A Linguistic Model for Protein Structure Generation

Abstract

Computational generation of protein backbone geometry is a critical upstream step in de novo protein design, but the existing state of the art approaches typically requires large pretrained structure prediction networks, significant computing in frastructure, or complex probabilistic frameworks. Here we present a lightweight autoregressive generative model based on a Long Short-Term Memory network with a Mixture Density Network output head, trained directly on backbone phi and psi torsion angle sequences which were extracted from a non-redundant subset of the Protein Data Bank. Angles were encoded as sine-cosine pairs to handle circular periodicity, and a variable-length end-of-sequence mechanism was incorporated to enable generation of proteins of realistic and variable length. The Mixture Density Network output head with eight Gaussian mixture components was introduced to prevent mode collapse onto the mean of the Ramachandran distribution. This enabled the model to generate sequences with realistic secondary structure composition. A total of 1000 novel backbone scaffolds were generated and reconstructed as three dimensional polyalanine structures using EZScafold, then scored using Rosetta after energy minimisation. Generated scaffolds achieved significantly lower normalised Rosetta total scores than a control set constructed by randomly sampling angles from the training data pool (p < 0.001), demonstrating that the model captures meaningful sequential dependencies in backbone geometry beyond the marginal angle distribution alone. This demonstrate that a simple, accessible, and interpretable recurrent architecture can be used as an effective backbone generation engine pro ducing geometrically and energetically meaningful scaffolds ready for downstream sequence design pipelines.