Researchers from the Skolkovo Institute of Science and Technology posed a practical test to AlphaFold, the renowned AI system developed by Google DeepMind. The challenge, described in a study published in PLOS ONE, shifted the focus from predicting protein structure to assessing how tiny genetic changes affect protein stability. This shift highlighted a core question in structural bioinformatics: can an AI that excels at one task also illuminate other, closely related aspects of molecular behavior?
Structural bioinformatics explores the architectures of proteins, RNA, and DNA and their interactions with other molecules. The information gathered through these studies serves as the backbone for drug development and the design of proteins that do not exist in nature. A foundational assumption in the field has long been that if one knows a protein’s amino acid sequence, it should be possible to infer its three‑dimensional shape in a living system, and from that shape, predict its function and stability. This idea has driven decades of research, with progress measured in incremental steps toward a more complete, predictive understanding of biomolecules.
Half a century after the inception of this line of inquiry, AlphaFold has demonstrated a remarkable ability to predict protein structures from sequences. Developed by Google DeepMind, the system has been celebrated for capturing many aspects of how proteins fold, offering researchers a powerful computational proxy for the physical process. For some, this achievement suggested that the neural network might be learning the underlying physics of protein folding rather than merely memorizing data or patterns. The excitement surrounding these capabilities has led to speculation that AI could eventually answer the broader questions that still challenge structural bioinformatics.
To test the limits of AlphaFold in a concrete way, scientists at Skoltech’s School of Bioinformatics for High School Students, alongside other researchers, constructed a focused task. They deployed AlphaFold, running on the Zhores supercomputer, to predict the effect of single amino acid substitutions on protein stability. The experiment required replacing one amino acid in a given protein with another and evaluating whether the mutation would stabilize or destabilize the protein, and to what extent. The aim was not to predict folding for a new sequence but to quantify the mutation’s impact on the protein’s structural integrity.
When the predictions were compared with experimental measurements, the results did not align. AlphaFold produced estimates that, in some cases, diverged significantly from observed stability changes caused by single mutations. The discrepancy underscored a crucial limitation: while AlphaFold excels at structural prediction for unmutated sequences, extending that competence to predictive tasks about stability after mutations is not straightforward. The study’s authors emphasized that AlphaFold’s creators never claimed broad applicability beyond structure prediction based on amino acid sequences, a point they reiterated in their conclusions.
The researchers used the divergence as a diagnostic signal rather than a critique of AI. It highlighted a broader truth about machine learning in biology: models trained to perform one well‑defined task may struggle when asked to address related but fundamentally different questions. The study therefore serves as a reminder that progress in AI does not automatically translate into universal problem solving across all facets of structural biology. It also calls for careful, task‑specific evaluation of AI tools before extending their use to new biological predictions.
As the field moves forward, scientists acknowledge both the promise and the current boundaries of AI in structural biology. Advances in computational methods, improved datasets, and more nuanced modeling approaches may eventually enable reliable predictions of how individual mutations influence protein stability. For now, the Skoltech work provides a clear, practical demonstration: the leap from predicting how a protein folds to forecasting how a point mutation alters stability is not yet bridged by AlphaFold’s existing framework. The takeaway is not discouragement but a measured, evidence‑driven approach to integrating AI into biological research.