Agentic_AI

Applications of Artificial Intelligence in Bioinformatics

Agentic Proteomics

Agentic Proteomics is a Streamlit application for accession-guided structural proteomics analysis.
The app combines:

• a dropdown-first UI for selecting a protein accession from a dataset
• retrieval-augmented generation (RAG) using accession-specific rows from VNMX_LiP_DA.csv
• a LangGraph ReAct workflow for deciding when to call structure tools
• live structure-source queries against:
  • PDB
  • AlphaFold DB
  • AlphaFill
  • SWISS-MODEL / 3D-Beacons
• deterministic ranking logic to choose the best candidate structure

Project goal

Given a protein accession from the dataset, the app:

1. retrieves accession-specific experimental context from the CSV
2. asks an LLM to reason over that context
3. allows the LLM to call structure tools when needed
4. collects returned structure candidates
5. ranks the candidates using explicit code-based scoring
6. returns the best available structure source with an explanation

Current workflow

The app uses a LangGraph state machine with the following major steps:

• load_accession_context
• react_model
• tool_node
• collect_candidates
• validate_and_rank
• finalize_answer

The ReAct loop continues until the model no longer requests tools.

Architecture

UI

• Streamlit app
• accession selected from a dropdown built from PG.ProteinAccessions

RAG

• accession-specific rows retrieved from VNMX_LiP_DA.csv
• selected columns are converted to compact markdown and passed to the model

ReAct agent

• local Ollama model via ChatOllama
• tools bound to the model:
  • query_pdb
  • query_alphafold
  • query_alphafill
  • query_swiss_model

Ranking

Candidate structures are scored deterministically using:

• coverage
• resolution
• confidence
• ligand/cofactor context
• source-specific priority

Graph state

The graph state tracks:

• accession
• rag_context
• messages
• structure_candidates
• best_candidate
• final_answer
• errors

Example graph flow

START
  ↓
load_accession_context
  ↓
react_plan
  ├── if tools requested → execute_tools
  └── if done / no tools → validate_and_rank
               ↓
       validate_and_rank
               ↓
         finalize_answer
               ↓
              END