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

Figure 1：Based on biomimetic principles, the SemMol model utilizes a "Dynamic Center Library" (DCL) and a soft matching mechanism to deeply learn molecular structural and semantic features from large-scale data. Experiments demonstrate that the model consistently outperforms existing state-of-the-art methods in both molecular representation learning and property prediction. Figure 2：The SemMol model projects multi-dimensional molecular representations into a unified space and constructs a Dynamic Center Library (DCL), utilizing the soft matching mechanism (ACSM) to achieve one-to-many associations between molecules and semantic centers. Compared to traditional one-to-one matching, this method enables learning from a broader knowledge distribution, significantly enhancing the model's generalization ability and discriminative power.

Pre-trained Models

Our pre-trained SemMol mode can all be downloaded via Baidu Netdisk: Download link: https://pan.baidu.com/s/1nIcHZoealZG2kRBem0949w Extraction code: abcd Please download and extract the files before training or fine-tuning.

🚀 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.

---

🗂️ Project Structure

├── config/           # Centralized configuration (config.json)
├── core/             # Core algorithms: center library, clustering, pseudo-pair logic
├── data/             # Example datasets (CSV, SMILES, targets)
├── datasets/         # Data loading, splitting (scaffold/random), and processing
├── img/              # Images and figures for reports or publications
├── model/            # Model components: embedding, fusion, projector, pseudo-pair
├── model_config/     # Model download and usage instructions
├── weight/           # Pretrained weights and training history
├── train.py          # Main training & evaluation script
├── environment.yml   # Conda environment for full reproducibility
└── README.md         # This documentation

---

⚙️ Installation & Environment

1. Clone the repository
2. Create environment (recommended)

   conda env create -f environment.yml
   conda activate A

3. Install additional dependencies (if needed)

   pip install -r requirements.txt

---

📊 Datasets

• Place your CSV datasets in data/. Each file should contain a smiles column and the appropriate target column (see config/config.json).
• Supported datasets: BBBP, ESOL, Lipophilicity, Tox21, etc.
• Data splitting: Scaffold split (chemically-aware) and random split are both supported.

---

🧩 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.

---

🏃‍♂️ Training & Evaluation

Basic usage:

python train.py \
    --task-type classification \
    --data-path path \
    --target-column Class/reg \
    --batch-size 32 \
    --lr 5e-4 \
    --epochs 50 \
    --hard-negative-k -1

Advanced examples: (see config/config.json for more)

python train \
    --task-type regression \
    --data-path path \
    --target-column   -- \
    --normalize-targets \
    --batch-size 64 \
    --lr 5e-4 \
    --epochs 50 \
    --hard-negative-k 32 \
    --hard-negative-ratio 0.3

Model weights and training history are saved in weight/ after each run.

---

🏗️ Extending the Platform

• Add new datasets: Place in data/ and update config/config.json.
• Custom models: Implement in model/ and reference in config.
• New data splits or augmentations: Add to datasets/.
• Custom loss or metrics: Extend in core/ or train.py.

---

📥 Pretrained Models

See model_config/File Description for download links and usage instructions for pretrained models.

---

🧪 Reproducibility & Best Practices

• All random seeds, splits, and hyperparameters are controlled via config.
• Use environment.yml for full environment reproducibility.
• For large-scale or production runs, see the production_mode and debug_mode settings in config.

---

📚 References & Citation

If you use this platform in your research, please cite the original authors and relevant papers.

---

🤝 Contributing & Support

• Pull requests and issues are welcome!
• For questions, suggestions, or bug reports, please open an issue.

---

Contact: For collaboration or consulting, please reach out via GitHub or email.