AI is changing the pace of drug discovery, but the harder lesson is that faster predictions do not remove the need for better experiments. As companies generate more possible drug candidates, the physical lab becomes more important, not less.
The central issue is the data loop. AI can propose molecules and narrow the field, but those ideas still need to be tested, characterized, purified, and fed back into models with enough quality and structure to improve the next round.
Why AI matters in early drug discovery
Drug discovery remains expensive, slow, and risky. Since the 1950s, the cost of developing new pharmaceuticals has roughly doubled every nine years, a pattern known as Eroom’s Law. Today, bringing a new drug to market takes an average of 10-15 years and costs anywhere from $1 billion to $2.5 billion, with failure rates upward of 90%.
That pressure explains why AI has become such a major focus for pharmaceutical companies. If teams can identify, test, and optimize new chemical compounds earlier and more efficiently, they may reduce the risk of expensive failures later in development.
One of the clearest early uses is hit identification. In this stage, researchers look for molecular entities that bind to a disease-related target, such as a protein. A successful hit gives scientists a starting point for more testing and refinement.
Paul Belcher, director of protein research strategy at Cytiva, describes a shift from empirical screening toward predictive design. Instead of relying only on physical screening of libraries, drug companies are using AI to design candidates from scratch and predict how they may interact with targets before committing them to R&D.
The lab bottleneck is getting more visible
AI can reduce the number of low-quality candidates that reach physical testing. It can also increase the number of promising hits available for follow-up. But that creates a new operational challenge: lab teams must handle a growing set of more diverse, AI-generated compounds.
Traditional screening workflows were designed to find hits at scale. They often used binary or threshold-based techniques that produced low-fidelity data: a basic yes-or-no result. That kind of workflow is not enough when teams need to understand complex candidates in more detail.
Belcher’s view is that AI still cannot reliably predict kinetics or developability for new compounds. That means every AI-generated candidate still needs lab validation. The result is a demand for technologies that can produce richer information at higher throughput.
This is where AI-driven drug discovery meets a very physical constraint. A model may move quickly, but instruments, samples, purification workflows, and characterization steps still determine whether a prediction becomes useful evidence.
Better models need better negative data
The quality of AI depends heavily on the quality and breadth of the data used to train it. According to Belcher, many earlier models trained on publicly available datasets are reaching a data wall. When models use the same available data, they can arrive at similar conclusions and deliver diminishing returns.
The problem is not only access. Many datasets were not built for AI, so they may lack the structure, labels, and diversity needed for accurate and less biased models.
Publication bias makes the issue worse. Scientific literature and public datasets tend to emphasize positive results. Failed experiments and compounds that do not bind are often missing, even though that information could help models learn what does not work.
Belcher argues that this missing negative data limits prediction quality. Without a broader view of both success and failure, models have less ability to avoid bias and make reliable distinctions.
Data integrity is becoming a drug discovery problem
AI also raises the stakes for data integrity. Fabricated or manipulated scientific data has long been a concern, but the risk becomes more serious when flawed data is used to train models.
The source article points to Western blots as one example. These are used to identify proteins in blood or tissue samples, and they are among the common targets for manipulation in biomedical research.
Belcher cites research by Dutch microbiologist Elisabeth Bik, who found that almost 4% of biomedical papers contained duplicated or manipulated images. That finding was from 2016, before generative AI made fabrication easier.
Some vendors are developing tools to address this. Belcher points to Cytiva’s Image Integrity Checker, which uses secure hash algorithms, described in the source as the same technology used in blockchain, to detect whether scientific images have been altered.
Autonomous labs depend on integration
The longer-term vision is a lab that can run with minimal human intervention. Belcher describes AI-driven dark labs, or labs-in-the-loop, that operate around the clock through cycles of prediction, testing, optimization, and feedback.
For that future to work, labs need consistent data and infrastructure. Instruments cannot remain isolated if the goal is to move information cleanly from the wet lab into computational systems and back again.
Integrated infrastructure could help labs generate FAIR data: findable, accessible, interoperable, and reusable. That kind of data could support individual reports while also training future AI models.
AI-driven drug discovery is still early. The source notes that no drug discovered primarily through AI-driven design has yet received full FDA approval, although Belcher expects that to change in the next two to three years.
The largest ambition is full in silico prediction of efficacy and toxicity, which would reduce the need for much of the physical wet lab work. But the source also points to barriers beyond model maturity, including regulatory hurdles and cost challenges. A Stanford study found that the cost of training frontier AI models has more than doubled every year since 2016.
For now, the practical path is balance. AI can speed design and prioritization, while lab systems provide the validation and data quality needed to make those predictions trustworthy.