Transformer-based Protein Structure Prediction from Sparse Experimental Constraints

Structural biology is still struggling with structure
prediction from sequence, especially when researchers have
poor experimental data to compare to. To accurately predict
structures, the study presents a transformer-based deep
learning architecture that we expect will approximate
geometry-aware attention from sparse experimental data only,
as in incomplete cryo-EM maps or NMR-based distance
limitations. With an average error of 1.85 Å RMSD and a TMScore
of 0.92 on benchmark cases with sparse experimental
constraints, the suggested model performed better than other
state-of-the-art techniques such as AlphaFold2 (2.15Å RMSD,
0.89 TM-Score; sampled 61.2% of the input experimental
constraints) and DMPfold2 (2.48Å RMSD, 0.86 TM-Score;
sampled 79.5% of the input experimental constraints). In
addition, the systems enable highly feature-rich encrypted
inference on edge-devices using model compression, packed
encoding, and hardware-aware optimizations bringing the
average inference time down to ~30 seconds per protein. These
results demonstrate the potential of fusion transformers with
experimental priors to enable rapid, accurate, and privacysensitive
protein structure prediction and to enable a family of
real-time biomedical applications.