Headlines claiming that “AI is reshaping drug discovery” are becoming commonplace at a time when AI is all the hype. Insilico Medicine, one of the biotech companies leading the wave of AI-driven drug research, announced that they discovered a novel drug molecule using AI in less than 18 months and that it cost them 10% as much as a conventional program. Pharmaceutical companies and researchers are promising a revolution in drug discovery with the help of AI, but how much of that promise has actually been realized?
It’s not that the claims are completely baseless. The power of AI in understanding biology was demonstrated in 2020 with the release of AlphaFold, an AI system that can predict the shape of proteins. Use of this software has become the standard in many fields of biomedical research, ranging from cancer mechanisms to protein dynamics. Andrea Pauli, a biochemist at the Research Institute of Molecular Pathology in Vienna, says that her lab uses AlphaFold for virtually every project. In 2018, her research team was lost on how a protein in zebrafish recognized sperm cells. This all changed when they used AlphaFold to predict a protein structure that helped explain the missing pieces. The success of the platform held great promise for what was coming in the field of AI-driven drug discovery.
Novel drug molecules, lower costs, and faster clinical trials are just some of the benefits being sold to us now that AI platforms are becoming more advanced. Conscience is dedicated to promoting AI for advancing drug discovery while fostering honest conversations about where the field currently stands. To evaluate the extent to which AI has accelerated drug discovery, we first need to understand its potential applications across the drug discovery timeline.
A multipurpose tool
AI-driven drug discovery can mean many different things because of how complex the drug discovery pipeline is. Newer versions of AlphaFold, like AlphaFold 3 (released in 2024), can predict how a drug-like molecule interacts with a protein in just seconds – a task that would otherwise take years of lab work to discover experimentally. Other AI models take this further, helping design entirely new drug candidates by testing a wide selection of molecular structures with increased efficiency, or by combining molecular fragments in new ways. For example, it may be known that molecule A and molecule B can both bind to a protein. An AI model may then design a molecule that is a hybrid of both, molecule A-B, that can still bind to the protein but would cause different downstream biological effects. This new molecule can then be tested and potentially developed as a novel drug candidate.
Optimizing drug molecules for a specific target is another way AI can contribute to drug discovery. A well-trained AI can tell scientists how subtly tweaking a molecule’s structure would lead to changes in absorption or toxicity when used in biology. It can also search for specific biological targets, like proteins or receptors, that could lead to disease. With these targets in mind, researchers can then design drugs that suppress their function to hopefully combat the disease.
Beyond the hype
While AI can serve many functions in drug discovery, reports about its actual impact can be misleading. Derek Lowe, a drug discovery chemist with over 35 years of experience, published an article in 2024 critiquing a scientific review paper for overstating the novelty of certain AI-discovered targets. He highlighted that if the biological target was already known to be associated with the disease, then can we really deem it to be “discovered by AI”? Success achieved by models identifying drug candidates or targets based on well-established knowledge may inflate the power of AI in this process, since re-identifying something already known is a much weaker test than finding something genuinely new. This is why prospective validation — testing predictions before the real-world outcome is known — matters so much for judging whether a model is actually uncovering something new. Conscience’s CACHE Challenges are designed around this kind of prospective testing. Participants use AI to predict novel molecules that could bind to a specific disease-linked target, and those predictions are then validated experimentally in the lab, providing independent, real-world evidence of how likely they are to work as drug candidates.
This kind of real-world testing matters because predictions don’t always hold up once they leave a lab, which is why benchmarking is essential. How do you quantify whether a predicted ‘hit’ will perform once it leaves the lab and enters a real, biological system? Biology is extremely complex – systems in the body like cell interactions can drastically influence the efficacy of a drug. So even though AI may identify many hit compounds in an isolated lab context, these molecules often struggle to perform when put into a biological one. There remains a disconnect between the chemical data AI models are being trained on and the biological problems they are expected to solve. It’s exactly this kind of rigour gap that BEACON was built to close, bringing independent, evidence-based assessment to AI-driven predictions across biomedical science.
One major reason for this poor biological performance is a lack of biologically contextualized data for models to be trained on. The interconnected systems in biology mean that a change in one variable can lead to a thousand perturbations elsewhere in the system. Such high variability built into biological data means that data points are often only applicable to hyper-specific scenarios. For example, while patient A and B might both be suffering from type I diabetes, their age, lifestyle, social class, and a multitude of genetic factors can all affect how they respond to a drug. Combined with the difficult task of consistent naming or labelling of data between researchers, generating all-encompassing and reliable data is a challenging task.
AlphaFold offers a useful point of comparison for what’s possible when that data problem is solved. What makes it work so well? AlphaFold had access to around 170,000 experimentally validated protein structures from the Protein Data Bank. Their model had vast amounts of reliable data to be trained on and didn’t involve a slew of biological variability either. These kinds of existing conditions were fundamental to what made AlphaFold so successful, but it’s a luxury that most drug discovery researchers don’t have.
What now?
AI has accelerated individual processes in the drug discovery pipeline, but getting a drug into market is still a couple steps away. As of July 2026, no AI-discovered drug has yet reached market — though progress is being made. Insilico Medicine’s Rentosertib, an AI-designed treatment for a progressive lung disease, has become the first AI-designed drug to reach Phase III trials, the furthest any AI-discovered candidate has progressed so far. However, the limited number of success cases illustrates a design gap that remains: AI systems for drug discovery have not been designed and optimized for biology, and there is simply not enough meaningful data to train on. Identifying these critical weaknesses is the first step. Now, changes need to be made for future AI-driven research to achieve real clinical impact.
AI models for drug discovery need to be trained for their end product rather than a promising statistic early on in the pipeline. Biological context needs to be incorporated into AI models to help inform which molecules will perform well in later stages of a clinical trial. This means using data from existing clinical trials to iterate on future drugs, evaluating a drug’s performance by biologically relevant metrics, validating models based on its use-case, and generating detailed and diverse datasets that capture the cause-and-effect of biological systems.
The challenges we are facing with AI-driven drug discovery highlight just how much room there is to improve. Realizing its potential will require a significant cultural shift towards a biology-informed perspective to design AI tools — one built on rigorous, independent testing, as programs like CACHE and BEACON are already demonstrating. Open sharing of knowledge and cross-discipline collaboration will be key to making meaningful strides in the coming years. It’s this kind of work that Conscience exists to support, in the pursuit of getting new treatments (including ones discovered with the help of AI) to the patients who need them.
