new molecular design, robots manufacture and test the design, and the experimental results then inform the next AI prediction. It’s a continuous feedback loop of hypothesis, test, and learn.
This is what self-driving laboratories make happen by bringing together active learning algorithms, robotic tools, liquid handlers, and microfluidic systems. Rather than having a researcher pick all of the experiments, Bayesian active learning will, for example, guide the model to the most informative experiments based on prior conditions, so the algorithm can then use the results to improve its predictions and determine what to test next.
Google Research demonstrated this approach in aptamer discovery. Aptamers are short DNA or RNA molecules that bind to specific molecular targets. Following three experimental design-test rounds, AI-designed alpha-thrombin aptamers achieved a confirmed on-target hit rate of 42.1% and specificity of 98.5%. Such findings demonstrate how a program’s feedback may be used to inform design over iterations. Aptamer discovery is distinct from small-molecule development but the underlying concept remains valid. AI becomes increasingly beneficial when continuous flow of biological data serves to refine its predictions and close the gap between confidence and reality.
Balancing Algorithmic Promise with Clinical Reality
AI can expedite drug discovery, but the results that it proposes can’t replace biological data, or interpretation by medical experts. For drug makers, the real power will be in the ability to develop a reasonable infrastructure to support those tools. That will begin with robust biological information, good governance, and effective validation processes. Increasing the number of model parameters is worthless if the data is not good.
Foundation models, labs and domain experts will be brought closer together in the future but focus must equally be given to oversight before adoption. WHO identified transparency, bias, fairness, accountability, privacy and possible harm to be the common issues in health-related AI research. We should aim for accelerated discovery, not speed at the cost of patient safety.
