Semantic-level multimodal molecular learning inspired by biological concept formation via soft matching

Fig. 1 | From instance-level to semantic-level alignment in multimodal molecular representation learning. a, Instance-level (left) and semantic-level (right) alignment paradigm. b, An example of human perception, where the simultaneous presence of bread, meat, and vegetables naturally leads to the recognition of a hamburger (left); an illustration of the proposed semantic-level soft matching mechanism for molecular understanding (right). c, Classification performance comparison across multiple benchmark datasets for representative molecular representation models. Fig. 2 | Conceptual framework of the semantic-level multimodal molecular representation learning of SemMol. a, DCL construction. Initial semantic centers for 2D and 3D modalities are constructed from 1D, 2D, and 3D representations via intra-batch clustering and refined using mini-batch K-means with Exponential Moving Average (EMA) updates in the DCL. b, ACSM mechanism. b1, Positive sample generation: retrieve the nearest semantic centers from 2D/3D center libraries relative to the anchor and construct positive samples using similarity-weighted fusion. b2, Negative sample generation and debiasing: filter out easily distinguishable negatives using a similarity threshold while retaining hard negatives to improve discriminative ability. c, Semantic alignment training by using the constructed positive and negative samples.

The most important supplementary file is provided at the following link: https://github.com/3505675604/SemMol/blob/main/Supplementary_Materials/Supplementary%20Materials.pdf

🚀 Features

• Semantic-level learning: By introducing the anchor-center soft matching (ACSM) mechanism and the dynamic center library (DCL), cross-modal semantic alignment is achieved, avoiding the risk of overfitting caused by one-to-one instance matching. This method can simultaneously capture the global skeleton structure and local functional group characteristics of the molecule, improving the model's generalization ability and interpretability.
• Multi-Modal Learning: Integrates chemical structure, text, and other modalities for improved prediction.
• Flexible Task Support: Handles both classification and regression tasks with dynamic configuration.
• Advanced Pseudo-Pair Generation: Supports hard negative mining, adaptive temperature, and memory bank for contrastive learning.
• Streaming & Incremental Clustering: Online center library with streaming K-means and FAISS acceleration.
• Configurable & Reproducible: All settings managed via a single JSON config; supports experiment reproducibility.
• Extensible Architecture: Modular codebase for easy extension of models, data pipelines, and loss functions.
• Pretrained Model Integration: Easy download and usage of state-of-the-art pretrained models.

---

⚙️ Installation & Environment

1. Clone the repository

   git clone https://github.com/3505675604/SemMol.git

2. Create environment (recommended)

   conda env create -f environment.yml
   conda activate A

🏃‍♂️ Training, Fine-tuning, and Resources Training :Run the following commands to start pre-training:

   chmod +x /data/FL/Semol/scripts/start_Pre_DPP.sh
   /data/FL/Semol/scripts/start_Pre_DPP.sh

Fine-tuning :Fine-tune the pretrained model with:

   python finetune.py

Pre-trained Models and Datasets We provide the pre-trained SemMol model (trained on 1M molecules) and the datasets used for pre-training and downstream fine-tuning:

• Pre-trained Models (1M molecules): https://huggingface.co/Lin-Glory/SemMol_model
• Datasets (for pre-training and fine-tuning): https://huggingface.co/datasets/Lin-Glory/SemMol_datasets/tree/main/dataset

Note: Please download and extract all files before starting training or fine-tuning.

Outputs：All model checkpoints, logs, and training histories are saved in:

   Save_model/

---

📊 Datasets

-Input format: CSV files containing a smiles column and corresponding target column(s) -Supported benchmarks: -BBBP -ESOL -Lipophilicity -Tox21 -Data split strategies: -Scaffold split (default, chemically-aware) -Random split

🧩 Configuration

All experiment, model, and data settings are managed in config/config.json:

• Data: File paths, target columns, split type, normalization
• Model: Architecture, fusion strategy, dropout, pretrained paths
• Training: Epochs, batch size, learning rate, scheduler, seed
• Loss: Loss function, pseudo-pair and alignment weights
• Pseudo-Pair: Hard negative mining, memory bank, temperature
• Early Stopping: Patience, monitored metric

See in-file comments and descriptions for all options.

---

📈 Reproducibility -Fixed random seeds for all experiments -Centralized configuration management -Modular and extensible codebase for easy customization

📚 Citation If you use SemMol in your research, please cite the original authors:

@article{semmol,
  title={SemMol:Semantic-level multimodal molecular learning inspired by biological concept formation via soft matching},
  author={Anonymous},
  journal={arXiv preprint},
  year={202X}
}