Evaluating 2D and 3D Graph Encoders for Infrared-Based Organic Structure Verification

This repository contains the code to support the paper "Evaluating 2D and 3D Graph Encoders for Infrared-Based Organic Structure Verification".

Setup

make setup
source .venv/bin/activate
export PYTHONPATH="$PWD/src"

Create a .env file in the repository root. A minimal example is:

PASGEN_DIR=$PWD
CALCULATIONS_DIR=$PWD/pasgen_data
PROCESSED_DATA_DIR=$PWD/data/processed
ZENODO_RECORD=20069940

Data

The data and model archive is available from Zenodo record 20069940. The pipeline downloads it automatically unless pasgen_data/pasgen_upload.zip is already present.

The archive contains prediction JSON files and trained checkpoints for:

• mpnn
• painn
• ensemble
• chemprop_ir
• original_pasgen

The cross-validation results use the matching _10_folds directories.

The archive does not include experimental spectra. Those are reconstructed locally from the supplied JDX/mol files when the manuscript pipeline runs.

Reproduce the manuscript

make paper

This runs scripts/run_manuscript_pipeline.sh and then builds the LaTeX manuscript under manuscript/build/.

To run only the analysis from existing predictions:

SKIP_PREDICTIONS=1 make paper

To skip xTB/DFT comparisons:

SKIP_CALC_MODELS=1 make paper

Baselines

The manuscript includes Chemprop-IR, the original PASGeN model, and a contrastive learning baseline. Their entry points are:

scripts/install_baseline_deps.sh
scripts/run_chemprop_ir_benchmark.sh
scripts/run_original_pasgen_benchmark.sh
scripts/run_kanakala_clip_fixed_benchmark.sh

Chemprop-IR and the original PASGeN source snapshots are under third_party/. The contrastive learning baseline code is under src/pasgen/baselines/kanakala_s2s/.

Layout

src/pasgen/     core code
scripts/        workflow entry points
configs/        pipeline configuration
manuscript/     LaTeX sources, tables, and figures
assets/         metadata and embeddings
data/sample/    tiny smoke-test dataset
third_party/    external baseline source snapshots