Tokyo University of Science Develops ‘DeepAFM’ AI Method for Protein Motion Analysis

Key Takeaways

• DeepAFM denoises HS‑AFM images with ~0.1 nm error
• Model classifies 19 SecA conformations with 93.4% accuracy
• Transfer learning extends method to other proteins without retraining
• Synthetic HS‑AFM dataset generated from molecular dynamics simulations
• Approach aligns with upcoming exascale platforms like Fugaku NEXT

Pulse Analysis

The breakthrough of AlphaFold in 2018 proved that artificial intelligence can predict static protein structures with near‑experimental accuracy, reshaping structural biology. Yet proteins are dynamic machines, constantly shifting shape to perform their functions. High‑speed atomic force microscopy captures these motions at the single‑molecule level, but the raw data are riddled with noise and temporal distortion, limiting quantitative analysis. DeepAFM bridges this gap by training a convolutional network on synthetic HS‑AFM images derived from molecular dynamics simulations, teaching the model to separate true structural signals from artefacts.

In the inaugural study, the TUS team focused on SecA, a bacterial translocase that toggles between closed and wide‑open states. By generating millions of labeled images representing 19 distinct conformations, the network learned to both denoise the input and predict the underlying shape. Test results showed sub‑nanometer fidelity (≈0.1 nm) and a 93.4% correct‑state identification rate, outperforming traditional fitting methods that often overfit noisy data. When applied to real experimental HS‑AFM frames, DeepAFM’s predictions matched independent biochemical measurements, confirming its practical utility.

Beyond the immediate performance gains, DeepAFM signals a broader shift toward AI‑augmented microscopy in drug discovery and biophysics. Accurate, high‑throughput conformational mapping can reveal transient binding pockets and allosteric sites, informing lead optimization. The method’s transfer‑learning capability means new protein targets can be analyzed with minimal additional training, reducing time‑to‑insight. As exascale systems like Fugaku NEXT come online, the computational demands of generating synthetic datasets and training deeper models will become trivial, paving the way for routine, AI‑driven interpretation of dynamic molecular imaging across the life‑sciences ecosystem.