InVirtuoGen: Uniform-Source Discrete Flows for Drug Discovery

📄 Paper: Refine Drugs, Don't Complete Them: Uniform-Source Discrete Flows for Fragment-Based Drug Discovery

InVirtuoGen is a generative framework for molecule design and property optimization, built on discrete flow matching with a uniform source. It supports de novo generation, fragment-constrained generation, and reinforcement learning–based property optimization.

---

🧩 Available Checkpoints

• invirtuo_gen.ckpt Standard model for general use.

• invirtuo_gen_big.ckpt Larger model variant for optimal performance.

---

🚀 Quick Usage

from huggingface_hub import hf_hub_download
import torch

# Download a checkpoint
ckpt_path = hf_hub_download("invirtuolabs/InVirtuoGen", filename="invirtuo_gen.ckpt")

from in_virtuo_gen.models import InVirtuoFM
model = InVirtuoFM.load_from_checkpoint(ckpt_path, map_location="cpu")

python -m in_virtuo_gen.generate \
    --ckpt_path checkpoints/invirtuo_gen.ckpt \
    --num_samples 1000 \
    --batch_size 200

---

📊 Tasks Supported

• De Novo Molecule Generation – high quality and diverse outputs
• Fragment-Constrained Generation – motif extension, linker design, scaffold decoration, superstructure generation
• Target Property Optimization – property-driven search with reinforcement learning
• Lead Optimization – structure-based optimization using docking

---

🔗 Full Reproducibility

For environment setup, benchmarks, and reproduction of results from the paper, please see the GitHub repository.

---

📄 Citation

If you use this work, please cite:

@misc{kaech2025refinedrugsdontcomplete,
    title={Refine Drugs, Don't Complete Them: Uniform-Source Discrete Flows for Fragment-Based Drug Discovery},
    author={Benno Kaech and Luis Wyss and Karsten Borgwardt and Gianvito Grasso},
    year={2025},
    eprint={2509.26405},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url={https://arxiv.org/abs/2509.26405},
}

---

📧 Contact

📬 Email: [email protected]

---