How AI Is Changing Drug Discovery
A stage-by-stage look at where machine learning enters the pharmaceutical pipeline
Drug discovery is slow, expensive, and has a high failure rate. Bringing a single drug to market takes an average of 10–15 years and costs over $2 billion. AI is being applied at nearly every stage of this process to reduce time, cost, and attrition.
Target Identification
Before you can design a drug, you need to know what biological target to go after — typically a protein involved in disease. Traditionally this relied on academic literature and manual hypothesis generation. AI approaches now mine genomic, transcriptomic, and proteomic datasets to identify novel targets. Insilico Medicine's PandaOmics platform, for example, uses transformer models trained on multi-omics data to rank potential drug targets by disease relevance.
Hit Finding & Virtual Screening
Once you have a target, you need molecules that bind to it. Traditional high-throughput screening tests millions of physical compounds, which is expensive and limited to existing chemical libraries. Virtual screening uses computational models to predict binding from molecular structure. Docking tools like Glide and AutoDock have existed for decades, but newer ML-based approaches like DiffDock use diffusion models to predict binding poses with higher accuracy and speed.
Lead Optimization
A "hit" molecule that binds a target is rarely a drug. It needs to be optimized for potency, selectivity, solubility, metabolic stability, and safety. This is where generative chemistry models are making significant impact. Models like those in Insilico's Chemistry42 platform propose molecular modifications that optimize multiple properties simultaneously — a multi-objective optimization problem that is difficult for human chemists to navigate efficiently.
ADMET Prediction
ADMET stands for Absorption, Distribution, Metabolism, Excretion, and Toxicity. These properties determine whether a drug candidate will work in the human body. Predicting ADMET early saves years of wasted development. Graph neural networks and transformer models trained on experimental ADMET data can now predict many of these properties from molecular structure alone, flagging problematic compounds before they enter expensive animal studies.
Structure Prediction
Understanding the 3D structure of a protein target is critical for rational drug design. Experimental methods like X-ray crystallography and cryo-EM are accurate but slow and expensive. AlphaFold, developed by DeepMind, demonstrated that deep learning could predict protein structures from amino acid sequences with near-experimental accuracy. This has made structural information available for nearly every known protein, enabling structure-based drug design at unprecedented scale.
Clinical Trial Design
AI is also being applied to clinical trials: identifying patient populations most likely to respond, optimizing trial protocols, and predicting outcomes. This is a less mature application area, but companies like Unlearn.AI and Recursion are working to reduce trial timelines using synthetic control arms and predictive modeling.
Where We Are Today
As of 2025, several AI-designed drug candidates have reached clinical trials. Insilico Medicine's INS018_055 for pulmonary fibrosis is in Phase 2, making it the most advanced fully AI-discovered and AI-designed compound. Exscientia, Recursion, and Relay Therapeutics also have AI-derived programs in the clinic. The technology is no longer theoretical — the question is how much faster and cheaper AI can make drug discovery compared to traditional approaches.