Molexar (Molecular Exalted Architect)

Molexar is a unified multimodal molecular foundation model for drug design that supports unconditional generation and multi-condition generation, including molecular-property conditioning, molecular-pharmacophore conditioning, protein-sequence conditioning, and protein-pocket conditioning, while also accommodating arbitrary custom conditions.

Resources

• Official website: https://molexar.com
• Molexar source code: https://github.com/fairydance/Molexar
• Fragment-SELFIES source code: https://github.com/fairydance/Fragment-SELFIES
• Unconditional base model: https://huggingface.co/fairydance/molexar-10m-base
• Universal multi-condition model: https://huggingface.co/fairydance/molexar-10m-omni

Features

• Fragment-SELFIES molecular representation with BRICS fragment tokens
• Unconditional base-model generation and conditional generation in one model class
• Continuous, discrete, and vector condition encoders
• Gemma2 features including RoPE, grouped-query attention, sliding-window attention, and softcapping
• Full-parameter SFT for multi-condition molecular generation
• Training and inference entrypoints for single-GPU and multi-GPU workflows through Accelerate

Installation

git clone https://github.com/fairydance/Molexar.git
cd Molexar

conda create -n molexar python=3.13
conda activate molexar

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu130
python -m pip install -e ".[train,data]"

python -c "import fragment_selfies; import molexar; print('Molexar environment ready')"

Fragment-SELFIES is required for SMILES conversion and generated molecule decoding. Molexar depends on the published fragment-selfies package, so pip install -e . installs it automatically from PyPI. Install Molexar in editable mode in every environment that runs training, inference, or auxiliary embedding scripts.

Repository Layout

Molexar/
├── docs/                     # architecture, data, training, and inference guides
├── examples/train/           # training wrappers and shared shell helpers
├── images/                   # image assets used by documentation
├── models/configs/base/      # config_10m_256h_16l.json
├── models/tokenizer/         # tokenizer.json
├── scripts/                  # training, inference, tokenizer, data, and embedding utilities
└── src/molexar/              # package source

Datasets are not included in this repository. Released model files are hosted on Hugging Face at fairydance/molexar-10m-base and fairydance/molexar-10m-omni. Pass local dataset and model paths explicitly through CLI flags or environment variables.

Released Models

• fairydance/molexar-10m-base - unconditional base model for de novo and fragment-constrained generation
• fairydance/molexar-10m-omni - universal multi-condition model for molecular-property, pharmacophore, protein-sequence, and protein-pocket conditioning

Base Pretraining

DATA_PATH=/path/to/pretrain.fragment_selfies \
OUTPUT_DIR=/path/to/output/pretrain_base_10m_256h_16l \
examples/train/pretrain_base.sh

Equivalent core command:

python scripts/run_training.py --task pretrain \
  --config_path models/configs/base/config_10m_256h_16l.json \
  --tokenizer_path models/tokenizer \
  --train_data_path /path/to/pretrain.fragment_selfies \
  --output_dir /path/to/output/pretrain_base_10m_256h_16l \
  --batch_size 32 \
  --epochs 2

Multi-Condition SFT

BASE_MODEL_PATH=/path/to/pretrain_base_10m_256h_16l/final_model \
MOLECULE_CONTEXT_PATH=/path/to/molecule_context.fragment_selfies \
PROPERTIES_PATH=/path/to/molecule_properties.csv \
PHARMA_FP_PATH=/path/to/gobbi_pharma_fps.npy \
SAIR_INDEX_DIR=/path/to/SAIR/index \
SAIR_STRUCTURES_DIR=/path/to/SAIR/structures_processed \
SAIR_LIGAND_FRAGMENT_SELFIES_PATH=/path/to/SAIR/ligands.fragment_selfies \
PLINDER_ROOT=/path/to/PLINDER/2024-06/v2 \
PLINDER_INDEX_DIR=/path/to/PLINDER/2024-06/v2/index \
PLINDER_LIGAND_FRAGMENT_SELFIES_PATH=/path/to/PLINDER/ligands.fragment_selfies \
OUTPUT_DIR=/path/to/output/sft_universal_multi_10m_256h_16l \
examples/train/sft_universal_multi.sh

Use NUM_PROCESSES, MIXED_PRECISION, USE_FSDP, and FSDP_STRATEGY environment variables to adapt the wrappers to your hardware.

Inference

python scripts/run_inference.py --mode base \
  --model_path /path/to/model/final_model \
  --tokenizer_path models/tokenizer \
  --num_samples 10 \
  --convert_to_smiles \
  --canonical \
  --output_file /path/to/output/base_samples.jsonl \
  --output_format jsonl

Conditional generation accepts JSON, NPY, PKL, direct scalar flags, reference SMILES for molecular properties or pharmacophore fingerprints, protein sequences for ESM embeddings, and pocket PDB files for GVP conditioning. See docs/inference.md for examples.

Documentation

• docs/architecture.md - model architecture and condition injection
• docs/data_preparation.md - expected input file formats
• docs/training.md - pretraining and SFT commands
• docs/inference.md - base and conditional generation examples

Citation

@misc{lin2026molexarunifiedmultimodalmolecular,
      title={Molexar: A Unified Multimodal Molecular Foundation Model for Drug Design},
      author={Haoyu Lin and Yiyan Liao and Jinmei Pan and Xinliao Ling and Luhua Lai and Jianfeng Pei},
      year={2026},
      eprint={2606.25865},
      archivePrefix={arXiv},
      primaryClass={q-bio.BM},
      url={https://arxiv.org/abs/2606.25865},
}

License

Molexar is released under the MIT License. See LICENSE for details.