helixVision — Self-Hosted Structure Prediction

Protein structure prediction in pure Rust f64 — no cloud, no PyTorch, no CUDA. AlphaFold primitives validated. strandGate capacity: 20-30 structures/day.

Repository: sporeGarden/helixVision (absorbed coralForge)
License: scyBorg (AGPL-3.0-or-later + ORC + CC-BY-SA 4.0)
Formerly: coralForge (syntheticChemistry/coralForge, now archived)
Compute Capacity: strandGate (RTX 3090) can predict 20-30 structures/day via Nest Atomic CAS pipeline


What It Is

helixVision is self-hosted protein structure prediction running locally on consumer hardware in pure Rust, with full f64 precision, complete data ownership, and cryptographic provenance. No cloud APIs. No PyTorch. No CUDA SDK. No data sent to Google.

Implemented All 6 AlphaFold primitives are implemented and individually validated (154 checks, 1e-10 tolerance vs NumPy). Architecture-ready End-to-end pipeline (FASTA → structure → confidence) is designed but not yet wired. “AlphaFold-quality” refers to the target — primitive-level parity is demonstrated, full-pipeline parity is not yet benchmarked.


The Isomorphism

AlphaFold’s architecture decomposes into 6 universal primitives — the same primitives BarraCuda already has as validated WGSL shaders:

AlphaFold OperationPrimitive Decomposition
Triangle multiplicationBatched outer product (GEMM) + sigmoid gating + reduction
Triangle attentionScaled dot-product attention + pair bias + softmax
Invariant Point AttentionQ·K^T/√d attention + L2 distance + softmax
Diffusion denoising (AF3)Scale + add per step
Confidence heads (pLDDT, PAE)Linear (GEMM) + softmax + weighted sum

helixVision does not introduce new computation. Every operation is a composition of primitives that already exist, are tested, and run on consumer GPUs.


How It Composes

LayerWhatPrimal
MathWGSL f64 shaders for all 6 primitivesbarraCuda
CompilationWGSL → native GPU binary (NVIDIA + AMD)coralReef
DispatchHardware discovery, execution, routingToadStool
ProvenanceEd25519 signing of every predictionBearDog
HistoryAppend-only prediction logloamSpine
OrchestrationPipeline routing via Neural APIbiomeOS

The pipeline: FASTA sequence → MSA search → Feature embedding → Evoformer × 48 → Structure module × 8 → Coordinates → Confidence (pLDDT, PAE, pDE) → Provenance chain.


Performance Targets Architecture-ready

These are design targets, not benchmarked results. Primitive-level validation is complete; end-to-end pipeline benchmarks are pending Phase C–D.

MetricCloud AlphaFoldhelixVision target
Precisionf32 (PyTorch default)f64 (native or DF64)
Cost per prediction~$0.01 (cloud API)~$0.0001 (electricity)
LTEE full analysis (8.3M predictions)~$83,000~$1,000 (6 months, 4× RTX 4070)
Data sovereigntyData sent to GoogleData stays local
ProvenanceNoneEd25519 signed, full chain
DependenciesPyTorch, JAX, CUDARust + wgpu (zero C deps)

Validation Status

PhaseStatusScope
A–B ReproducedAll AlphaFold2/3 primitives decomposed, implemented in Rust, validated to 1e-10 vs NumPy, GPU-accelerated
C PlannedWire BarraCuda GEMM to Evoformer operations
D PlannedEnd-to-end pipeline (FASTA → structure → confidence → provenance)
E PlannedLTEE structural evolution analysis (8.3M predictions)
F PlannedStandalone helix-vision crate on crates.io

Full roadmap: Structure Prediction Roadmap


What It Enables

  • Drug discovery: Structure-based docking from sequence alone (no $50K/yr Schrodinger license)
  • Metagenomic structural census: Predict protein structures for entire microbial communities
  • Vaccine/antigen design: Signed provenance chain from target selection to final construct
  • Enzyme engineering: Computational enzyme design via structure prediction + P≠NP enzyme thesis
  • LTEE analysis: 8.3M structure predictions across 75,000 generations of E. coli evolution

See also: Structure Prediction Roadmap, neuralSpring for validation evidence, barraCuda for the math engine.