Molclaw Mol Opt Target

Molecule Optimization for Protein Target Binding

step 1 — Baseline Evaluation

Evaluate the source molecule against the target protein. Run all tools below and collect results.

1.1 Target protein retrieval

Use skill molclaw-protein-sequence-retrieve and molclaw-protein-structure-retrieve to obtain target protein sequence and structure, respectively.

1.2 Binding Activity Evaluation

Utilize molclaw-boltz2-affinity and molclaw-quickvina-docking to evaluate binding between the source molecule (ligand) and target protein (receptor).

Record the baseline metrics:

1.3 Interaction Analysis

Convert the Boltz2 complex structure and analyze protein-ligand interactions:

response = await client.session.call_tool("convert_complex_cif_to_pdb", arguments={"cif_file_path": complex_cif_file})
complex_pdb = client.parse_result(response)["output_file"]

response = await client.session.call_tool("analyze_protein_ligand_interactions", arguments={"complex_file": complex_pdb, "ligand_identifier": "auto"})
interaction_data = client.parse_result(response)
interactions = interaction_data["interactions"]       # H-bonds, hydrophobic, pi-stacking, etc.
pocket_summary = interaction_data["summary"]           # contacted residues, pocket composition
clashes = interaction_data["clashes"]                  # steric clashes to resolve

From the interaction data, identify:

• Key interactions to preserve: Strong H-bonds and salt bridges that anchor the ligand — these must NOT be disrupted by modifications
• Weak or missing interactions: Residues within contact range but forming only weak interactions — opportunities to strengthen
• Steric clashes: Ligand atoms too close to protein residues — must be resolved
• Pocket composition: Hydrophobic vs. polar character of the surrounding residues — guides what type of modifications to attempt

1.4 Source Molecule Property Profiling

Use skill molclaw-admet to calculate precise molecular properties for the source molecule (do NOT rely on LLM estimation).

1.5 Structural Analysis Report

Based on the above data, generate a report covering:

• Baseline scores from 1.2
• Interaction profile from 1.3: list key H-bonds (residue, distance, strength), hydrophobic contacts, pi-stacking, salt bridges, and any clashes
• Property profile from 1.4
• Modification site candidates (2-4 positions): for each, state:
  • Which interaction data motivates this choice (e.g., "no H-bond to nearby ASP189", "clash with VAL138", "uncontacted polar residue SER195")
  • Which strategy from step 2 could apply
• Existing concerns: Lipinski violations, ADMET red flags, structural alerts

step 2 — Generate Optimized Molecules

Based on the Structural Analysis Report from step 1, design 1-3 optimized molecules. Every modification MUST cite a specific residue and interaction from step 1.

Generation approach — LLM design + REINVENT supplement:

• Primary: LLM-guided design (1-3 molecules per round) based on interaction analysis
• Supplement: If LLM designs produce invalid SMILES repeatedly, or if more diversity is needed, use reinvent_mol2mol_sampling with prior_type="scaffold_generic" and n=10-15 to generate additional candidates. Filter these by scaffold preservation and similarity, then dock all valid molecules.
• For ADMET-focused optimization: If step 1 reveals ADMET deficiencies alongside binding targets, use reinvent_similarity_optimization to generate candidates optimized for specific property ranges (QED, LogP, TPSA), then evaluate binding via docking.

Rules:

• Each molecule: only 1-2 structural changes from source
• For fused/polycyclic rings: ONE change only, insert substituents via () within the SMILES ring path, never append atoms after ring closure digits
• Do not disrupt strong interactions identified in step 1
• Do not introduce PAINS substructures or known toxicophores
• Preserve stereochemistry markers (@, @@, /, \) unless directly modifying that center
• If retrying: do NOT repeat any previously failed strategy. Use a different modification site or approach.

Output per molecule:

{
  "InteractionTarget": "Which residue/interaction this targets, citing step 1 data",
  "Modification": "Exact structural change",
  "SMILES": "optimized molecule SMILES",
  "Confidence": "High/Medium/Low"
}

Example: Source c1ccc(-c2ccc(NC(=O)C)cc2)cc1, Vina -6.8. Step 1 shows ASP189 at 4.1Å with no H-bond; terminal phenyl has only weak hydrophobic contacts with LEU134.

{
  "InteractionTarget": "ASP189 at 4.1Å, currently no H-bond. Adding pyridine N as H-bond acceptor.",
  "Modification": "Replace terminal benzene with pyridine: c1ccccc1 → c1ccncc1",
  "SMILES": "c1ccnc(-c2ccc(NC(=O)C)cc2)c1",
  "Confidence": "High"
}

step 3 — Validate & Filter

# 1. Validity check
response = await client.session.call_tool("is_valid_smiles", arguments={"smiles_list": candidate_smiles_list})
valid = [r["smiles"] for r in client.parse_result(response)["valid_res"] if r["is_valid"]]
# If an LLM-designed molecule is invalid: note the error, regenerate in step 2 with a simpler edit. Do NOT retry the same SMILES pattern.

# 2. Drug-likeness filter
response = await client.session.call_tool("calculate_mol_drug_chemistry", arguments={"smiles_list": valid})
passed = [m["smiles"] for m in client.parse_result(response)["metrics"] if m["lipinski_rule_of_5_violations"] <= 1]

# 3. Similarity filter
response = await client.session.call_tool("calculate_smiles_similarity", arguments={"target_smiles": source_smiles, "candidate_smiles_list": passed})
final = [r["smiles"] for r in client.parse_result(response)["similarities"] if 0.4 <= r["score"] <= 0.95]

step 4 — Evaluate & Iterate

Evaluate all filtered candidates with skill molclaw-boltz2-affinity and molclaw-quickvina-docking. Compare against source:

Success: Vina improves ≥ 0.5 kcal/mol OR Boltz2 affinity_pred_value meaningfully improves → report best molecule. Done.

Insufficient improvement: Record what was tried and the scores. Return to step 2 with a different modification site and strategy. Remember all previous attempts — do not repeat failed approaches.

Retry limit: Maximum 3 retries (4 total rounds). If no molecule meets success criteria after all rounds, report the best-performing molecule across all attempts with a comparison table and ADMET profile (pred_mol_admet).

⚠ Computation-First Principle (L3 Principle 13)

ALL values in baseline evaluation and optimization rationale MUST come from tool calls performed in this session. Do NOT fill in values from LLM training knowledge. CORRECT: "ADMET-AI predicts CYP3A4 prob=0.72; the methoxyethoxy group is a likely O-demethylation site (agent analysis based on tool output)." WRONG: "CYP3A4 inhibition is likely due to the methoxyethoxy group" (without first confirming from ADMET-AI).

Each molecular modification MUST cite specific computed data (docking scores, interaction residues, ADMET values) as the basis. Literature-derived rationale must be explicitly labeled.

⚠ Mandatory File Downloads (L3 Principles 14-15)

At each evaluation round, download: Boltz-2 complex CIF files, docking pose PDBQT files, interaction analysis images, and ADMET result JSON files. These are Category A outputs.

⚠ Data Integrity in Optimization Trajectory (L3 Principle 11)

Every number in the comparison table (Vina score, Boltz2 value, QED, ADMET) must be the exact value returned by the tool call — never rounded, estimated, or recalled from memory. Record the tool return value directly.

⚠ Residue Numbering (L3 Principle 17)

When interaction analysis reports residue IDs from Boltz-2 predicted structures, apply residue numbering mapping before interpreting which pocket residues are involved. See molclaw-residue-mapper.