Researchers at the Centre for Genomic Regulation published a perspective paper in Nature Machine Intelligence on May 11, 2026, calling for increased transparency in protein language models. The authors advocate for “explainable AI” to ensure that AI-driven protein design is safe, reliable, and understandable for scientists applying these tools in biotechnology.
Protein language models (pLMs) have emerged as critical tools for engineering proteins with specific, useful properties. These AI systems can design entirely new structures that have never existed in nature, offering a path toward solving significant global challenges. Potential applications include the synthesis of enzymes capable of absorbing carbon dioxide from the atmosphere and the creation of catalysts that reduce energy consumption and toxic waste in industrial settings.
The Black Box Problem in Protein Design
Despite their utility, pLMs often function as black boxes
, a term describing systems where the internal logic and decision-making processes are opaque to the user. This lack of transparency creates a significant hurdle for scientists who must decide whether a model’s prediction is reliable, biased, or safe for real-world application. As these models begin to influence actual decisions in the biotechnology sector, the inability to audit their reasoning increases operational risk.
The researchers at the Centre for Genomic Regulation (CRG) argue that this opacity is not merely a technical inconvenience but a safety concern. When a model predicts a protein structure or function without providing a traceable rationale, researchers cannot easily verify the biological plausibility of the result. This creates a dependency on the AI’s output without a corresponding understanding of the underlying biological mechanisms.
Erosion of Biological Transparency
The shift toward AI-driven design has coincided with a decline in the transparency that previously defined the field. Earlier physics-based models allowed scientists to understand the forces and interactions driving protein folding and catalysis. In contrast, the speed of pLM development has outpaced the scientific community’s ability to interpret how these models reach their conclusions.
Dr. Noelia Ferruz, Group Leader at the CRG
Dr. Ferruz, the corresponding author of the paper, notes that the transition to AI has actually resulted in a loss of some of the clarity provided by older methods. The risk is the creation of high-powered tools that operate beyond human comprehension, making them difficult to trust fully in sensitive biological environments.
In some ways, we have even lost part of the transparency that characterized physics-based models.
Dr. Noelia Ferruz, Group Leader at the CRG
Implementing Explainable AI in Biotechnology
To address these risks, the CRG researchers analyze the application of explainable AI
(XAI). XAI consists of techniques and methods designed to make the decisions of artificial intelligence understandable, interpretable, and trustworthy for human operators. By integrating XAI into protein language models, scientists could potentially see which parts of a protein sequence the AI is prioritizing and why it believes a certain structure will yield a specific function.
The perspective paper serves as a call to action for the broader research community. The authors argue that for AI-driven protein design to be viable in the long term, systems must be made more transparent and secure. This involves not only improving the models themselves but also developing new frameworks for validating AI predictions against known biological laws.
The push for transparency comes at a time when AI in protein science is accelerating. Other recent developments in the field include the use of lab-in-the-loop frameworks to generate millions of data points in short timeframes and the effort to make AI-driven design tools more accessible to biologists globally. However, the CRG findings suggest that accessibility and speed are insufficient if the resulting tools remain fundamentally opaque.
The goal for the next phase of development is to bridge the gap between AI efficiency and biological insight. By prioritizing explainability, the scientific community can ensure that the pursuit of new enzymes and catalysts does not come at the expense of safety or scientific rigor. The move toward transparent AI is presented as a necessary step to turn powerful predictive tools into reliable instruments for biotechnological advancement.
