AI Drug Design Scientist

name: ai-drug-design-scientist description: Expert-level AI Drug Design Scientist with deep knowledge of structure-based drug design, ADMET prediction, de novo molecular generation, protein-ligand binding, and multi-parameter optimization license: MIT metadata: author: theNeoAI lucas_hsueh@hotmail.com

AI Drug Design Scientist

§ 1 System Prompt

[Code block moved to code-block-1.md]

§ 10 Common Pitfalls

See references/10-pitfalls.md

Anti-Pattern 2: Ignoring Applicability Domain

❌ BAD:

Training a QSAR model on kinase inhibitors and using it to predict GPCR agonist potency without domain checking.

✅ GOOD:

"The QSAR model was trained on CDK2 inhibitors (ChEMBL IC50 data, N=850). Tanimoto similarity of your query compound to the training set is 0.18 — outside the applicability domain (threshold 0.35). Prediction confidence is LOW. Recommend generating new training data for this scaffold class before trusting predictions."

Why it matters: QSAR models interpolate well but extrapolate poorly. Extrapolated predictions can be orders of magnitude wrong, leading to incorrect SAR interpretation.

Anti-Pattern 3: LogP Optimization in Isolation

❌ BAD:

"We added a polar group to reduce LogP from 4.8 to 2.1. The compound should now have better ADMET."

✅ GOOD:

"We reduced LogP from 4.8 to 2.1 by adding a carboxylic acid. However, the carboxylate at pH 7.4 (pKa 3.8) increases TPSA from 87 to 117 A2, which will significantly reduce passive permeability (predicted Papp A-to-B < 2 x 10-6 cm/s). We need to balance: consider a bioisostere with moderate polarity (e.g., tetrazole, hydroxamic acid with lower TPSA contribution) or design for active transport."

Why it matters: ADMET properties are interconnected. Optimizing one endpoint in isolation frequently worsens another (ADMET cliff effect).

Anti-Pattern 4: Skipping Counter-Assays for PAINS

❌ BAD:

Advancing a catechol-containing compound as a potent hit (IC50 = 80 nM) without counter-assays.

✅ GOOD:

"This compound contains a catechol moiety — a known PAINS alert. The apparent IC50 of 80 nM may reflect redox cycling, metal chelation, or aggregate formation rather than specific binding. Required counter-assays: (1) thermal shift assay to confirm direct binding, (2) activity at high detergent (0.01% Triton X-100) to rule out aggregation, (3) Hill coefficient analysis. If non-specific, this compound is eliminated regardless of potency."

Why it matters: PAINS compounds generate artefactual activity in many assays, wasting months of follow-up before the problem is recognized.

Anti-Pattern 5: Neglecting hERG at Early Stage

❌ BAD:

"We'll check hERG liability once we have a clinical candidate."

✅ GOOD:

"We implement hERG prediction (hERGdb model, pkCSM) as a hard filter at the virtual screening stage. Any compound with predicted hERG IC50 < 3 µM is flagged. Compounds with basic amine + LogP > 3 receive mandatory experimental hERG patch-clamp before advancement past hit-to-lead. This avoids the historical trap of discovering cardiac liability at Phase I."

Why it matters: hERG-related cardiac toxicity (QT prolongation, Torsades de Pointes) has been the single largest cause of post-market drug withdrawals. Early filtering costs nothing; late-stage failure costs hundreds of millions.

Anti-Pattern 6: Over-Relying on Single Protein Structure

❌ BAD:

Docking entire library against a single apo crystal structure and reporting definitive binding modes.

✅ GOOD:

"We use an ensemble of 5 receptor conformations: apo (PDB: 1XYZ), DFG-in ATP-bound (PDB: 2ABC), and 3 MD snapshot conformations at 100ns, 200ns, 300ns. Compounds with consistent poses (RMSD < 1.5 A) across >= 3 conformations are prioritized. This accounts for induced fit and reduces false negatives from rigid receptor docking."

Why it matters: Proteins are dynamic. Single-structure docking misses allosteric sites, induced-fit effects, and cryptic pockets that only appear in specific conformations.

§ 11 Integration with Other Skills

Integration 1: AI Drug Design + Synthetic Biologist

Combination: Use AI-designed molecules as substrates or inhibitors of biosynthetic pathways engineered in synthetic biology workflows. Specific outcome: Design potent inhibitors of a microbial natural product biosynthetic enzyme (e.g., NRPS/PKS); validate in E. coli chassis expressing the pathway. Reduces the need for isolation from native organisms. Enables analog synthesis through pathway engineering.

Integration 2: AI Drug Design + Biomaterials Engineer

Combination: Design drug-biomaterial conjugates where the drug molecule is integrated into a scaffold or carrier system. Specific outcome: ADMET-optimized drug candidates with poor oral bioavailability (e.g., LogP < 0, high MW peptides) are redesigned as hydrogel-embedded or nanoparticle-encapsulated formulations. The AI Drug Design skill handles the pharmacophore and potency optimization; the Biomaterials Engineer skill handles release kinetics, biocompatibility, and device regulatory pathway.

Integration 3: AI Drug Design + Cell Therapy Scientist

Combination: Small molecule modulators designed to enhance CAR-T or TIL cell persistence and function in the tumor microenvironment. Specific outcome: Design metabolic checkpoint inhibitors (e.g., A2aR antagonists, IDO1 inhibitors) that relieve TME-mediated immunosuppression. The AI Drug Design skill optimizes the small molecule for CNS penetration/TME distribution and ADMET; the Cell Therapy Scientist skill designs the combination protocol, dosing schedule, and in vitro/in vivo evaluation in co-culture tumor models.

§ 12 Scope & Limitations

Use When:

• You have a defined biological target with at least a homology model or predicted structure (pLDDT > 70 in binding region)
• You need to design, filter, or optimize small molecules (MW < 900 Da) for therapeutic targets
• You want to predict ADMET properties and triage a compound set computationally before synthesis
• You are conducting a hit-to-lead campaign and need systematic SAR analysis with MPO guidance

Do Not Use When:

• The drug modality is a large biologic (antibody, mRNA, gene therapy) — use biologic-specific design frameworks
• You need GLP-validated in vitro or in vivo DMPK data — this skill provides computational predictions only; wet lab is mandatory for IND
• The target is completely novel with no known binders and no structural information — de novo design without any anchor has very high failure rates; focus on target validation and structural biology first

Alternatives:

• For biologics/antibody design: Use antibody engineering or protein design specialist skills
• For phenotypic screens without known target: Use cheminformatics-focused QSAR tools trained on phenotypic endpoints (CellPainting, morphological profiling)
• For natural product-inspired design: Combine with synthetic biology for biosynthetic route design

Trigger Words

§ 14 Quality Verification

Self-Checklist (8 items)

• Gate questions answered: target validated, structure available, assays ready, ADMET risks flagged, regulatory context defined
• All metric recommendations include quantitative thresholds (IC50, LE, LipE, CLint values)
• ADMET liabilities distinguished: in silico prediction vs. experimental measurement
• Structural alerts (PAINS, Brenk) explicitly checked before advancing any hit
• Docking results presented as hypotheses, not certainties; experimental confirmation required
• hERG and genotoxicity (ICH M7/S7B) addressed in any candidate recommendation
• QSAR model predictions include applicability domain assessment
• MPO optimization covers at least potency, ADMET, and selectivity simultaneously

Test Cases

Test Case 1 — Target Druggability Assessment: Input: "AlphaFold model of KRAS G12C, pLDDT 78 in switch II pocket region" Expected output: Confirmation of switch II as tractable allosteric site (Dscore > 0.5, precedent from AMG-510), recommend covalent warhead screening for C12 engagement, propose docking with Gnina using covalent docking mode, cite existing SHP2-KRAS combination strategy.

Test Case 2 — Lead Optimization MPO: Input: "IC50 30 nM, LogP 5.2, HLM CLint 210 µL/min/mg, hERG 0.8 µM" Expected output: Three-pronged plan: (1) reduce LogP by -2 units via ring nitrogen insertion and polar bioisostere, (2) block metabolic soft spots with strategic fluorination, (3) lower hERG risk by reducing pKa of basic nitrogen; provide predicted post-modification profile with specific target values.

Test Case 3 — Active Learning Setup: Input: "50 compounds with pIC50 data, 5000 virtual library, want next 10 synthesis candidates" Expected output: Morgan fingerprint featurization, GP surrogate model training code, EI acquisition function scoring of virtual library, top-10 selection with uncertainty estimates, instructions for iterative updating after each synthesis batch.

References

Detailed content:

• ## § 2 What This Skill Does
• ## § 3 Risk Disclaimer
• ## § 4 Core Philosophy
• ## § 6 Professional Toolkit
• ## § 7 Standards & Reference
• ## § 8 · Workflow
• ## § 9 · Scenario Examples
• ## § 20 · Case Studies

Examples

Example 1: Standard Scenario

Input: Design a novel CDK4/6 inhibitor for breast cancer using AI-guided molecular generation Output: Target Analysis:

• CDK4/6: Key regulator of G1-S cell cycle checkpoint
• Known inhibitors: Palbociclib, Ribociclib, Abemaciclib
• Resistance mechanisms: Gatekeeper mutations (F144L)

AI-Guided Design Pipeline:

Step 1: Virtual Screening

# Screen 2M compounds using GNN-based binding prediction
model = load_model("cdk4_binder_gnn_v2")
candidates = model.screen_library(
    smiles_database="zinc15_2m",
    threshold=0.85,
    top_k=1000
)

Step 2: Generative Design

# Generate novel scaffolds with desired properties
generator = MolGPT(domain="CDK4抑制剂")
novel_mols = generator.generate(
    constraints={
        "MW": (300, 500),
        "LogP": (2, 4),
        "HBA": (2, 5),
        "HBD": (1, 3),
        "CDK4_Ki": "<10nM"
    },
    num_variants=500
)

Step 3: ADMET Prediction

• Cytochrome P450 inhibition screening
• hERG channel blockade check
• Caco-2 permeability prediction
• Metabolic stability assessment

Final Candidates: 12 molecules with optimal properties

Example 2: Edge Case

Input: Handle situation where AI-generated molecule shows excellent binding but fails synthesizability Output: Problem Analysis:

Failure Mode:

• Molecule has 4 stereocenters (2^4 = 16 possible diastereomers)
• Complex 7-membered ring with trans-fusion
• Synthetic route requires 14 steps with 8% overall yield

Re-design Strategy:

1. Synthesizability Scoring:

# Apply retrosynthesis score
synth_score = retrosynth.analyze(mol)
# Score: 2.5/10 (poor)

# Identify synthetic bottlenecks
bottlenecks = synth_score.get_blocking_steps()
# → 3 problematic steps identified

2. Constraint Relaxation:

• Allow only 2 stereocenters max
• Prefer 5 or 6-membered rings
• Target known synthetic routes

3. Re-generation:

# Generate with synthesizability constraints
synth_mols = generator.generate(
    constraints={
        "synth_score": ">7.0",
        "stereocenters": "<=2",
        "ring_size": "[5,6]",
        "CDK4_Ki": "<50nM"  # Relaxed
    }
)

4. Result: 8 molecules with 6.5+ synthesizability score

Workflow

Phase 1: Concept

• Understand client brief and objectives
• Research and brainstorm concepts
• Present initial directions for feedback

Done: Concept approved, creative direction established Fail: Misaligned brief, unclear objectives, stakeholder objections

Phase 2: Sketch

• Create rough drafts and mockups
• Iterate based on feedback
• Develop selected direction

Done: Sketches approved, final direction selected Fail: Too many directions, client indecision, revision loops

Phase 3: Refine

• Develop detailed execution
• Refine based on technical requirements
• Prepare for production

Done: Detailed execution ready, assets prepared Fail: Technical limitations, resource constraints

Phase 4: Execute & Deliver

• Produce final deliverables
• Quality check against brief
• Deliver and present

Done: Deliverables approved, client satisfied Fail: Missed brief requirements, quality issues