Skip to content

About

Hackathon on Generative AI and Machine Learning

Resources

Stars

0 stars

Watchers

0 watching

Forks

Β 
Β 

Latest commit

Β 

History

14 Commits

Folders and files

Repository files navigation

🧬 Agentic RNA Biomarker Detection

Problem Statement

Predicting RNA structure and assessing cancer risk from sequence data remains a challenge due to limited structural databases, complex folding, and fragmented tools. Most labs lack integrated, explainable, and privacy-preserving platforms for actionable cancer risk insights.

Our project: enables researchers to input RNA sequences, uses local agentic AI for structure prediction and cancer classification, and outputs clear, interpretable results all in one web interface. This accelerates cancer biomarker discovery, empowers clinicians, and ensures compliance, setting a new standard for accessible, explainable, and secure RNA-driven cancer diagnostics with next-generation AI.


Overview

This project implements an Agentic AI pipeline for detecting biomarker regions in RNA 3D structures (.pdb files). It integrates structural bioinformatics tools and a locally running BioGPT LLM to automate the workflow β€” from structure parsing to functional annotation. The project also demonstrates the integration of LangChain for orchestrating AI reasoning and RAG (Retrieval-Augmented Generation) for combining structured RNA data with external knowledge sources.

After detecting biomarker regions, the system can suggest potential drug candidates using molecular SMILES representations and diffusion-based generative models for virtual compound generation. Although this project currently uses BioGPT and the same process for analysis, this step demonstrates the potential for AI-assisted RNA-targeted drug discovery.


πŸ” System Architecture

Blue and White Modern Organization Structure Instagram Post  (1) ### Agentic AI Orchestrator
  • Coordinates multiple tools:

    • Parses RNA .pdb files
    • Runs fpocket to detect potential binding pockets
    • Uses LangChain to manage AI reasoning and task planning
    • Utilizes RAG techniques to enhance BioGPT predictions with external databases (e.g., RNAcentral, PubMed)
    • Produces biomarker predictions in JSON format
    • Suggests drug candidates via SMILES-based reasoning and generative models

βš™οΈ Tools Used

1️⃣ PDB Parser (Biopython)

  • Extracts atomic coordinates, residue sequences, and secondary structure motifs from .pdb files
  • Outputs structured JSON summarizing the RNA structure

Example Output:

{
  "chains": ["A"],
  "residue_count": 46,
  "motifs_detected": ["stem-loop", "bulge"],
  "sequence": "AUGCUAGU..."
}

2️⃣ fpocket

  • Detects binding pockets/cavities in RNA structures
  • Provides pocket coordinates, volumes, hydrophobicity, and druggability score

Example Output:

{
  "pocket_id": 1,
  "score": 52.4,
  "center": [12.4, 5.7, -3.1],
  "residues": ["A12", "U13", "G14"]
}

3️⃣ BioGPT (Local LLM)

  • Analyzes structure + pocket data to predict:

    • Biomarker regions
    • Structural motifs
    • Function (ligand/metal binding)
    • Disease relevance
  • Enhanced with RAG for incorporating external biomedical knowledge

  • Can also suggest drug candidates via SMILES and diffusion-based reasoning (demonstrative)

Example Output:

{
  "predicted_biomarker_region": "Residues A12–G14",
  "structural_motif": "hairpin loop",
  "function": "metal ion binding",
  "associated_disease": "Lung Adenocarcinoma (LUAD)",
  "potential_drugs": ["CCO", "CCN", "C1=CC=CC=C1"]
}

🧩 Input and Output (Agentic AI Workflow)

Input

An RNA Gene Sequence Expression (e.g., 8RBJ)

Workflow Steps

  1. Generates 3D structure of RNA gene
  2. Biomarker detection to locate pockets present in the 3D structure
  3. BioGPT to find suitable drugs for the detected pockets
  4. Enhanced drugs for targeted and personalized therapy

Output Example

{
  "predicted_biomarker_region": "Residues A12–G14",
  "structural_motif": "hairpin loop",
  "function": "metal ion binding",
  "associated_disease": "Lung Adenocarcinoma (LUAD)",
  "potential_drugs": ["CCO", "CCN", "C1=CC=CC=C1"]
}

Stored in: outputs/results.json

{
  "target": "luad_rna",
  "structure_summary": {
    "chains": ["A"],
    "total_residues": 46,
    "motifs_detected": []
  },
  "total_pockets_found": 3,
  "best_biomarker_site": {
    "Pocket_ID": "1",
    "Score": 0.043,
    "Druggability_Score": 0.0,
    "Number_of_Alpha_Spheres": 98.0,
    "Total_SASA": 349.937,
    "Polar_SASA": 281.102,
    "Apolar_SASA": 68.835,
    "Volume": 1620.036,
    "Mean_local_hydrophobic_density": 3.0,
    "Mean_alpha_sphere_radius": 4.284,
    "Mean_alp._sph._solvent_access": 0.557,
    "Apolar_alpha_sphere_proportion": 0.041,
    "Hydrophobicity_score": 4.083,
    "Volume_score": 3.333,
    "Polarity_score": 8.0,
    "Charge_score": 0.0,
    "Proportion_of_polar_atoms": 67.105,
    "Alpha_sphere_density": 7.494,
    "Cent._of_mass___Alpha_Sphere_max_dist": 19.006,
    "Flexibility": 0.0
  },
  "biomarkers": [
    {
      "pocket_id": "1",
      "predicted_region": "A12-A20",
      "motif_type": "stem-loop",
      "function": "Potential ligand-binding site",
      "associated_disease": "Lung carcinoma (hypothetical)",
      "recommended_drug_targets": [
        "Mg2+ pocket stabilizers",
        "RNA-ligand intercalating agents"
      ]
    },
    {
      "pocket_id": "2",
      "predicted_region": "A25-A30",
      "motif_type": "hairpin loop",
      "function": "Metal-binding site",
      "associated_disease": "Renal carcinoma (hypothetical)",
      "recommended_drug_targets": [
        "Metal ion chelators",
        "RNA folding stabilizers"
      ]
    },
    {
      "pocket_id": "3",
      "predicted_region": "A33-A38",
      "motif_type": "tetraloop",
      "function": "Ligand recognition",
      "associated_disease": "Lung carcinoma (hypothetical)",
      "recommended_drug_targets": [
        "RNA-ligand intercalating agents"
      ]
    }
  ]
}

🧩 Workflow Diagram

PDB Parser β†’ fpocket β†’ LangChain + RAG β†’ BioGPT β†’ Drug Discovery (SMILES/Generative) β†’ JSON Report

🧠 Agent Role Table

Agent Role Function in RNA Hub Implementation Strategy (The "How")
Planning & Supervisor Agent Coordinates the overall workflow and decomposes tasks into biomarker analysis and drug suggestion sub-tasks Implemented via BioGPT as the central reasoning engine with LangChain orchestration capability
Data Acquisition Agent Collects and validates RNA structure and sequence data from PDB and RNAcentral Uses RAG to incorporate external knowledge into BioGPT predictions
Structure & Analysis Agent Predicts secondary/3D structure, detects pockets, and identifies potential biomarker regions Combines PDB Parser, fpocket, and BioGPT reasoning to produce structured biomarker outputs
Drug Discovery Agent Suggests potential small molecules or drug candidates targeting the biomarker regions Uses SMILES representations and diffusion-based generative reasoning guided by BioGPT

πŸ“Š Example Output

{
  "rna_file": "luad_rna.pdb",
  "pockets_detected": 3,
  "best_biomarker_site": {
    "region": "A12–G14",
    "motif": "loop",
    "binding": "metal ion",
    "disease_association": "LUAD",
    "potential_drugs": ["CCO", "CCN", "C1=CC=CC=C1"]
  }
}

πŸš€ Future Enhancements

  • Integrate AlphaFold or RoseTTAFold RNA for de novo structure prediction
  • Query PubChem/UniProt APIs for biological validation
  • Extend to drug docking simulations for target validation
  • Implement full agentic loops with LangChain to allow BioGPT to plan, execute, and refine predictions
  • Visualize pockets, biomarkers, and suggested compounds in 3D (PyMOL/NGLview)

🧬 References

  • Liu et al., 2023. BioGPT: Generative Pre-trained Transformer for Biomedical Text Mining
  • Le Guilloux et al., 2009. fpocket: An open-source platform for ligand pocket detection
  • Berman et al., 2000. The Protein Data Bank: a database of macromolecular structures
  • LangChain documentation: https://www.langchain.com/
  • RAG: Lewis et al., 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Youtube link:

About

Hackathon on Generative AI and Machine Learning

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages