chapter twelve

12 Transformers for Protein Structure Prediction

 

This chapter covers

  • Predicting protein structure directly from sequence
  • Why proteins fold into stable 3D shapes
  • Generating structures with modern generative models
  • Capturing long-range residue interactions with attention
  • Using protein language models in downstream biology tasks

In Chapter 11, we used graph neural networks (GNNs) to predict how small molecules bind to proteins. We treated protein structures as fixed scaffolds, with assistance from multiple sequence alignment and position scoring matrices to determine contacts between residues and form a protein graph. As we saw, determining drug-target binding affinity depends critically on knowing three-dimensional protein structures.

The importance of protein structure prediction extends to drug discovery as a whole. For example, when investigating why a cancer becomes resistant to treatment, we may need to understand how mutations could alter the shape of the target protein. Structure determines function, and without structure, drug design gravitates closer to guesswork.

12.1 Conjoined Problems: Structure Prediction & Protein Design

12.1.1 Protein Structure Essentials

12.1.2 Why Protein Folding is a Hard Modeling Problem

12.2 SimpleFold: An End-to-End Example

12.2.1 Stage 1: Configuration

12.2.2 Stage 2: Loading Pretrained Models

12.2.3 Stage 3: Inference Pipeline

12.2.4 Stage 4: Flow Matching

12.2.5 Stage 5: From Sequence to Structure

12.2.6 Comparing Approaches: PLM-Based vs. MSA-Based

12.2.7 Retrospective

12.3 Protein Language Models

12.3.1 Why Transformers? From Local to Global Attention

12.3.2 The Encoder-Only Transformer Architecture

12.3.3 Training a Small Protein Language Model

12.3.4 Downstream Applications: Antibody Classification

12.3.5 The Evolutionary Scale Modeling (ESM) Family of Models

12.4 Summary

12.5 References