appendix-d

Appendix D. Technical Deep Dive into Protein Structure Prediction

 

This appendix provides comprehensive technical details for readers who want to reproduce, extend, or deeply audit the approaches described in Chapter 12. Each section is designed to be consulted independently as a reference. Every model we’ll discuss (SimpleFold end to end, the ESM-2 protein language model that conditions it, AlphaFold2, and the newer all-atom systems surveyed at the close) presents a different answer to one question: given a protein sequence (or a set of interacting molecules), what three-dimensional structure should we hand to a downstream drug-discovery task such as docking, binding-site analysis, or hit triage, and how far should we trust it?

Our roadmap is as follows: we’ll cover the biophysics that constrain any solution, then a complete SimpleFold implementation, then how to train and use protein language models for several applications, and finally the modern landscape (AlphaFold2 and AlphaFold3, the open Boltz models, and diffusion-based design) that situates all of it in current practice.

D.1 Protein Biophysics and Chemistry

This section expands on the structural biology concepts introduced in Chapter 12, providing the detailed chemical and physical foundations necessary for a deep understanding of protein folding.

D.1.1 Amino Acid Chemistry

D.1.2 Hydrogen Bonding Mechanics

D.1.3 Torsion Angles and the Ramachandran Plot

D.1.4 The Physics of Folding

D.2 SimpleFold End-to-End Example Details

D.2.1 Protein Representation Formats

D.2.2 Stage 1: Configuration

D.2.3 Stage 2: Loading Pretrained Models

D.2.4 Stage 3: Inference Pipeline

D.2.5 Stage 5: Executing Protein Structure Prediction

D.2.6 Alignment-based Validation

D.2.7 Metrics, Metrics, Metrics

D.2.8 Understanding CAMEO and CASP Benchmarks

D.3 Training Protein Language Models

D.3.1 Tokenization: Converting Proteins to Numbers

D.3.2 Scaled Dot-Product Attention

D.3.3 Encoding Position Information

D.3.4 Encoder Layer: Attention + Feed-Forward

D.3.5 Training our Protein Language Model

D.3.6 When to Train Your Own Model

D.4 ESM-2 Case Study: Predicting Mutation Effects

D.5 Advanced SimpleFold Architecture

D.5.1 Invoking the Bitter Lesson in Structural Biology

D.5.2 Flow Matching: From Noise to Native Structure

D.6 Timestep Conditioning with Diffusion Transformer Blocks

D.6.1 SwiGLU Activation Functions

D.6.2 Rotary Position Embeddings

D.6.3 QK Normalization

D.6.4 Diffusion Transformer (DiT) Blocks

D.7 Deconstructing SimpleFold: Encoder-Trunk-Decoder