AI Drug Discovery Hits a Data Wall, Demands Lab Integration

The pharmaceutical industry's AI gamble

Artificial intelligence has become the pharmaceutical sector's primary strategy for reversing decades of rising costs and lengthening timelines. Developing a new drug now takes 10 to 15 years and costs between $1 billion and $2.5 billion, with failure rates exceeding 90 percent. Since the 1950s, drug development costs have doubled roughly every nine years—a trend known as Eroom's Law.

AI's early promise lies in hit identification, the process of screening molecular libraries against disease targets to find binding candidates. Rather than physically testing compounds at scale, companies now use AI to design drug candidates computationally and predict their interactions before committing resources to lab work.

"AI does away with that," says Paul Belcher, director of protein research strategy at Cytiva, a global life sciences company. "And it can help eliminate low-quality candidates before you have to physically test them, saving time and resources."

The missing half of the dataset

But AI models trained on publicly available data are encountering what Belcher calls a data wall. These datasets suffer from publication bias—they contain only successful experiments. Failed compounds, negative results, and binding failures remain locked in lab notebooks, never feeding back into model training.

"Most publicly available datasets and scientific publications focus exclusively on positive results," Belcher explains. "No one wants to share their failures. This bias is almost like having one hand tied behind your back."

Without comprehensive negative data, models cannot learn to avoid bias or reliably predict which compounds will fail. The result: AI can generate more hit candidates, but every one still requires wet-lab validation. Traditional screening workflows were built for scale, not for profiling diverse, AI-generated compounds in detail. Lab teams now face mounting pressure to characterize a growing volume of complex candidates.

Data integrity and the fabrication problem

Generative AI has made data fabrication trivial, compounding existing integrity concerns. Research by microbiologist Elisabeth Bik found that nearly 4 percent of biomedical papers contained duplicated or manipulated images—before generative tools existed. Western blots, a standard protein identification technique, are among the most commonly manipulated.

"Manipulated or faked data has always been a problem in science, but in the AI world, especially when used to train models, it could have potentially disastrous consequences," Belcher notes.

Some vendors are deploying verification tools. Cytiva's Image Integrity Checker uses blockchain-style hash algorithms to detect tampering in scientific images. Publishing houses are showing interest in adopting such tools as standard practice.

The integration bottleneck

Belcher envisions fully autonomous labs operating around the clock, cycling through prediction, testing, and optimization while feeding results back into AI models. But most labs lack the interoperable systems and structured data infrastructure required.

"Today, a lot of the instruments in labs are standalone," Belcher says. "You can have the best technology in the world, but if it's a closed ecosystem—if the user can't get the data out—it doesn't do any good."

Integrated systems could enable labs to generate FAIR data—findable, accessible, interoperable, and reusable—at scale, closing the loop between computational design and physical validation.

Why it matters

No AI-discovered drug has yet received full FDA approval, though Belcher expects that milestone within two to three years. The pharmaceutical industry is under intense pressure to improve success rates before clinical trials, where the majority of costs accumulate. AI's ability to compress early-stage timelines depends entirely on data quality and lab infrastructure—two areas where most organizations remain fragmented. Without access to negative results and integrated systems, AI models will continue hitting performance ceilings, limiting their impact on the industry's cost and timeline crisis.

The cost equation

A Stanford study found that training costs for frontier AI models have more than doubled annually since 2016. Belcher acknowledges the tension but remains optimistic: "As long as the cost of compute doesn't ever outweigh the cost of clinical development, I think AI is going to be an advantage."

These details were first reported by MIT Technology Review.