Structure Prediction Roadmap: helixVision — Sovereign AlphaFold-Quality

Note: 🧬👁️ helixVision was previously known as coralForge. The codebase originated in syntheticChemistry/neuralSpring and is moving to sporeGarden/helixVision as a standalone product. Source code references may still use coral_forge module names during transition.

Sovereign protein structure prediction on consumer hardware. No cloud. No PyTorch. No CUDA. No data leaves the lab.

Last Updated: March 31, 2026
License: CC-BY-SA 4.0

The Goal

AlphaFold2/3 revolutionized structural biology. It is also a Google DeepMind product: requires PyTorch/JAX, CUDA, cloud APIs, and sends sequence data to external servers. For any lab handling pre-publication sequences, patient genomics, or proprietary protein engineering — this is a non-starter.

🧬👁️ helixVision is the 🔧🦎 ecoPrimals path to sovereign structure prediction: AlphaFold2/3-quality results running locally on consumer hardware in pure Rust, with full data sovereignty and cryptographic provenance.

Where We Are Now (Phase A–B: Complete)

The Isomorphism Proof

AlphaFold’s “novel” neural architecture decomposes into 6 universal primitives — the same primitives used everywhere else in machine learning:

This is the key insight: AlphaFold does not introduce any new category of computation. Every operation is a composition of GEMM, attention, normalization, nonlinearity, reduction, and gating — the same primitives BarraCuda already has as validated WGSL shaders.

Current Validation

154 checks passing (62 Python + 55 Rust + 37 GPU):

Precision: All tolerances named (e.g., CROSS_LANGUAGE 1e-10, FOLDING_EPS 1e-10). 15 DF64 WGSL shaders validated with max diff < 1e-6 vs f64 CPU (GELU 5.6e-7, SDPA 1.1e-7). f64 canonical throughout.

What This Means

Every building block of AlphaFold2 and AlphaFold3 has been:

1. Decomposed into universal primitives
2. Implemented in pure Rust
3. Validated against NumPy baselines to 1e-10 tolerance
4. Accelerated to GPU via BarraCuda WGSL shaders
5. Verified on consumer hardware (RTX 4070, Titan V)

The blocks are proven. The pipeline is next.

The Phases Ahead

Phase C — BarraCuda Integration (Next)

Wire 🧬👁️ helixVision primitives to BarraCuda canonical operations:

• GemmF64::execute_gemm_ex() for all GEMM ops
• GPU attention via existing BatchedScaledDotProduct
• GPU LayerNorm via existing BatchedLayerNorm
• Rayon for CPU parallelism on non-GPU paths

Target: ~50% wall-time reduction for the Pairformer block.

Phase D — End-to-End Pipeline

FASTA sequence → MSA search → Feature embedding
  → Evoformer × 48 → Structure module × 8
  → Coordinates → Confidence (pLDDT, PAE, pDE)
  → Provenance chain (BearDog signing, loamSpine cert)

Remaining components:

• MSA search: MMseqs2 port or sovereign k-mer search
• Template search: PDB template library (public, ~200 GB)
• Recycling loop: Evoformer output fed back N times
• Amber relaxation: Optional energy minimization post-prediction

Validation gate: LDDT > 0.7 on at least one CASP target (e.g., T1024).

Phase E — LTEE Structural Evolution Analysis

The primary scientific application. Lenski’s Long-Term Evolution Experiment: 75,000+ generations of E. coli under glucose-minimal constraint, frozen at 500-generation intervals. 🧬👁️ helixVision predicts structures at each timepoint and population.

Scale: ~8.3 million predictions (4,600 genes × 150 timepoints × 12 populations).

Questions only 🧬👁️ helixVision can answer:

• Do independently evolved populations converge on the same structural solutions? (Structural convergence beyond sequence convergence)
• Do structural changes follow power-law dynamics? (Constrained evolution prediction)
• Can Ara-3 citrate utilization precursors be identified retroactively from structural evolution?
• Do hitchhiker mutations have structural consequences?
• Does genome streamlining (gene loss) produce compensatory structural changes in retained genes?

Phase F — Standalone Publication

Standalone helix-vision crate on crates.io. Companion paper documenting the isomorphism proof, validation evidence, and LTEE application.

Performance Targets vs AlphaFold

The LTEE number is the most meaningful comparison: $83,000 and rate limits on cloud AlphaFold vs $1,000 electricity and unlimited local predictions. For a lab doing structural genomics at scale, this is the difference between “we can’t afford it” and “we already did it.”

Beyond AlphaFold: What Sovereign Structure Prediction Enables

Drug Discovery (Paper 12 + helixVision)

The Anderson-augmented MATRIX scoring pipeline (329/329 checks validated) currently uses published IC50 and pathway data. 🧬👁️ helixVision adds:

Drug candidate → helixVision structure → binding site geometry
  → Anderson tissue penetration model → combined score

Structure-based docking from sequence alone. No crystal structure required. No commercial docking software (Schrödinger ~$50K/yr, MOE ~$20K/yr).

Metagenomic Structural Census

💧♨️ wetSpring’s sovereign 16S pipeline identifies what organisms are present. 🧬👁️ helixVision predicts what their proteins look like:

Environmental sample → wetSpring 16S → community composition
  → Gene calling → helixVision structure → structural diversity index
  → Anderson W(structural) — disorder measured in protein space

This does not exist elsewhere. No one has applied Anderson localization to structural diversity of metagenomic communities.

Vaccine and Antigen Design

Structural prediction + provenance = a signed record of every design iteration from target selection to final construct:

Pathogen genome → helixVision structure → epitope identification
  → Antigen design → rhizoCrypt DAG (design history)
  → loamSpine cert (design certificate) → sweetGrass (attribution)

Enzyme Engineering (P≠NP Connection)

The P≠NP enzyme thesis (methodology/P_NP_ENZYME_THESIS.md) argues that enzymes are nature’s generative solutions to chemical NP problems. 🧬👁️ helixVision enables computational enzyme design:

Target reaction → retrosynthetic analysis → enzyme class identification
  → helixVision structure → active site engineering → validation

If you can predict structure from sequence, and you can design sequence for function, you have a sovereign enzyme engineering pipeline.

What Someone Else Could Pick Up

🧬👁️ helixVision is in 🧠♨️ neuralSpring (public, AGPL-3.0). The primitives are validated. Anyone with Rust and a GPU can:

1. Complete Phase C — wire BarraCuda GEMM to 🧬👁️ helixVision Evoformer (estimated 2–4 weeks for a competent Rust developer)
2. Build the MSA search — MMseqs2 is open-source; a Rust port is tractable (estimated 4–8 weeks)
3. Run Phase D validation — CASP targets are public, PDB is public, the pipeline is modular
4. Apply to their own domain — any lab with sequences and questions about structure can use the validated primitives

The ⚖️🔓 scyBorg license (AGPL-3.0) means: anyone who uses it must share their improvements. Every advance returns to the commons. The pipeline gets better for everyone, permanently.

Source: whitePaper/helixVision/ (20 documents), 🧠♨️ neuralSpring src/coral_forge/
Validation: 154/154 checks PASS (62 Python + 55 Rust + 37 GPU)
Repositories: syntheticChemistry/neuralSpring, ecoPrimals/barraCuda