CoCoFold2 v2.0.0 — Fixed-frame projection by default

CoCoFold2 v2.0.0 makes fixed-frame projection the default for new single-GPU and component-parallel particle-refinement runs. Compared with v1.0.1, rendered structures now remain in a shared 3D reference frame across views and components, rather than using the historical per-image recentering and normalization.

What changed since v1.0.1

• Fixed-frame projection for single- and multi-GPU refinement. Component projections use the same 3D origin before they are summed.
• Explicit projection origin. New fixed-frame runs require --projection-origin X_A Y_A Z_A in the placed CIF/map coordinate frame. The parameter guide and examples explain how to choose it.
• Legacy reproduction remains available. The initial CoCoFold2 manuscript results used the historical projection behavior. Specify --projection-frame legacy to reproduce that setting. Exact resumes retain their saved projection settings; use a new warm-start run to change frames.
• Optional GMM width initialization. --gmm-sdev-init-mode molmap can initialize widths from a requested physical resolution. The historical legacy width initialization remains the default.
• Updated documentation and project site. The GitBook user guide, command references, tutorials, and project homepage now reflect the current workflows and projection settings.
• Expanded public checks for projection-frame behavior, input validation, restart compatibility, and component-parallel calculations.

Upgrade note: Commands that relied on the old projection default must now either supply a verified --projection-origin for fixed-frame refinement or explicitly select --projection-frame legacy. A map header alone does not establish that the placed CIF and particle poses share the same coordinate frame.

CoCoFold2 v1.0.1

This release updates the project website, documentation, and citation
metadata following the publication of the CoCoFold2 preprint.

Updates

• Added the manuscript DOI and BibTeX citation to the README.
• Updated CITATION.cff with the manuscript citation and software version.
• Added the project website and tutorials, with selected paper figures.
• Clarified installation, validation scope, and release information.

No changes to model implementation or runtime behavior.

Links

• Preprint: https://doi.org/10.65215/LTSpreprints.2026.09.15.000338
• Project website: https://jwliaomath.github.io/CoCoFold2/

CoCoFold2 v1.0.0

CoCoFold2 v1.0.0

CoCoFold2 provides deterministic latent refinement of protein structure predictions using cryo-EM particle observations and a frozen Protenix-v1 diffusion prior.

Included workflows

• Initial prediction and diffusion-cache generation, with input preflight and failure summaries.
• Single-GPU refinement with configurable seeds, optional GMM amplitude/width freezing, and per-chain rigid placement.
• Component-parallel refinement with one cache per GPU, Independent/Contextual input preparation, and merged CIF output.
• Structured command/configuration/metric records, synchronized CIF/checkpoint output, complete-epoch resume, and legacy warm-start/export support.
• English installation and usage guides, a 7ZDT/7ZD5 map-derived smoke/refine example, a Contextual 6ZBH 1+3 example, and public CPU/Gloo regression tests.

Established scientific defaults are preserved: seed 42, legacy RNG behavior, GMM learning enabled, and learning rates 0.01/0.01/0.005 for latent bias/amplitudes/widths. The small example explicitly freezes both GMM parameter groups. Training requires a compatible aligned reference CIF; template-free export is available when cache topology is sufficient.

Installation

See the installation guide. The validated full environment uses Linux, Python 3.11, Protenix 1.0.2, PyTorch 2.7.1/CUDA 12.6, and NumPy 2.4.1. CUDA extension compilation requires a compatible toolkit/compiler; GNU 12.2.0 was used for the validated fresh installation.

Weights, common resources, sequence/MSA inputs and particle stacks are supplied separately. Random, fine-tuning and the main heterogeneity experiments are outside this release.

Validation and limits

The main-branch GitHub Actions run checks the independent public source, two Gloo CPU processes and source packaging. CPU tests use analytic substitutes where specified and do not establish real-model accuracy.

Author-provided server acceptance includes single-structure export/restart features, the 7ZDT/7ZD5 smoke and 1000-particle refinement, Contextual two-GPU checks (39/39), and the first four complete epochs of the Contextual 6ZBH case (69/69). A fresh Linux installation separately passed environment/CPU checks and a real-weight smoke retry (10/10) after compiler setup.

These are specific tested workflows, not an assertion that every documented option combination has been executed. A long Independent run and fresh-environment prediction generation are outside this acceptance. The fresh-installation smoke's reported FRC decreased; its functional pass does not demonstrate structural improvement. Map agreement and RMSD remain manual assessments. See validation scope.

Source and checksums

CoCoFold2-v1.0.0-source.zip contains the 118 public files from the tagged merge commit 6ec6adc1d877eb116300a414cddbda278bbada1d. source_manifest.json records the commit and each file's SHA-256. Verify the attached files after downloading them into one directory:

sha256sum -c SHA256SUMS.txt

GitHub's automatically generated source archives are separate downloads and are not covered by this checksum file. No weights or experimental particle stacks are included.

The citation title is CoCoFold2: scalable latent refinement of diffusion-based protein structure predictions from limited-particle cryo-EM data. The unavailable manuscript DOI is omitted. The code retains its Apache-2.0 license; upstream software and data retain their respective terms.

CoCoFold2 v0.1.0-beta — Research preview

Initial research preview of CoCoFold2 accompanying the manuscript.

This release provides the cryo-EM particle-guided refinement workflow and
a tutorial based on a real dataset. It is intended for research evaluation
and is not yet considered production-ready.