Chem-Brainstorm — Computational Chemistry Thinking Guide

A flexible brainstorming framework for comp chem problems. Not a rigid checklist — activate only the steps relevant to the problem at hand. For complex/recurring workflows, load the appropriate references/protocol-*.md.

Step 1 — Classify the Problem

Identify which mode(s) apply (can be multiple):

For simple one-off questions (e.g. "LogP of aspirin"), skip to Step 3 directly.

Step 2 — Audit Inputs

What is available?

Missing inputs → note what to fetch (Step 3 MCPs).

Step 3 — Tool Map

Match inputs and goal to available tools. Order = cost (⚡ fast → ⚡⚡⚡ expensive).

ALKYL Scripts (local, instant)

ALKYL Skills (conceptual / coding help)

MCPs (external data, use when local data insufficient)

Marketplace Skills (specialized)

Step 4 — Propose Directions

Generate 2–3 concrete directions, each annotated with:

• Cost estimate (⚡/⚡⚡/⚡⚡⚡)
• First concrete action (script call or MCP query)
• What the result will tell us

Format:

Direction 1 — [name] ⚡
→ Action: python chem_filter.py --smiles "..."
→ Tells us: drug-likeness baseline before investing further

Direction 2 — [name] ⚡⚡
→ Action: ChEMBL compound_search + get_bioactivity
→ Tells us: known activity landscape for this scaffold

Direction 3 — [name] ⚡⚡⚡
→ Action: chem_3d.py → diffdock against PDB:XXXX
→ Tells us: predicted binding pose and affinity

Step 5 — Sanity Checks (auto-activate if SMILES available)

Always run before investing in expensive steps:

chem_filter.py  → Lipinski, Veber, PAINS alerts
chem_analyze.py → SA score (> 6 = hard to synthesize), QED
chem_pka.py     → dominant form at physiological pH 7.4
chem_metabolism.py → CYP450 liabilities

Red flags that change the plan:

• SA score > 6 → consider simpler analog or retrosynthesis first
• PAINS alert → flag as assay interference risk
• net_charge ≠ 0 at pH 7.4 → affects permeability/docking

Step 6 — Literature Anchor (activate for HYP mode or novel targets)

bioRxiv.search_preprints(query="[target/method]", date_range="2024-2026")
→ find recent methods, negative results, benchmark conditions

ChEMBL.target_search(query="[gene]")
→ confirm target is druggable, find reference ligands

OpenTargets.search_entities(query="[target]")
→ genetic evidence, disease associations, tractability score

Decision Tree

Got SMILES?
  └─ Yes → run sanity checks (Step 5) first
  └─ No  → fetch from ChEMBL/PubChem or start from concept

Simple property question?
  └─ Yes → Step 3 scripts directly, no need for full workflow

Need external data?
  └─ Yes → MCPs before running expensive local tools

Complex / reproducible workflow?
  └─ Yes → load references/protocol-*.md

Related Skills

• daylight-theory — write correct SMARTS/SMIRKS for queries and transforms
• rdkit — implement any cheminformatics step in Python
• synkit — if the problem involves reaction mechanisms or networks