tooluniverse-pharmacovigilance

COMPUTE, DON'T DESCRIBE

When analysis requires computation (statistics, data processing, scoring, enrichment), write and run Python code via Bash. Don't describe what you would do — execute it and report actual results. Use ToolUniverse tools to retrieve data, then Python (pandas, scipy, statsmodels, matplotlib) to analyze it.

Pharmacovigilance Safety Analyzer

Systematic drug safety analysis using FAERS adverse event data, FDA labeling, PharmGKB pharmacogenomics, and clinical trial safety signals.

KEY PRINCIPLES:

1. Report-first approach - Create report file FIRST, update progressively
2. Signal quantification - Use disproportionality measures (PRR, ROR)
3. Severity stratification - Prioritize serious/fatal events
4. Multi-source triangulation - FAERS, labels, trials, literature
5. Pharmacogenomic context - Include genetic risk factors
6. Actionable output - Risk-benefit summary with recommendations
7. English-first queries - Always use English drug names in tool calls

When to Use

Apply when user asks:

• "What are the safety concerns for [drug]?"
• "What adverse events are associated with [drug]?"
• "Is [drug] safe? What are the risks?"
• "Compare safety profiles of [drug A] vs [drug B]"
• "Pharmacovigilance analysis for [drug]"

Clinical Reasoning Framework

Reasoning Strategy 1: On-Target vs Off-Target Thinking

Ask: is this adverse effect a predictable extension of the drug's mechanism (on-target), or something the mechanism doesn't explain (off-target)? On-target effects are dose-dependent and predictable. Off-target effects are often idiosyncratic and harder to predict.

How to apply this:

1. Look up the drug's primary mechanism of action (use ChEMBL or DailyMed label)
2. For each reported adverse event, ask: "Does this follow logically from what the drug does to its target?" If yes, it is on-target toxicity — expect dose-dependence and manage with dose reduction
3. If the adverse event cannot be explained by the primary mechanism, consider off-target receptor binding or reactive metabolite formation. These require different management (drug discontinuation, not dose adjustment)
4. Use KEGG pathway data to identify metabolic routes that could produce toxic intermediates

Reasoning Strategy 2: Timeline as Diagnostic Tool

When did the adverse event start relative to drug initiation? The timeline alone narrows the mechanism:

• Hours = anaphylaxis, immediate hypersensitivity, or direct pharmacological overshoot
• Days = serum sickness, cytotoxic reactions, cumulative pharmacological effects
• 1-6 weeks = delayed hypersensitivity (SJS/TEN, DRESS), organ accumulation
• Months = chronic toxicity, cumulative organ damage
• Years = long-term cumulative effects

How to apply this: When reviewing FAERS case reports, always check the time_to_onset field. If the reported timeline is biologically implausible for the proposed mechanism, suspect confounding or misattribution. A reaction appearing years after drug start is unlikely to be immune-mediated but could be chronic accumulation.

Reasoning Strategy 3: Dose-Dependent vs Idiosyncratic Classification

This distinction determines monitoring strategy and management:

• Dose-dependent (Type A): Predictable from pharmacology. Dose-response relationship exists. Can be managed by dose reduction. These are on-target toxicities pushed too far.
• Idiosyncratic (Type B): Not predictable from pharmacology alone. No clear dose-response. Often immune-mediated or due to metabolic idiosyncrasy (e.g., genetic variation in drug metabolism). Drug must be stopped — dose reduction will not help.
• Mixed: Some reactions are dose-dependent in most patients but become idiosyncratic in genetically susceptible individuals. When you see a "Type A" reaction occurring at unexpectedly low doses, suspect a pharmacogenomic contributor.

How to apply this: When evaluating a safety signal, classify it as Type A or B. This determines whether you recommend dose adjustment (Type A) or drug avoidance with potential pharmacogenomic screening (Type B).

Reasoning Strategy 4: The Naranjo Algorithm for Causality Classification

When investigating a suspected drug adverse event, the Naranjo algorithm asks: (1) Did the event appear after the drug was given? (2) Did it improve when the drug was stopped? (3) Did it reappear when restarted? (4) Could other causes explain it? Score each question to classify causality.

Reasoning Strategy 5: The Rechallenge Question

Did the event recur when the drug was restarted? Positive rechallenge is the strongest evidence for causation in an individual case. But rechallenge is often unethical for serious reactions, so absence of rechallenge data doesn't exonerate the drug.

How to apply this: When reviewing case narratives or FAERS reports, check for dechallenge (did the event resolve when the drug was stopped?) and rechallenge (did it recur on re-exposure?). A positive dechallenge + positive rechallenge is near-definitive. Negative dechallenge weakens the causal link considerably.

Reasoning Strategy 5: Disproportionality Reasoning

A signal in FAERS means the drug-event pair is REPORTED more than expected. It does not mean the drug CAUSES the event. Think about reporting biases:

• Serious events get reported more than mild ones
• New drugs get reported more than old ones (Weber effect)
• Drugs prescribed to sick populations get events attributed to them that may reflect the underlying disease
• Media attention or regulatory alerts create reporting spikes

How to apply this: Always ask — what is the base rate of this event in the untreated population? A high PRR for "cardiac arrest" in a drug used by ICU patients may reflect the patient population, not the drug. Cross-reference with clinical trial placebo-arm rates when available.

Reasoning Strategy 6: When to Use Tools vs Reason

Use FAERS/OpenFDA tools to QUANTIFY a signal you have already hypothesized based on mechanism. Do not mine FAERS without a hypothesis — you will find spurious associations.

The correct sequence:

1. Reason about mechanism first (what adverse events are plausible given this drug's pharmacology?)
2. Form specific hypotheses (e.g., "this drug may cause QT prolongation because it blocks hERG channels")
3. Query tools to test each hypothesis (FAERS for reporting frequency, DailyMed for label warnings, PharmGKB for genetic risk factors)
4. Interpret results in context (is the signal consistent with the mechanism? Is the timeline plausible? Are there confounders?)

Reasoning Strategy 7: Pharmacogenomic Risk Assessment

Rather than memorizing gene-drug pairs, apply this reasoning framework:

1. Identify the drug's metabolic pathway (use KEGG or DailyMed label): Which CYP enzymes metabolize it? Is it a prodrug requiring activation?
2. Assess the consequence of altered metabolism: For active drugs, poor metabolizers accumulate the drug (toxicity risk). For prodrugs, poor metabolizers fail to activate (efficacy failure). Ultra-rapid metabolizers show the opposite pattern.
3. Check for immune-mediated risk: If the drug is associated with severe cutaneous reactions (SJS/TEN, DRESS) or hypersensitivity syndrome, query PharmGKB for HLA associations. These are population-specific.
4. Use PharmGKB evidence levels to guide action: Level 1A/1B (guideline-based) = actionable now. Level 2A/2B = may inform. Level 3 = not clinically actionable yet.

Query PharmGKB_search_drugs(query=...) and CPIC_list_guidelines to get current pharmacogenomic annotations rather than relying on memorized associations, which may be outdated.

Critical Workflow Requirements

Report-First Approach (MANDATORY)

1. Create [DRUG]_safety_report.md FIRST with all section headers and [Researching...] placeholders
2. Apply mechanistic reasoning first (on-target toxicity, time-to-onset, dose vs. idiosyncratic, PGx)
3. Progressively update as data is gathered
4. Output separate data files: [DRUG]_adverse_events.csv and [DRUG]_pharmacogenomics.csv

Citation Requirements (MANDATORY)

Every safety signal MUST include source tool, data period, PRR, case counts, and serious/fatal breakdown.

Tool Parameter Reference (CRITICAL)

Workflow Overview

Phase 0: Mechanistic Reasoning (BEFORE tools)
  On-target toxicity, time-to-onset, dose vs idiosyncratic, PGx risk

Phase 1: Drug Disambiguation
  -> Resolve drug name, get identifiers (ChEMBL, DrugBank)

Phase 2: Adverse Event Profiling (FAERS)
  -> Query FAERS, calculate PRR, stratify by seriousness

Phase 3: Label Warning Extraction
  -> DailyMed boxed warnings, contraindications, precautions

Phase 0: Mechanistic Reasoning (DO THIS BEFORE TOOLS)

1. Identify drug class and primary mechanism (use DailyMed label or ChEMBL)
2. Apply on-target vs off-target thinking (Strategy 1) to predict plausible adverse events
3. Estimate expected time-to-onset for each predicted event (Strategy 2)
4. Classify each as dose-dependent vs idiosyncratic (Strategy 3)
5. Formulate specific, testable safety hypotheses to guide tool queries (Strategy 6)

Phase 1: Drug Disambiguation

1. Search DailyMed via DailyMed_search_spls(drug_name=...) for NDC, SPL setid, generic name
2. Search ChEMBL via ChEMBL_search_drugs(query=...) for molecule ID, max phase
3. Document: generic name, brand names, drug class, mechanism, approval date

Phase 2: Adverse Event Profiling (FAERS)

1. Query FAERS_count_reactions_by_drug_event(drug_name=..., limit=50) for top events
2. For each event, get detailed breakdown (serious, fatal, hospitalization counts)
3. Calculate PRR: (A/B) / (C/D) where A=drug+event, B=drug+any, C=event+any_other, D=total_other
4. Apply signal thresholds: PRR > 2.0 (signal), > 3.0 (strong signal), case count >= 3

Severity classification:

• Fatal (highest priority), Life-threatening, Hospitalization, Disability, Other serious, Non-serious

FAERS filter_serious_events -- MedDRA Spelling (CRITICAL)

FAERS_filter_serious_events uses MedDRA preferred terms which follow British English spelling conventions. Common examples:

The adverse_event parameter should use the exact MedDRA preferred term spelling. When in doubt, first query FAERS_count_reactions_by_drug_event to see the exact event names as they appear in the FAERS database, then use those exact strings.

Additional FAERS notes:

• adverse_event is now correctly appended to the OpenFDA query in _filter_serious_events
• FAERS_stratify_by_demographics: adverse_event is optional — when omitted, stratification covers all events for the drug. Sex codes: 0=Unknown, 1=Male, 2=Female

See SIGNAL_DETECTION.md for detailed disproportionality formulas and example output tables.

Phase 3: Label Warning Extraction

1. Get label via DailyMed_get_spl_by_setid(setid=...)
2. Extract: boxed warnings, contraindications, warnings/precautions, drug interactions
3. Categorize severity: Boxed Warning > Contraindication > Warning > Precaution

Phase 3.5: SIDER Structured Label Side Effects (Faster Alternative/Complement to Raw Label Parsing)

SIDER (SIDER_search_drug, SIDER_get_drug_side_effects, SIDER_get_drug_indications, SIDER_get_drugs_for_side_effect, SIDER_search_side_effect) pre-extracts and structures side effects and indications directly from drug package inserts/labels — a curated, static snapshot, not a real-time feed like FAERS. Use it when you want the label's own reported side-effect frequencies (e.g. "diarrhoea, very common, 9.6%-57.5%") without hand-parsing DailyMed's raw label text.

1. SIDER_search_drug(drug_name=...) -> resolve to a sider_drug_id (PubChem CID)
2. SIDER_get_drug_side_effects(drug_name=... or sider_drug_id=..., limit=...) -> MedDRA-coded side effects with the drug's own reported frequency and, when the label includes a placebo arm, the placebo-arm frequency for direct comparison
3. SIDER_get_drugs_for_side_effect(side_effect_name=... or meddra_code=...) / SIDER_search_side_effect(side_effect_name=...) -> reverse lookup: which other drugs report this side effect, for class-level comparison

Frequency field semantics (read carefully before reporting a number):

• frequency is copied verbatim from the label's "Data for drug" column — it may be a percent, a percent range, a qualitative term (postmarketing, rare, frequent), or a combination (uncommon, 5% - 7%). It is null when the label reports no frequency at all — do not report null as "0%" or "not reported as a risk."
• placebo_frequency is null for the majority of side effects because most SIDER labels have no placebo arm at all. Null means "SIDER has no placebo rate for this entry," not "the placebo rate is zero" and not "the placebo rate equals the drug's rate." Never state or imply a placebo comparison you don't actually have a placebo_frequency value for.

Known SIDER data-quality gotcha (verified live, not a bug to route around silently): SIDER_get_drug_indications is not a clean "what this drug treats" list — it is extracted by text-mining the label's indications section and label text more broadly, so it can include conditions that are actually risks, contraindications, or complications rather than approved uses. Live example: metformin's SIDER "indications" list includes Diabetes mellitus and Type 2 diabetes mellitus (correct) alongside Coma, Lactic acidosis, Diabetic ketoacidosis, and Myocardial infarction (these are risks/warnings associated with metformin, not things it treats). Always sanity-check a SIDER "indication" against known pharmacology before stating it as an approved use — cross-reference against DailyMed's actual "Indications and Usage" section (Phase 3 above) or OpenTargets_get_drug_indications_by_chemblId rather than reporting a SIDER indication list at face value.

SIDER_get_drugs_for_side_effect results are not pre-sorted by frequency or clinical relevance — treat the returned order as arbitrary and re-sort yourself if ranking by reported frequency matters for the analysis.

Phase 4: Pharmacogenomic Risk

1. Search PharmGKB_search_drugs(query=...) for clinical annotations
2. Document actionable variants with evidence levels (1A/1B/2A/2B/3)
3. Note CPIC/DPWG guideline status

PGx Evidence Levels:

Phase 5: Clinical Trial Safety

1. Search search_clinical_trials(intervention=..., phase="Phase 3", status="Completed")
2. Extract serious AE rates, discontinuation rates, deaths
3. Compare drug vs placebo rates

Phase 5.5: Pathway & Mechanism Context

1. Query KEGG for drug metabolism pathways
2. Analyze target pathways for mechanistic basis of AEs
3. Document pathway-AE relationships

Phase 5.6: Literature Intelligence

1. PubMed: PubMed_search_articles(query='"[drug]" AND (safety OR adverse OR toxicity)')
2. BioRxiv/MedRxiv: Search for recent preprints (flag as not peer-reviewed)
3. OpenAlex: Citation analysis for key safety papers

Phase 6: Signal Prioritization

Signal Score = PRR x Severity_Weight x log10(Case_Count + 1)

Severity weights: Fatal=10, Life-threatening=8, Hospitalization=5, Disability=5, Other serious=3, Non-serious=1

Categorize signals:

• Critical (immediate attention): High PRR + fatal outcomes
• Moderate (monitor): Moderate PRR + serious outcomes
• Known/Expected (manage clinically): Low PRR, in label

Cross-check against mechanistic prediction: A signal not predicted mechanistically warrants additional scrutiny (possible confounding, reporting bias, or genuinely novel finding).

Output Report

Save as [DRUG]_safety_report.md. See REPORT_TEMPLATES.md for the full report structure and example outputs.

Evidence Grading

Fallback Chains

Completeness Checklist

See CHECKLIST.md for the full phase-by-phase verification checklist.

References

• FAERS: https://www.fda.gov/drugs/questions-and-answers-fdas-adverse-event-reporting-system-faers
• DailyMed: https://dailymed.nlm.nih.gov
• PharmGKB: https://www.pharmgkb.org
• ClinicalTrials.gov: https://clinicaltrials.gov
• OpenFDA: https://open.fda.gov
• KEGG Drug: https://www.genome.jp/kegg/drug
• Tool documentation: TOOLS_REFERENCE.md