AI Struggles to Decode Gene Activity Despite Promises of Precision

Recent research highlights significant limitations in artificial intelligence (AI) tools when applied to the complex field of biology. A study conducted by researchers from Heidelberg, including Constantin Ahlmann-Eltze, Wolfgang Huber, and Simon Anders, found that AI software designed to predict gene activity under various conditions failed to outperform basic prediction methods.

While AI has shown promise in specific biological applications, such as designing enzymes to break down plastics and creating proteins to counteract snake venom, the notion that AI can universally analyze biological data is premature. The study examined several AI packages aiming to forecast changes in gene activity, yet their performance was disappointing.

Complexities of Gene Activity

The human genome consists of approximately 20,000 genes, but not all are active in every cell. Gene activity varies based on cell type and environmental conditions. For instance, some genes remain consistently active, while others are conditionally activated. Over the years, extensive research has mapped gene activities across various cell types and conditions, providing a foundation for training AI models.

The researchers utilized what are known as “single-cell foundation models,” which are trained on gene activity data from individual cells, aiming to predict how gene activity might change when certain genes are altered. This approach is crucial as it can offer insights into the complex relationships between different genes and their activities.

Despite the advanced nature of these AI models, the study revealed they struggled significantly in predicting intricate changes in gene interactions. When examining the effects of altering one or two genes using CRISPR technology, the AI systems did not surpass the predictions made by a very simple model that assumed no changes would occur.

Findings and Implications

The researchers conducted experiments where they activated one or two genes and used the resulting data to assess the AI’s predictive capabilities. They compared the AI predictions against two basic models: one that predicted no changes and another that proposed additive effects from simultaneous gene activations. The results were clear: all AI models exhibited a prediction error that was substantially higher than the additive baseline.

In their conclusion, the researchers noted, “As our deliberately simple baselines are incapable of representing realistic biological complexity yet were not outperformed by the foundation models, we conclude that the latter’s goal of providing a generalizable representation of cellular states and predicting the outcome of not-yet-performed experiments is still elusive.”

While the findings indicate that AI’s current capabilities in understanding gene activity are limited, this does not suggest that future advancements are impossible. The study serves as a reminder of the challenges inherent in biological research and underscores the necessity for continued experimentation and refinement of AI applications in this field.

As excitement grows around AI’s potential, this research emphasizes the importance of cautious optimism. The complexity of biology cannot be underestimated, and the path to developing AI systems that can reliably interpret biological data remains a significant challenge. The full details of the study are published in Nature Methods in 2025.