MoLA: Molecular Multimodal Layerwise Adaptive Network

Official/author-maintained partial code release for:

MoLA: Molecular multimodal layerwise adaptive network for molecular property prediction
Jiayi Li, Zihang Zhang, Zhenyu Lei, Jiujun Cheng, Lianbo Ma, Cong Liu, Shangce Gao
Knowledge-Based Systems, Volume 338, 115563, 2026.

• DOI: https://doi.org/10.1016/j.knosys.2026.115563
• ScienceDirect: https://www.sciencedirect.com/science/article/pii/S0950705126003059
• Share Link: https://authors.elsevier.com/c/1md~S3OAb9Kb5p

Highlights

• Multimodal molecular representation learning with:
  • molecular graph features
  • Morgan fingerprint features
  • SMILES sequence features
  • optional MoLFormer embeddings
• Layerwise adaptive fusion with cross-layer attention.
• Reproducible training script for MoleculeNet classification tasks.
• Built-in ablation switch for MoLFormer:
  • --use-molformer (default)
  • --no-molformer

Repository Structure

.
|-- main_Cla_MoLA.py        # training/evaluation entry
|-- model_MoLA.py           # MoLA architecture
|-- prepare_molformer_embeddings.py  # build MoLFormer embeddings (reference utility)
|-- prepare_MoLFormer.py    # backward-compatible launcher
|-- environment.yml         # base conda environment specification
|-- setup_env.sh            # one-command environment bootstrap
|-- utils.py                # metrics (PRC-AUC / ROC-AUC)
|-- datasets/
|   |-- featurized/         # DeepChem cached data
|   `-- Molformer_OUTPUT/   # optional MolFormer embeddings
|-- result_Cla/             # training logs
`-- weights/                # checkpoints

Environment

Requirements

• conda (or mamba)
• Python 3.10
• CUDA 12.1 (for GPU training with the default setup)
• Linux/macOS shell (or Git Bash on Windows) for setup_env.sh

Quick Setup (recommended)

bash setup_env.sh
conda activate chem

Data Preparation

The training script uses DeepChem MoleculeNet loaders and expects:

1. DeepChem cache directory:
   • ./datasets/featurized/...
2. (Optional) MoLFormer embedding file:
   • ./datasets/Molformer_OUTPUT/<dataset_name>/Molformer_Emb_2025.h5
3. H5 keys:
   • train_fp, valid_fp, test_fp

Build MoLFormer embeddings (reference utility)

This repository provides prepare_molformer_embeddings.py to compute embeddings and save:

• datasets/Molformer_OUTPUT/<dataset_name>/Molformer_Emb_2025.h5

Notes:

• This script is a project helper for MoLA and is provided for reference.
• It must be used together with the official IBM MoLFormer release:
  • https://github.com/IBM/molformer

Run:

python prepare_molformer_embeddings.py

Quick Start

1) Train with MoLFormer embeddings

python main_Cla_MoLA.py \
  --use-molformer \
  --run-times 1 \
  --epochs 300 \
  --batch-size 256

2) Train without MoLFormer embeddings

python main_Cla_MoLA.py \
  --no-molformer \
  --run-times 1 \
  --epochs 300 \
  --batch-size 256

Reproducibility Notes

• Dataset loaders currently configured:
  • bace_classification, bbbp, clintox, muv, sider, tox21
• main_Cla_MoLA.py does not require --datasets; it runs the built-in dataset list.
• Seed control:
  • --seed (default: 2025)
• Typical output paths:
  • logs: result_Cla/<model_name>/result_<model_name>_<run_id>.txt
  • checkpoints: weights/<model_name>_<dataset>_<run_id>.pth

Main Arguments

• --model-name: experiment name prefix
• --run-times: repeated runs
• --epochs: maximum epochs
• --batch-size: mini-batch size
• --embed-dim: hidden width
• --num-layers: encoder depth
• --max-len: SMILES max length
• --fp-bits: Morgan fingerprint bits (default 2048)
• --lr: learning rate
• --weight-decay: weight decay
• --scheduler-patience: LR scheduler patience
• --early-stop-patience: early stop patience
• --device: cuda or cpu
• --use-molformer / --no-molformer: whether to use MoLFormer embeddings

Current Release Scope

This repository is a partial release and currently includes limited prepared data (mainly BACE-related files in this snapshot). It focuses on core model/training logic and reproducible experiments.

Citation

If you use this codebase, please cite:

@article{li2026mola,
  title   = {{MoLA}: Molecular multimodal layerwise adaptive network for molecular property prediction},
  author  = {Li, Jiayi and Zhang, Zihang and Lei, Zhenyu and Cheng, Jiujun and Ma, Lianbo and Liu, Cong and Gao, Shangce},
  journal = {Knowledge-Based Systems},
  volume  = {338},
  pages   = {115563},
  year    = {2026},
  doi     = {10.1016/j.knosys.2026.115563}
}