Skip to main content
Back

SimplexFold Challenges AlphaFold2 by Teaching AI to See Protein Geometry in 3D

Chris Hayduk open-sources SimplexFold, an AlphaFold2 research fork that adds triangle and tetrahedron states to the protein folding trunk, guided by 150+ hours of GPT-5.5 autoresearch.

  • Chris Hayduk open-sourced SimplexFold, a research fork of AlphaFold2 that adds learnable triangle and tetrahedron states to the model trunk.
  • The core idea: proteins have three-body and four-body geometric constraints that pairwise edge representations alone struggle to capture efficiently.
  • A new SimplicialEvoformer block maintains sparse face tensors F_ijk and tetra tensors U_ijkl, with gated message passing back into the pair and single streams.
  • Hayduk used GPT-5.5's Goal Mode for 150+ hours of autonomous structural biology research to develop the architectural concepts behind the project.
  • The repo includes full PyTorch implementation, tiny/medium/full configs, a parameter-matched benchmark config, and a detailed ablation plan.
  • Released under MIT license; the scientific claim (higher-order cells improve folding accuracy) is still being benchmarked, not yet proven.

Protein folding models have been graph-based almost by default: residues are nodes, pairwise contacts are edges, and everything else gets approximated from there. SimplexFold, a new open-source research fork by Chris Hayduk, challenges that assumption by asking what happens when you give the model explicit, learnable states for triangles and tetrahedra formed by residue groups , not just pairs.

The project is also a real-world demonstration of a new kind of AI-assisted research workflow. Hayduk ran GPT-5.5's Goal Mode for over 150 hours straight, using it to explore topologically inspired architectural changes to AlphaFold2. The code and research vision that emerged from that process are now public under an MIT license.

Why pairs aren't enough

AlphaFold2 predicts protein 3D structure from sequence by integrating multiple sequence alignments (MSAs), structural templates, an iterative attention-based network called the Evoformer, and a structure module that uses Invariant Point Attention to model spatial relationships. The Evoformer is already doing sophisticated geometric reasoning , it uses triangle multiplication and triangle self-attention to make pair features more geometrically consistent. But there's a gap: those triangle operations write back into an edge tensor. They don't maintain persistent learned states for the triangle (i, j, k) or the tetrahedron (i, j, k, l) as first-class objects.

SimplexFold's core bet is that proteins are not just collections of pairwise contacts. Folding is full of three-body and four-body constraints: backbone angles, torsions, sheet geometry, hydrophobic-core packing, and local residue-contact motifs. A pair feature can say "residue