T-conditioned_molecular_representation

This repository implements temperature-conditioned molecular representations for predicting organic solubility across varying temperatures.

1. Code

   • preprocessing
     • csv_to_pkl.py: Converts .csv data into .pkl format for efficient loading
     • custom_dataset_single.py: Dataset class for Chemprop-based models
     • fastprop_csv_dataset.py: Dataset class for Fastprop-based models
     • feature_vectors.py: Feature construction for molecular graph representation
   • models
     • chemprop: model architecture information (Chemprop, Chemprop w/FiLM)
     • fastprop: model architecture information (Fastprop, Fastprop w/FiLM)

2. 47_pairs_result_graph

   • Chemprop
   • Fastprop

---

Code Overview

1. Data Preprocessing (code/preprocessing)

This folder prepares raw solubility data for training. First,

• Converts SMILES into molecular graph representations using RDKit
• Constructs solute–solvent pairs
• Includes temperature (T) and target solubility (logS)
• Saves processed dataset as a .pkl file

Second,

• Custom PyTorch dataset for solute–solvent systems
• Molecular graphs encoded via message-passing features
• Includes:
  • Atom features
  • Bond features
  • Molecular descriptors

2. Model Architecture (code/models)

The models are based on Directed Message Passing Neural Network (D-MPNN) and descriptors with temperature conditioning vial FiLM. This allows continuous modulation of molecular representations with respect to temperature.

Key Components

• Molecular Encoders

  • Separate encoders for solute and solvent

• Temperature Encoding

  • Scalar temperature is expanded using Gaussian RBF

• FiLM (Feature-wise Linear Modulation)

  • Learns feature-wise scaling (γ) and shifting (β) based on temperature
  • Applied to both solute and solvent representations

• Prediction Head

  • Outputs predicted log solubility

---

Application Study Results

The file 47_pairs_result_graph.zip contains visualization results for 47 solute–solvent pairs.

Each plot shows:

• Experimental solubility (ground truth)
• SAFT-γ Mie predictions
• Baseline model predictions (Chemprop, Fastprop)
• T-conditioned model predictions (Chemprop w/FiLM, Fastprop w/FiLM)

File naming: 'SoluteName_SolventName.png'

---

Citation

If you use this work, please cite: [Will be updated]