Spectral Prediction

This repository contains hybrid quantum neural networks for spectral predictions. The architecture is an extension of chemprop-IR described in the paper Message Passing Neural Networks for Infrared Spectral Predictions and available in the Chemprop-IR GiHub repository.

Docker

Docker provides a nice way to isolate the code and environment. To install and run our code in a Docker container, follow these steps:

1. Install Docker from https://docs.docker.com/install/
2. cd /path/to/project
3. docker build -t hqnn .
4. docker run -it --gpus all -v $(pwd):/project hqnn /bin/bash

Note: To enable GPU support, you must use a CUDA-enabled base image in the Dockerfile. You can replace the default line with a CUDA/cuDNN/Ubuntu base image that matches your environment. You can find options at https://hub.docker.com/r/nvidia/cuda/tags. Make sure the CUDA version also matches the PyTorch version you are using. You can refer to the compatibility chart here: https://pytorch.org/get-started/previous-versions/

Training

To train a model, run:

cd /path/to/project/
python ./chempropIRZenodo/chempropIR/train.py --data_path <path> --dataset_type spectra --checkpoint_path <checkpoint> --save_dir <dir>

where <path> is the path to a CSV file containing a spectral dataset,dataset_type is set to spectra. <checkpoint> is the path to a pretrained model file (e.g., model.pt). This argument is used when you want to fine-tune an existing model instead of training from scratch, and <dir> is the directory where model checkpoints will be saved.

The example spectral dataset format in a CSV file:

smiles,400,402,404,...,3996,3998,4000
OCc1cc(F)cc(F)c1,0.000423112823390177,0.0004251490935022775,...,0.0004340529928113568

We recommend using the training configuration provided in /path/to/project/recommended_config.json. You can specify a config file using --config_path <config>. We also suggest using a GPU, which can be selected using the --gpu <index> where <index> indicates the GPU device ID to use (if available).

To use the HQNN (Hybrid Quantum Neural Network) model, you must include the --qnn flag.
This enables quantum layers in the network. You can also control the number of quantum hidden layers using the --qnn_layer <int> argument. If the --qnn flag is not specified, a standard FFNN (Feedforward Neural Network) model will be used instead.

An example script for training is provided at: /path/to/project/scripts/run_train.sh To use it, run the following commands:

cd /path/to/project
./scripts/run_train.sh

Predicting

To load a trained model and make predictions, run chempropIRZenodo/chempropIR/predict.py and specify:

• --test_path <path> Path to the data to predict on. Format this file as a .csv file with only a single column, with the header row with the entry smiles and every subsequent row entered with the SMILES you would like to predict.
• --checkpoint_path <path> Path to a model checkpoint file (.pt file).
• --preds_path Path where a CSV file containing the predictions will be saved.
• --qnn If the model checkpoint is HQNN model. An example script for training is provided at: /path/to/project/scripts/run_pred.sh To use it, run the following commands:

cd /path/to/project
./scripts/run_predict.sh

We suggest using a GPU, which can be selected using the --gpu <index> where <index> indicates the GPU device ID to use (if available).