ProtCross Desktop 0.2.0

ProtCross Desktop 0.2.0

Unsigned testing build. These Windows and macOS desktop artifacts are official ProtCross release assets for the 0.2.0 desktop workflow, but they are not code-signed or notarized. Windows SmartScreen and macOS Gatekeeper may warn that the publisher cannot be verified.

Downloads

• Windows 10/11 x64: ProtCross_Desktop_0.2.0_x64-setup.exe
• macOS 12+: ProtCross_Desktop_0.2.0_macos.dmg
• Checksums: SHA256SUMS.txt

What is included

• Tauri + React desktop interface.
• Mol* protein viewer with pocket center and residue table support.
• Local CPU prediction backend setup flow.
• Optional GPU or existing conda backend detection path.
• Bundled ProtCross checkpoint/PCA assets from the stable 0.1.2 asset bundle.

ESM-C weights are not bundled. Users must confirm the ESM-C license in the app and then download or import the ESM-C weights separately.

ProtCross 0.1.3

Summary

ProtCross 0.1.3 is a release-preparation and user-facing hardening release. It focuses on making clean installs, prediction outputs, training entry points, runtime assets, and repository layout reliable before broader distribution.

Highlights

• protcross is now the canonical package name, with deprecated evopoint_da compatibility aliases retained for 0.1.x users.
• protcross predict now writes a downstream-friendly result package by default: annotated structure, extended scores TSV, pocket JSON, and summary JSON.
• Prediction outputs now include pocket centers, residue lists, probabilities/scores, residue counts, asset version, threshold, clustering metadata, truncation metadata, and output paths.
• Structure writeback is safer for docking/MD workflows: scored protein residues are handled explicitly, altloc atoms are updated consistently, HETATM collisions are avoided, and multi-model warnings are recorded.
• Runtime asset resolution is centralized and now supports manifest metadata, SHA verification, refresh, offline mode, explicit overrides, and clearer cache behavior.
• Original 0.1.0-era paper reproduction scripts are archived under reproduction/legacy/; the main README now focuses on maintained 0.1.3 workflows.
• Root-level runtime assets and placeholder data are removed from source control. Assets are downloaded, supplied explicitly, or read from cache.
• Release packaging now includes source package data, CI, wheel/sdist smoke checks, and deprecated import compatibility tests.

Validation

• PR CI on main: passed
• conda run -n pyg5070 pytest -q: 95 passed
• python -m build: passed
• python -m twine check dist/*: passed
• Clean wheel install and CLI help smoke: passed
• PyPI upload: https://pypi.org/project/protcross/0.1.3/

Notes

This package release ships code only. The default runtime asset bundle remains the stable 0.1.2 checkpoint/PCA bundle and is installed through protcross setup-assets or automatically during prediction unless offline mode is requested.

ProtCross 0.1.2

Release checkpoint update for external prediction and benchmark use.

Assets:

• protcross-0.1.2-binding-moad-final.ckpt: Binding MOAD-trained release checkpoint. SHA256 ccb56884b21402a027bfae9d4779f38c8f534513d980a96d7cd78c9931748b65.
• pca_esmc_128_binding_moad_0.1.2.pkl: matching PCA reducer. SHA256 0f4e11806a622642c07dad539cec4216030220c1b5f3fc44c7926a2f6bca4d62.

Use this checkpoint for practical ProtCross prediction and benchmark reports. Use the 0.1.1 best-epoch=59.ckpt asset only for reproducing the original paper setup.

ProtCross 0.1.1

ProtCross 0.1.1 engineering-focused reproducibility release.

Highlights:

• Installable PyPI package metadata and console entry points.
• Lightweight single-structure prediction API and CLI that writes scores to PDB B-factors.
• One-command runtime asset setup with checkpoint and PCA assets from this release, plus ESM-C weights from Hugging Face.
• Refactored src package layout, compatibility wrappers, English documentation/logging, examples, and pytest coverage.

Runtime assets attached here:

• best-epoch.59.ckpt (saved locally by the installer as best-epoch=59.ckpt)
• pca_esmc_128.pkl

ESM-C weights are not redistributed in this release. protcross setup-assets downloads them from EvolutionaryScale's Hugging Face model repository.