Source: https://github.com/aipoch/medical-research-skills

When to Use

• Screening large compound libraries to quickly triage for drug-like candidates (e.g., Lipinski/Veber + alerts).
• Flagging problematic chemotypes (e.g., PAINS, reactive/toxicophores, curated structural alerts) before follow-up assays.
• Prioritizing lead-optimization candidates with stricter criteria (lead-like rules, demerit systems, complexity caps).
• Enforcing property constraints (MW/logP/TPSA/rotatable bonds) for target-specific design windows (e.g., CNS).
• Identifying molecules containing specific functional groups/scaffolds (e.g., Michael acceptors, hinge binders) for SAR or risk assessment.

Key Features

• Drug-likeness and medchem rule sets: Lipinski (Ro5), Veber, Oprea, CNS, lead-like (soft/strict), Rule of Three, REOS, Golden Triangle, etc.
• PAINS and structural alert filtering: curated alert catalogs and pattern-based screening.
• Curated industrial filter sets: e.g., NIBR filters; Lilly demerit scoring with pass/fail thresholds.
• Functional-group detection via SMARTS-based group matchers (hinge binders, phosphate binders, Michael acceptors, reactive groups, custom patterns).
• Named catalogs of curated structures (functional groups, protecting groups, reagents, fragments) for matching and annotation.
• Molecular complexity metrics (e.g., Bertz/Whitlock/Barone-style) and threshold-based complexity filters.
• Constraint-based filtering for property windows (MW/logP/TPSA/RB, etc.).
• Query language to combine heterogeneous criteria (rules + alerts + numeric thresholds) into a single expression.

Dependencies

• medchem (latest)
• datamol (latest)
• pandas (latest, for tabular workflows)

Example Usage

# End-to-end, runnable example:
# 1) load SMILES
# 2) apply Ro5 + Veber
# 3) apply common structural alerts
# 4) compute complexity and filter
# 5) export a CSV with decisions

import pandas as pd
import datamol as dm
import medchem as mc

smiles_list = [
    "CC(=O)OC1=CC=CC=C1C(=O)O",  # aspirin
    "CN1C=NC2=C1C(=O)N(C(=O)N2C)C",  # caffeine
    "c1ccccc1",  # benzene
]

df = pd.DataFrame({"smiles": smiles_list})
mols = [dm.to_mol(smi) for smi in df["smiles"]]

# 1) Drug-likeness rules
rule_filter = mc.rules.RuleFilters(rule_list=["rule_of_five", "rule_of_veber"])
rule_res = rule_filter(mols=mols, n_jobs=-1, progress=False)
df["passes_rules"] = rule_res["pass"]

# 2) Structural alerts
alerts = mc.structural.CommonAlertsFilters()
alert_res = alerts(mols=mols, n_jobs=-1, progress=False)
df["has_alerts"] = alert_res["has_alerts"]

# 3) Complexity (example threshold)
complex_filter = mc.complexity.ComplexityFilter(max_complexity=500)
complex_res = complex_filter(mols=mols, n_jobs=-1, progress=False)
df["passes_complexity"] = complex_res["pass"]

# 4) Final decision
df["keep"] = df["passes_rules"] & (~df["has_alerts"]) & df["passes_complexity"]

# 5) Save results
df.to_csv("medchem_screening_results.csv", index=False)
print(df)

Implementation Details

• Rule evaluation (medchem.rules)

  • Rules are implemented as callable checks over SMILES or RDKit-like molecule objects (commonly via datamol).
  • RuleFilters(rule_list=[...]) applies multiple rules and returns a structured result (typically including an overall pass plus per-rule details).
  • Typical use: start broad (Ro5/Veber), then tighten (CNS/lead-like) as project constraints become clearer.

• Structural alerts (medchem.structural)

  • Alert systems are primarily SMARTS/pattern-based matchers curated from literature/industrial practice.
  • CommonAlertsFilters, NIBRFilters, and LillyDemeritsFilters provide different philosophies:
    • Common alerts: general-purpose red flags.
    • NIBR: curated industrial filter set.
    • Lilly demerits: assigns penalties per matched rule; a common convention is reject if total demerits > 100.

• Complexity (medchem.complexity)

  • Complexity scores approximate synthetic difficulty / structural intricacy using established heuristics (e.g., Bertz/Whitlock/Barone-style metrics).
  • ComplexityFilter(max_complexity=...) converts a numeric score into a pass/fail gate for library triage.

• Constraints (medchem.constraints)

  • Property windows (MW/logP/TPSA/rotatable bonds, etc.) are applied as hard filters.
  • Use constraints to encode target-specific design hypotheses (e.g., CNS-like space) rather than universal “good/bad” judgments.

• Groups and catalogs (medchem.groups, medchem.catalogs)

  • Group detection is SMARTS-driven and returns boolean matches and/or match details (substructure hits).
  • Named catalogs provide curated sets for consistent annotation and matching across projects.

• Parallelization

  • Most batch APIs accept n_jobs; set n_jobs=-1 to use all available CPU cores for large libraries.