End-to-End Protein Normal Mode Frequency Predictions Using Language and Graph Models and Application to Sonification

The prediction of mechanical and dynamical properties
of proteins
is an important frontier, especially given the greater availability
of proteins structures. Here we report a series of models that provide
end-to-end predictions of nanodynamical properties of proteins, focused
on high-throughput normal mode predictions directly from the amino
acid sequence. Using neural network models within the family of Natural
Language Processing and graph-based methods, we offer atomistically
based mechanistic predictions of key protein mechanical features.
The models include an end-to-end long short-term memory (LSTM) model,
an end-to-end transformer model, a graph-based transformer model,
and an equivariant graph neural network. All four models show exceptional
performance, with the graph-based transformer architecture offering
the best results but at the cost of requiring a graph structure as
input. Conversely, the LSTM and transformer models offer end-to-end
sequence-to-property prediction capabilities, providing efficient
avenues for protein engineering, analysis, and design. We compare
our results against published data based on a Principal Neighborhood
Aggregation graph neural network, revealing that the transformer model
offers better performance while also being able to predict a large
set of the first 64 normal mode frequencies, simultaneously. The use
of the end-to-end transformer model may facilitate other downstream
applications through the use of transfer learning, and it offers a
comprehensive prediction of dynamical properties without any structural
knowledge, directly from the amino acid sequence. We demonstrate a
potential application in scientific sonification, where the normal
mode frequencies are transposed to generate audible signals for a
detailed analysis of subtle changes of protein sequences.