knf 1.0.8

Automated Descriptor Engine for SNCI, SCDI, and 9D KNF

NCIForge

NCIForge is a physics-informed descriptor pipeline for non-covalent interactions. It computes KNF descriptors, SNCI/SCDI metrics, and KUID-family indexing artifacts from molecular structure files using xTB + NCI backend + KNF post-processing.

Current package version: 1.0.8
Release milestone tag: v1

What It Covers

• Generate KNF feature vectors (f1..f9) for single files and batches
• Generate full KUID identifiers (instance-level)
• Generate KUID-Intensive identifiers (topology-family level)
• Build reusable lookup artifacts (prefix, reverse, and bridge indexes)
• Recompute universal KUID outputs from existing batch folders (--universal-kuid)
• Generate canonical atlas bundle files (--atlas-bundle)

Requirements

• Python >=3.8
• External tools in PATH:
  • xtb
  • obabel
• Multiwfn only when using --nci-backend multiwfn

Optional:

• torch for Torch NCI backend (CPU or CUDA)

Install

Fastest local install (from this repo):

Windows PowerShell:

./scripts/install_nciforge.ps1

macOS/Linux:

bash ./scripts/install_nciforge.sh

This opens an interactive CLI wizard that asks:

• install scope: global or local virtual environment
• PyTorch mode: cpu, gpu, or skip
• whether to auto-setup external dependencies (xtb, obabel, etc.)

Manual source install:

git clone https://github.com/Prasanna163/KNF.git
cd KNF
pip install -e .

Install with Torch extra:

pip install -e ".[torch-nci]"

CLI Commands

Primary CLI:

• nciforge

Compatibility alias (still supported):

• knf

Example Runs

Single molecule:

nciforge example.mol

Batch run:

nciforge ./molecules --processing multi --workers 4 --force

GPU run (Torch CUDA):

nciforge ./molecules --gpu

--gpu enables smart GPU routing by default:

• Runs Torch NCI on CUDA first.
• If a molecule hits CUDA OOM, that molecule is re-routed to CPU automatically.
• The next molecule retries GPU automatically.
• Uses memory-friendlier float32 by default.

CPU run (no GPU):

nciforge ./molecules --cpu

--cpu forces CPU execution. If PyTorch is unavailable, KNF auto-falls back to the Multiwfn CPU backend (when Multiwfn is installed).

Split into batches:

nciforge ./molecules --batches 4

Recompute universal KUID calibration from existing batch outputs:

nciforge ./existing_runs --universal-kuid

Generate canonical atlas submission bundle:

nciforge ./molecules --atlas-bundle

Outputs

Single-run outputs include:

• knf.json (contains kuid and kuid_intensive sections)
• kuid_calibration.json

Batch root outputs include:

• batch_knf.json
• batch_knf_unified.csv
• kuid_calibration.json
• kuid_intensive_calibration.json
• kuid_prefix_index.json
• kuid_topology_prefix_index.json
• kuid_instance_prefix_index.json
• kuid_full_topology_bridge.json
• kuid_full_topology_bridge.csv
• kuid_reverse_index.json
• kuid_reverse_index.csv
• kuid_topology_reverse_index.json
• kuid_topology_reverse_index.csv
• kuid_family_stats.json
• kuid_family_stats.csv
• kuid_intensive_family_distribution.csv
• kuid_intensive_family_distribution.png

With --water, water-suffixed variants are emitted (for example *_water.json, *_water.csv).

Atlas Submission Bundle

When --atlas-bundle is supplied, NCIForge writes:

• submission_bundle/atlas_submission.csv
• submission_bundle/manifest.json

If prior batch outputs already exist, running --atlas-bundle can parse those existing CSV outputs and generate the bundle without recomputing the KNF pipeline. Legacy batch CSV names are still supported for upgrade compatibility.

After bundle creation, submission runs clean up auxiliary analysis/index artifacts in the same results root to keep exports lightweight.

atlas_submission.csv includes:

• molecule_name, charge, spin
• f1..f9, SNCI, SCDI, SCDI_variance
• backend, device, xtb_version, knf_core_version
• nci_grid_spacing, nci_grid_padding, water_mode
• KUID_raw, KUID_Cluster
• KUID_Intensive_raw, KUID_Intensive_Cluster
• instance_hash (sha256(f1..f9,charge,spin,xtb_version,nci_grid_spacing,nci_grid_padding)[:8])

Incremental Resume

When batch_knf_unified.csv already exists (legacy names are also supported, including older atlas_submission.csv), and --force is not set, existing rows are reused and only new files are processed.

Docker

Container workflows are documented in README.DOCKER.md.

Releasing

Release steps are documented in RELEASE.md.

License

MIT (LICENSE)