FARScore: Molecular Synthetic Accessibility Predictor

Fragment Assembly autoRegressive based synthetic accessibility scorer to accelerate drug discovery

🎯 What Makes FARScore Different

FARScore revolutionizes synthetic accessibility prediction through Fragment Assembly autoRegressive pretraining. Unlike traditional approaches that directly learn synthesis patterns, FARScore first masters molecular construction fundamentals—understanding how molecules are assembled from fragments—then applies this knowledge to predict synthetic accessibility.

Two-Stage Learning:

• Stage 1: Pretrain on 9.2M unlabeled molecules to learn molecular assembly patterns
• Stage 2: Finetune on 800K labeled molecules for synthetic accessibility prediction

This mirrors human chemical intuition: experienced chemists understand molecular construction before assessing synthetic difficulty.

✨ Key Features

• Easy Integration - Simple CSV input/output format
• Batch Prediction - One-click synthetic accessibility scoring
• High Accuracy - Achieves SOTA performance on multiple test sets with key metrics including accuracy, AUROC and specificity.

🌐 Online Service

Instant molecular synthesis prediction in the cloud. Simply upload your CSV file with SMILES and receive AI-powered synthetic accessibility scores in seconds.

🚀 Quick Start

1. Installation

    # Clone repository
    git clone https://github.com/simmzx/FARScore.git
    cd ../FARScore

    # Create environment and install dependencies
    conda create -n FARScore python=3.8
    conda activate FARScore
    pip install -r requirements.txt

2. Prepare Data

Create CSV file with "smiles" field:

molecule_id	smiles
Palbociclib	CC1=C(C(=O)N(C2=NC(=NC=C12)NC3=NC=C(C=C3)N4CCNCC4)C5CCCC5)C(=O)C
(+)-Eburnamonine	[C@]12(C3=C4CCN1CCC[C@@]2(CC(=O)N3C1C4=CC=CC=1)CC)[H]

3. Run Prediction

CSV File Mode

    python farscore.py --input_file example.csv

Direct SMILES Mode

    # Single molecule
    python farscore.py --smiles "CCO"
    # Multiple molecules
    python farscore.py --smiles "CCO" "CC(=O)O" "c1ccccc1"

4. View Results

Output file will contain FARScore values:

molecule_id	smiles	farscore
Palbociclib	CC1=C(C(=O)N(C2=NC(=NC=C12)NC3=NC=C(C=C3)N4CCNCC4)C5CCCC5)C(=O)C	0.9453
(+)-Eburnamonine	[C@]12(C3=C4CCN1CCC[C@@]2(CC(=O)N3C1C4=CC=CC=1)CC)[H]	0.0286

FARScore Interpretation:

• Close to 1: Easy to synthesize

• Close to 0: Hard to synthesize

• Threshold 0.5: Binary classification cutoff

📖 Advanced Usage

Custom Pretraining and Finetuning task

Pretrain Model

    python farscore_pretrain.py \
        --dataset smiles.txt \
        --vocab fragment.txt 

Note: smiles.txt contains unlabeled molecules, fragment.txt is a fragment vocabulary generated by ./scripts/utils/mol/cls.py from smiles.txt for fragment assembly autoregressive pretrain.

Finetune Model

    python farscore_finetune.py \
        --input_model_file gnn_pretrained.pth \
        --dataset dataset.csv

Note: gnn_pretrained.pth is a model saved in pretraining stage, dataset.csv contains labeled molecules for finetune on specific downstream task.

🔧 Requirements

• Python 3.8-3.10
• CUDA-enabled GPU (recommended)
• Key dependencies: PyTorch, RDKit, DGL, DeepChem

📄 Citation

If this program is useful to you, please cite our paper:

📧 Contact

For questions, please contact: Xiang Zhang (Email: zhangxiang@simm.ac.cn)

---

🌟 Like this project? Give us a Star