mistral-7b-drug-prots

<div align="center">

mistral-7b-drug-prots

PharMistral · Stage I

</div>

Mistral-7B after joint pre-training on unpaired protein sequences and drug-like molecules. This is the intermediate checkpoint; the final sequence-conditioned model is mistral-7b-drug-prots_pair.

From PharMistral: Drug generation for novel protein targets by Mistral Large Language Model (Computer Methods and Programs in Biomedicine Update, 2026). Code: PharMistral.

Training

Stage I of a two-stage process on Mistral-7B: joint pre-training on unpaired protein and small-molecule sequences, then sequence-conditioned fine-tuning on high-affinity protein–ligand pairs from ChEMBL-26. Sequence length 1,500, batch size 4 × 4 gradient-accumulation steps on 4 GPUs, learning rate 5e-5 with cosine decay, fill-in-the-middle rate 0.5. Use the matching tokenizer drug-prot-tokenizer-v2-mistral. Training scripts: GitHub.

Artifacts from this paper

Intended use

Research use only. Generated molecules are computational hypotheses: they have not been synthesised or tested experimentally, and they require the usual medicinal-chemistry, safety and regulatory evaluation before any practical use.

Citation

@article{bendjeddou2026pharmistral,
  title   = {PharMistral: Drug generation for novel protein targets by Mistral Large Language Model},
  author  = {Asma Bendjeddou and Kamyar Zeinalipour and Dalal Bardou and Marco Maggini and Franco Scarselli and Monica Bianchini},
  journal = {Computer Methods and Programs in Biomedicine Update},
  volume  = {10},
  pages   = {100257},
  year    = {2026},
  doi     = {10.1016/j.cmpbup.2026.100257}
}