Connecting AI Agents to Scientific Tools with MCP: My OSI Hackathon 2026 Experience

Description and use

During OSI Hackathon 2026, the team built an MCP server enabling AI agents to access scientific tools, and used it to impute missing 26S proteasome data with 95.1 % partner‑based coverage.

Problem

Single‑cell proteomics experiments often miss 20–60% of protein measurements, requiring imputation. Traditional k‑NN methods rely only on existing data patterns, so the team explored whether biological knowledge, especially protein‑complex relationships, could improve imputation accuracy.

Outcome

The MCP integration allowed AI agents to access and invoke specialized scientific tools, enabling a real‑world imputation experiment that achieved 95.1% partner‑based coverage for missing 26S proteasome subunits.

Workflow

• Clone KnowledgeIsPower repo and activate Python virtual environment
• Run offline tests (19 passed, 27 xfailed) to verify MCP server tools
• Configure Claude Code CLI as an MCP client using STDIO transport
• Configure GitHub Copilot CLI as an MCP client using STDIO transport
• Use Copilot CLI to query get_complexes for PSMB5 and retrieve 26S proteasome subunits
• Prepare a 26S‑only .h5ad subset from Petrosius 2023 dataset
• Invoke impute_with_prior with Complex Portal prior to impute missing values
• Collect and interpret imputation metrics (total missing, partner‑based, kNN fallback, coverage)

Evidence

Our team explored whether biological knowledge, particularly relationships between proteins, could improve these estimates.

Source

We developed scp-graphfs-mcp, a scientific Model Context Protocol (MCP) server that enables AI agents to interact with protein-complex data, knowledge graphs, and imputation tools.

Source

The resulting metrics were: Total missing values 638, Imputed from protein partners 607, kNN fallback imputations 31, Partner-based coverage 95.1%

Source

I successfully configured the connection and verified that Claude Code recognized the MCP server.

Source