> ## Content Index
> Fetch the complete content index at: https://www.stirup.si/llms.txt
> Use this file to discover other available public pages before exploring further.

# AlphaFold's Limitations and the Role of Data
- URL: https://www.stirup.si/alphafold-s-limitations-and-the-role-of-data-68a69483/
- Published: 2026-10-13T15:34:03.000Z
- Updated: 2026-10-13T15:34:03.000Z
- Description: AlphaFold's progress in protein structure prediction is constrained by the availability and quality of training data, highlighting the need for continued data curation and integration with domain-specific insights.
- Author: Satish Das
- Tags: AI, News, Stirup Research

## AlphaFold's Limitations and the Role of Data

AlphaFold has made significant strides in predicting protein structures, but its limitations are tied closely to the availability and quality of training data. Pushmeet Kohli of Google DeepMind and Sal Candido of Biohub note that while AlphaFold leveraged existing datasets like the Protein Data Bank (PDB), the lack of substantial data augmentation limited its potential. Organizations and scientists have invested heavily in constructing these datasets, making them invaluable but also a bottleneck for further progress. Kohli suggests that there may be practical limits to how much data can be generated or how large a model can be trained, given resource constraints. As more data becomes available, the challenge shifts from simply collecting it to curating and selecting the most relevant information for model training. These factors highlight that while AlphaFold represents a major breakthrough, it is not a complete solution to the protein folding problem, and ongoing data curation remains essential.

## The Bitter Lesson of AI Scaling and Its Implications

Pushmeet Kohli of Google DeepMind and Sal Candido of Biohub discuss the limitations of AlphaFold in solving protein folding, emphasizing that AI scaling alone is not sufficient for deep biological understanding. Kohli references the 'bitter lesson' of AI scaling, a concept originally articulated by Rich Sutton, which suggests that large-scale data and model training often outperform handcrafted knowledge. However, this approach has not fully resolved the complexities of protein folding, revealing gaps in current AI capabilities. Kohli stresses the need to rethink AI strategies for biology, acknowledging that while AlphaFold has made significant progress, it has not achieved a complete solution. Candido highlights the remaining challenges in accurately predicting protein behavior in dynamic environments. Both experts agree that AI must be complemented by domain-specific insights and experimental validation to advance the field.

## Beyond Protein Structure Prediction: The Next Steps in AI and Biology

AlphaFold has made significant strides in predicting protein structures, but as Pushmeet Kohli of Google DeepMind notes, science operates by isolating problems and making progress step by step. While AlphaFold has solved the challenge of predicting protein structure with remarkable accuracy, it has not yet addressed the broader complexities of protein biology. Sal Candido of the Biohub highlights that these models are useful but do not fully solve the problems most researchers aim to tackle. The next steps in AI and biology involve understanding how proteins function within cellular environments, interact with other molecules, and contribute to disease processes. These challenges require integrating structural predictions with functional and dynamic biological data, a task that remains ongoing. As Brandon Anderson points out, the media often suggests that protein structure prediction has been fully solved, but the reality is more nuanced. The field continues to evolve, with AI serving as a powerful tool rather than a complete solution.

## Sources and reporting notes

- [Latent Space](https://www.latent.space/p/biohub-deepmind?ref=stirup.si)

Based on published source material. No independent hands-on testing is claimed.