Proprius Labs - AI-First Personal Identity Infrastructure

Most MCP servers are thin wrappers around REST APIs. PocketScout is different: it encodes domain expertise into tool design so Claude can reason like a medicinal chemist.

The Gap Between "I Have a Target" and "I'm Running RFdiffusion"

Before any computational protein design work begins, scientists spend hours to days manually gathering information across half a dozen browser tabs. They check UniProt for protein function, browse PDB for experimental structures, search ChEMBL for prior art, read papers for allosteric insights — then synthesize it all in their heads.

This reconnaissance step is where drug discovery campaigns quietly go wrong. A scientist picks the obvious binding site without checking that 200 compounds have already failed there. They miss an allosteric pocket described in a recent paper. They don't realize the binding site residues aren't conserved in mouse until their in vivo model fails six months later.

Why MCP Is the Right Pattern

Model Context Protocol is how AI assistants connect to external tools and data sources. But the life sciences MCP ecosystem is thin — most servers are hackathon-quality wrappers around REST APIs with no error handling, no scientific context in the tool descriptions, and no thought about how tools compose into real workflows.

The ecosystem has plumbing. What it lacks is domain expertise.

Tool Descriptions as Prompt Engineering

Here's the key insight: Claude selects and sequences tools based on their docstrings and parameter descriptions. Every description in PocketScout encodes scientific reasoning — not just what the tool returns, but when to use it, what the results mean in context, and how they gate downstream decisions.

A get_binding_sites tool that says "get binding sites" is useless. One that says "call this AFTER characterize_target, each site includes druggability assessment and modality recommendations" actually helps Claude reason like a medicinal chemist.

Six Tools, One Assessment

PocketScout provides six domain-informed tools that compose across five public APIs (UniProt, RCSB PDB, AlphaFold DB, ChEMBL, PubMed):

characterize_target: Protein function, family classification, subcellular location, disease associations, plus AlphaFold structure confidence to flag regions where predicted binding sites can't be trusted

get_related_structures: All experimental structures for the target across PDB, sorted by resolution, with ligand inventories. A target with 200 co-crystal structures is a fundamentally different design problem than one with a single cryo-EM map

get_binding_sites: Known pockets from co-crystallized ligands, filtered for crystallization artifacts (45 common buffer/cryo compounds), classified as orthosteric, allosteric, cofactor, or PPI — with modality-specific druggability notes

get_ligand_history: Bioactivity landscape from ChEMBL — how many compounds tested, potency distribution, clinical candidates — classified as crowded, moderate, emerging, or untargeted

check_conservation: Compares binding site residues between human and mouse orthologs. The step most scientists skip and the one that most often causes late-stage preclinical failures

search_target_literature: Focused PubMed searches for structural biology and drug design papers — allosteric mechanisms, resistance mutations, cryptic sites from MD simulations

The tools mirror how expert drug hunters actually evaluate targets: establish biological context, assess structural coverage, map known pockets, evaluate competitive landscape, check translatability, search literature.

Pre-Computed Interpretations

A raw list of ChEMBL bioactivities is hard for Claude to reason about. A structured competitive landscape assessment — "crowded, 200+ compounds with sub-micromolar activity" — lets Claude synthesize meaningfully.

Rather than returning raw resolution values (2.1 Angstroms) and expecting the LLM to know what that means, PocketScout returns "well-resolved (2.1A)." The tool does the domain translation so the model can focus on reasoning.

What I Chose Not to Build

The hardest design decisions were about what to leave out.

Why not pocket prediction? Tools like fpocket and P2Rank computationally predict novel binding sites, but they require CPU/GPU infrastructure that doesn't fit the MCP model of lightweight API-based tools. PocketScout focuses on known binding intelligence from experimental data. Pocket prediction belongs in a separate compute-oriented server — clear boundaries, clear scope.

Simplified conservation. Full multi-species multiple sequence alignment is computationally expensive and error-prone. Human vs. mouse sequence comparison using context-aware residue matching covers the most critical preclinical translatability question and handles indels between orthologs. The limitation is documented; the pragmatic choice is intentional.

Try It Live

PocketScout runs as a live MCP server. You can connect it to Claude right now:

Open claude.ai

Go to Settings > Connectors

Click "Add custom connector"

4. Paste the server URL: https://pocketscout-mcp.up.railway.app/mcp

Then try prompts like:

"Assess the binding landscape of PDB 8QXB for designing a protein binder to block viral entry"

"Is PCSK9 (Q8NBP7) already too crowded for a new binder program?"

"PDB 6OIM has multiple bound ligands. Which pocket is allosteric?"

Where PocketScout Wins

PocketScout shines on prompts where live data, coordinate-level computation, or current database queries matter — not general knowledge.

Ask Claude about a structure deposited after its training cutoff — it won't know it. PocketScout pulls the actual ligands, contacts, and resolution from PDB in real time. Ask about competitive landscape — Claude has general knowledge, but PocketScout pulls current ChEMBL bioactivity counts and potency distributions for a quantitative answer.

The Bigger Picture

The gap in the MCP ecosystem isn't more database connectors — it's scientifically-informed tool composition that reflects real decision-making. PocketScout is a working demonstration of that thesis: six tools, one prompt, deployed and usable today.

PocketScout is open source. View on GitHub or connect it to Claude right now.

Building an MCP Server for Drug Discovery