Appendix D. Technical Deep Dive into Protein Structure Prediction

This appendix gives a technical foundation for protein structure prediction by first grounding the problem in protein chemistry, folding physics, and the geometric constraints that any model must respect. It explains how amino acid properties drive folding, why hydrogen bonds stabilize helices and sheets, how torsion angles and Ramachandran-allowed regions define valid backbone conformations, and why folding is shaped by hydrophobic collapse, entropy, and kinetic traps. It also connects misfolding to disease and reviews older approaches such as molecular dynamics, Rosetta-style sampling, homology modeling, and threading, setting up the case for modern deep learning methods.

The text then walks through SimpleFold in detail, from protein representations and tokenization to transformer components, training, and inference. It describes how sequences and structures are encoded, how flow matching turns random noise into coordinates through iterative denoising, and how the model uses ESM-2 embeddings, timestep conditioning, rotary position embeddings, QK normalization, and DiT blocks to combine sequence and geometric information. A hierarchical encoder-trunk-decoder design processes atoms locally, residues globally, and atoms again for velocity prediction, while training combines flow-matching loss with a smooth LDDT objective, plus timestep resampling, gradient clipping, EMA weights, and stochastic sampling for ensemble generation and confidence estimation.

Finally, the appendix situates SimpleFold within the broader structure-prediction landscape. It contrasts AlphaFold2’s MSA-driven, highly engineered geometric inductive biases with SimpleFold’s simpler, more general generative approach, and explains when confidence metrics such as pLDDT, PAE, RMSD, TM-score, GDT-TS, and LDDT should be trusted or treated cautiously. It closes by showing how the field is moving toward diffusion-based all-atom models such as AlphaFold3, Boltz-1, Boltz-2, and RFdiffusion, which extend prediction from single proteins to complexes, ligands, affinities, and de novo design, making structure prediction one component in a larger drug-discovery pipeline rather than the end goal itself.

Figure D.1 Illustrates the rotational elements of the backbone alongside a Ramachandran plot specific to glycine. These allowed regions correspond precisely to the secondary structures we discussed earlier. There is a prominent region in Ramachandran space centered around Φ ≈ -60° and Ψ ≈ -45° that corresponds to right-handed alpha-helices. Another large region, centered around Φ ≈ -120° and Ψ ≈ +120°, corresponds to beta-sheets. A smaller region corresponds to left-handed helices, which are rare due to steric clashes with most sidechains but accessible to glycine. Note that the R-group for Glycine consists of just one H atom attached to the alpha Carbon. Other AAs have different R-groups.

Figure D.2 Overview of SimpleFold’s architecture from “SimpleFold: Folding Proteins is Simpler than You Think”. [9]

Figure D.3 AlphaFold2 model architecture from “Highly accurate structure prediction with AlphaFold.”

Figure D.4 Flow matching visualized in the protein structure prediction setting. (A) The biological folding process transforms an unfolded chain (t=0) into a folded structure (t=1). (B) The computational flow matching process mimics this by learning a velocity field that guides a distribution of random noise (source) to match the distribution of real protein structures (target) over a virtual timeline.

Figure D.5 Illustration of grouping operations in SimpleFold from “SimpleFold: Folding Proteins is Simpler than You Think.”

Figure D.6 Illustration of ungrouping operations in SimpleFold from “SimpleFold: Folding Proteins is Simpler than You Think.”

Figure D.7 Distribution of resampled timestep compared to uniform distribution from “SimpleFold: Folding Proteins is Simpler than You Think.”

Figure D.8 Illustration of pLDDT module training from “SimpleFold: Folding Proteins is Simpler than You Think.”

FAQ