license: cc-by-nc-4.0
This dataset is constructed for supervised fine-tuning (SFT) of Y-Mol, a multimodal foundation model for drug discovery. It contains high-quality instruction–response pairs that span a wide range of drug-related tasks, designed to align the model with expert-like capabilities in molecular reasoning and decision-making.
The dataset covers four key tasks:
- Molecular Property Prediction
Synthetic instruction–response pairs are generated using expert tools (e.g., ADMETlab, SwissADME) to simulate predictions of pharmacokinetic and physicochemical properties such as solubility, permeability, toxicity, and metabolic stability.
- Molecule Generation
Includes prompts and reference outputs for generating novel molecular structures conditioned on specific constraints such as bioactivity profiles, physicochemical rules, or scaffold constraints.
- Drug–Target Interaction (DTI) Prediction
Contains natural language prompts describing candidate drugs and targets, paired with binary or affinity-based interaction labels sourced or simulated from public databases and predictive models.
- Drug–Drug Interaction (DDI) Prediction
Captures interactions between drug pairs with both mechanistic and effect-level annotations, aiming to train the model to reason about combinatorial effects or contraindications.
Each instruction–response pair is designed to be LLM-friendly and structured to encourage step-by-step reasoning, enabling Y-Mol to generalize across multiple biomedical and cheminformatics tasks.