Reversing Eroom’s Law: How High-Fidelity Data Fuels AI in Pharmaceutical Research

The Economic Crisis of Drug Development

Since the middle of the 20th century, pharmaceutical innovation has been governed by Eroom’s Law, a trend where the cost of developing a new drug doubles roughly every nine years. Currently, bringing a therapy to market requires an investment between $1 billion and $2.5 billion, typically spanning 10 to 15 years. With failure rates exceeding 90%, the industry is increasingly leveraging artificial intelligence to compress these timelines and improve the quality of clinical candidates.

Paul Belcher, director of protein research strategy at Cytiva, notes that the primary financial burden lies within the clinical phase. By utilizing AI to refine risk assessment and candidate selection early in the process, companies intend to reach the clinic with higher-quality molecules. This shift represents a transition from empirical physical screening to predictive design, where researchers can now architect drug candidates from scratch and simulate their interactions with disease targets before initiating physical R&D.

The Bottleneck of Laboratory Validation

While AI excels at identifying potential hits, it cannot yet accurately predict the kinetics or manufacturability of new compounds. This creates a new operational challenge: labs must now validate a massive volume of diverse, AI-generated candidates. Traditional hit identification workflows were optimized for binary, low-fidelity results—simple yes-or-no data on whether a molecule binds to a target.

Modern AI-assisted research demands more sophisticated, information-rich technologies. Success depends on the ability to characterize and purify complex candidates at scale. Belcher suggests that the abundance of AI-generated leads is putting unprecedented pressure on laboratory infrastructure to move beyond simple threshold-based techniques toward higher-throughput, high-fidelity data collection.

The Data Wall and the Need for Failure

A significant obstacle to AI progress is what has been described as a data wall. Many current models rely on the same public datasets, which often lack the structure and diversity required for high-level accuracy. This problem is exacerbated by publication bias. Because scientific journals and researchers rarely document failed experiments, AI models are trained on an incomplete picture of reality.

• Public datasets focus almost exclusively on positive results, omitting the "negative data" that could help models avoid future errors.
• The absence of documented failures limits the model's ability to recognize patterns that lead to development dead-ends.
• AI models risk reaching similar, plateaued conclusions when fed identical, biased information.

Integrity concerns also mount as generative tools make data manipulation easier. Research from 2016 suggested that nearly 4% of biomedical papers involved manipulated images, such as Western blots. To combat this, vendors are developing verification tools like Cytiva’s Image Integrity Checker, which utilizes secure hash algorithms to ensure the authenticity of scientific imagery before it enters the training pool for future AI models.

Source: MIT Technology Review