Protein-AI-Research-Assistant

# Protein AI Research Assistant Protein AI Research Assistant is a domain-specific AI agent for protein sequence exploration, secondary-structure prediction, UniProt annotation lookup, and training-dataset analysis. The project started as a PyTorch LSTM-based protein secondary-structure prediction system. It now combines: - local Q3 and Q8 prediction tools backed by trained PyTorch checkpoints, - a FastAPI inference service for standalone serving, - a Google ADK agent for conversational workflows, - UniProt REST API tools for biological lookup, - a custom MCP server for guarded BigQuery access. This is not a general autonomous scientist. It is a focused protein research assistant that can chain retrieval, prediction, and dataset analysis in one workflow. ## Current Scope The assistant can currently: - predict Q3 secondary structure for one sequence, - predict Q8 secondary structure for one sequence, - run batch Q3 and Q8 predictions, - search UniProt for candidate protein entries, - retrieve a UniProt entry with sequence, function annotations, GO terms, and keywords, - compare a protein against the training dataset through BigQuery, - combine retrieved facts and model predictions in one natural-language response. Example workflow: 1. search UniProt for a reviewed human protein, 2. retrieve the protein sequence and functional annotations, 3. compare sequence length to the training dataset average, 4. run Q3 or Q8 prediction locally, 5. return a concise interpretation with confidence caveats. ## Architecture ```mermaid flowchart TD U["User"] --> A["Google ADK Agent<br/>ProteinResearchAgent"] A --> P["Direct Python Tools"] P --> Q3["predict_q3 / predict_q8"] P --> BQ["batch_predict_q3 / batch_predict_q8"] P --> UNI["search_uniprot / get_uniprot_entry"] Q3 --> M["Local PyTorch Inference"] BQ --> M M --> ART["Model Artifacts<br/>serving/artifacts"] UNI --> UR["UniProt REST API"] A --> MCP["MCP Toolset"] MCP --> S["protein_bq_mcp_server"] S --> GT["get_table_info"] S --> QT["query_protein_data"] S --> BI["BigQuery"] BI --> DS["Protein training dataset"] ``` Supporting components: - `serving/`: FastAPI app and model-serving code - `Protein_agent/`: ADK agent, prediction tools, UniProt tools, prompt, and internal schemas - `protein_bq_mcp_server/`: MCP server exposing read-only BigQuery operations - `protein_model/`: shared training/model code - `scripts/`: training and analysis utilities ## Model Performance Published benchmark results from the sequence-modeling work: | Dataset | Q3 Accuracy | Q8 Accuracy | |---------|-------------|-------------| | CB513 | 76.2% | 62.1% | | TS115 | 75.8% | 61.7% | | CASP12 | 74.9% | 60.3% | These scores reflect the trained sequence models, not end-to-end agent quality. ## Repository Layout ```text protein_struct_proj/ ├── Protein_agent/ │ ├── agent.py │ ├── agent-prompt.md │ ├── tools.py │ ├── uniprot_tools.py │ └── schemas.py ├── protein_bq_mcp_server/ │ └── server.py ├── protein_model/ │ ├── architecture.py │ ├── data_utils.py │ └── preprocess_training.py ├── serving/ │ ├── app/ │ │ ├── main.py │ │ ├── model.py │ │ ├── preprocess.py │ │ └── schemas.py │ └── artifacts/ ├── scripts/ │ ├── train.py │ ├── plots.py │ └── queries/ ├── dataset/ ├── Dockerfile ├── requirements.txt └── README.md ``` ## Installation ### Prerequisites - Python 3.11 recommended - Docker, if you want to run the FastAPI container - Google Cloud credentials, if you want BigQuery MCP access ### Setup ```bash git clone https://github.com/Mubarak-11/Protein_Secondary_Strucuture_Prediction.git protein_struct_proj cd protein_struct_proj pip install -r requirements.txt ``` If you are using a local virtual environment, activate it before installing dependencies. ## Running the FastAPI Service Local development: ```bash uvicorn serving.app.main:app --reload ``` Docker: ```bash docker build -t protein-serving . docker run -p 8000:8000 protein-serving ``` Available endpoints: - `GET /health` - `POST /predict/q3` - `POST /predict/q8` - `POST /predict/batch_q3` - `POST /predict/batch_q8` Example: ```bash curl -X POST http://localhost:8000/predict/q3 \ -H "Content-Type: application/json" \ -d '{"sequence": "MVLSPADKTNVKAAW"}' ``` ## Running the ADK Agent The agent uses: - local PyTorch prediction tools, - UniProt REST lookups, - an MCP server for BigQuery access. Before running the agent, make sure the required environment and credentials are available for: - Google ADK, - BigQuery access, - local model artifacts in `serving/artifacts/`. Typical local flow: ```bash cd protein_struct_proj # activate your virtual environment first # then start the ADK development workflow using Protein_agent/agent.py ``` The exact launch command may vary with your local ADK setup, but the root agent is defined in `Protein_agent/agent.py`. The agent is currently intended for local development and demo workflows rather than public production deployment. ## UniProt and BigQuery Workflows Recent improvements include: - richer UniProt entry retrieval with function annotations, GO terms, and keywords, - richer UniProt search results with reviewed-entry metadata, - grouped prediction regions for Q3 and Q8 outputs, - prompt updates that better separate retrieved facts from model predictions. This means the agent can now answer prompts such as: - “Find the reviewed UniProt entry for human myoglobin and summarize its function.” - “Compare this protein’s length to the average length in the training dataset.” - “Predict Q3 for the UniProt sequence and summarize the major structural regions.” ## Current Progress Implemented and working locally: - trained Q3 and Q8 PyTorch checkpoints, - FastAPI serving path, - ADK agent with local prediction tools, - MCP BigQuery integration, - UniProt search and entry lookup, - structured prediction regions for cleaner agent summaries, - internal Pydantic schemas that reflect tool output contracts. Not yet done: - public deployment of the ADK agent, - tighter canonical-entry ranking for ambiguous UniProt searches, - richer biological fields such as domains or subcellular location, - evaluation harnesses for systematic agent benchmarking, - polished public repo/demo assets. ## Future Iterations Planned next steps include: - improve UniProt candidate selection and ranking behavior, - expand biological annotations and protein metadata retrieval, - add better formatting or smoothing for structural region summaries, - improve evaluation and observability for agent responses, - add semantic protein search and retrieval workflows, - explore mutation analysis and structure-impact experiments, - prepare a cleaner public demo and repository presentation. ## Caveats - Prediction confidence is a model confidence score, not a calibrated biological probability. - The sequence models can disagree with known structural biology, especially when tertiary context matters. - BigQuery access is guarded for development use, but the current MCP guardrails are not a full public security boundary. ## License This project is licensed under the MIT License.