Cyanex / DrDTI-Reasoner

🧬 DrDTI-Reasoner

A biomedical LLM fine-tuned for Drug–Target Interaction (DTI) prediction and reasoning.

Built on Meta LLaMA-3 8B (4-bit quantized) and adapted via LoRA (Low-Rank Adaptation) on structured bioactivity datasets, DrDTI-Reasoner is designed to support computational drug discovery, bioactivity prediction, and molecular reasoning research.

---

📋 Table of Contents

• Model Details
• Capabilities
• Input Format
• Output Format
• Training Data
• Evaluation Results
• Limitations
• Intended Use
• Roadmap

---

⚙️ Model Details

---

🔬 Capabilities

DrDTI-Reasoner is trained to handle the following tasks:

• DTI Classification — Predict whether a drug interacts with a given protein target
• Bioactivity Estimation — Classify compounds as Active or Inactive
• Potency Approximation — Estimate pXC50 values where applicable
• Molecular Interpretation — Parse and reason over SMILES molecular representations
• Target Analysis — Interpret protein targets from UniProt IDs or gene names
• Assay-Aware Reasoning — Incorporate assay metadata (mechanism, technology, mode) into predictions
• Biological Explanation — Generate short, human-readable justifications for predictions

---

🧪 Input Format

The model expects structured biomedical input in the following format:

Drug (SMILES):       <SMILES string>
Target protein:      <UniProt ID or protein name>
Assay mechanism:     <optional>
Assay technology:    <optional>

Note: Input formatting quality directly affects prediction accuracy. Always validate SMILES strings and use standard UniProt identifiers when possible.

Example Input

Drug (SMILES):       CC1=CC=C(C=C1)S(=O)(=O)N
Target protein:      P00533 (EGFR)
Assay mechanism:     Inhibition
Assay technology:    Biochemical

---

📤 Output Format

The model returns structured predictions:

Active:  true | false
pXC50:   <numeric value> | null
Reason:  <short biological explanation of the predicted interaction>

Example Output

Active:  true
pXC50:   7.4
Reason:  The compound's sulfonamide group forms key hydrogen bonds with the EGFR
         ATP-binding pocket, suggesting moderate inhibitory activity consistent
         with the predicted pXC50.

---

📊 Training Data

DrDTI-Reasoner was trained on curated drug–target interaction datasets containing:

• Molecular structures in SMILES format
• Protein target identifiers ( UniProt IDs and gene names)
• Bioassay metadata (mechanism, technology, and mode)
• Binary activity labels (active / inactive)
• Potency values ( pXC50 ) where available

Key sources include ChEMBL and BindingDB bioactivity databases.

---

📈 Evaluation Results

Evaluated on a held-out test split from the ChEMBL bioactivity dataset:

⚠️ These are preliminary research metrics. Results may vary across target classes, assay types, and chemical scaffolds not well-represented in the training data. Scaffold-split benchmarking is planned for a future release.

---

⚠️ Limitations

Please read carefully before use:

• Research only — This is not a clinical or regulatory-grade system
• Probabilistic outputs — Predictions are model-generated estimates, not experimentally verified results
• Statistical SMILES understanding — Molecular interpretation is learned from data, not from physical simulation or quantum chemistry
• Format-sensitive — Prediction quality degrades with poorly formatted or non-standard inputs
• Not medically validated — This model has not been assessed for safety or efficacy in any clinical context

---

🚀 Intended Use

DrDTI-Reasoner is designed for:

✅ Computational drug discovery research
✅ Bioactivity prediction experiments
✅ Machine learning benchmarking in cheminformatics
✅ Educational use in bioinformatics and AI
✅ Early-stage hit identification and prioritization workflows

Not intended for:

❌ Clinical diagnosis or treatment decisions
❌ Regulatory or pharmaceutical approval processes
❌ Any application directly impacting human health

---

🗺️ Roadmap

• Improved molecular encoding — SMILES-aware tokenization and embedding strategies
• Protein sequence integration — Embedding protein context via ESM models
• Multi-task learning — Joint classification (active/inactive) and regression (pXC50) heads
• Scaffold-based evaluation — Benchmarking on scaffold-split datasets to assess generalization
• Calibrated confidence scores — Better-calibrated uncertainty estimates for pXC50 predictions

---

📦 Base Model & Framework

---

Disclaimer: DrDTI-Reasoner is a research tool. All predictions should be interpreted in the context of existing literature and validated experimentally before any downstream use.