translational-safety

name	translational-safety
description	Safety pharmacology assessment, CTCAE grading, organ-specific toxicity biomarkers, CYP liability analysis, structural alerts, and drug-drug interaction evaluation
version	1.0.0
tags	["safety","toxicology","CTCAE","hERG","hepatotoxicity","CYP","DDI","adverse-events","biomarkers"]

Translational Safety

This skill guides safety assessment for translational medicine, covering computational toxicology, safety biomarker interpretation, and adverse event classification.

CTCAE v5 Grading (Common Terminology Criteria for Adverse Events)

When clinical or laboratory safety data includes adverse event severity, use CTCAE v5 grading:

Grade	Description	Clinical Action
1	Mild; asymptomatic or mild symptoms	Observation only
2	Moderate; minimal intervention indicated	Outpatient management
3	Severe; hospitalization indicated	Dose modification/interruption
4	Life-threatening	Urgent intervention
5	Death related to AE	—

When grading lab values:

Hepatotoxicity (Drug-Induced Liver Injury)

• ALT/AST: Grade 1: >ULN–3x ULN, Grade 2: 3–5x ULN, Grade 3: 5–20x ULN, Grade 4: >20x ULN
• Total bilirubin: Grade 1: >ULN–1.5x ULN, Grade 2: 1.5–3x ULN, Grade 3: 3–10x ULN, Grade 4: >10x ULN
• Hy's Law: ALT >3x ULN AND bilirubin >2x ULN (without ALP >2x ULN) — high risk of fatal DILI
• RUCAM score: Use to assess causality (drug vs other cause)

Nephrotoxicity

• Creatinine: Grade 1: >ULN–1.5x baseline, Grade 2: 1.5–3x baseline, Grade 3: 3–6x baseline, Grade 4: >6x baseline
• eGFR: Monitor trend; acute decline >25% is concerning
• Proteinuria: Grade 1: +1, Grade 2: +2–3, Grade 3: nephrotic range

Cardiotoxicity

• Troponin: Any elevation above ULN warrants investigation
• LVEF: Grade 1: asymptomatic decline, Grade 2: <50% or decline >10%, Grade 3: symptomatic CHF
• QTc prolongation: Grade 1: 450–480ms, Grade 2: 481–500ms, Grade 3: >500ms or >60ms increase

Hematological Toxicity

• Neutrophils: Grade 1: <LLN–1500, Grade 2: 1000–1500, Grade 3: 500–1000, Grade 4: <500
• Platelets: Grade 1: <LLN–75000, Grade 2: 50000–75000, Grade 3: 25000–50000, Grade 4: <25000

Safety Target Panel Quick Lookup

Cortex ships a curated secondary-pharmacology off-target safety panel (~30 well-established liability targets, organ-tagged). Use this as a fast first-pass screen before reaching for live tools.

From the agent:

Screen identifiers — ChEMBL IDs, gene symbols, or UniProt accessions — against the panel with the safety-panel lookup available to you. It returns matched entries plus a summary by severity and organ. An unmatched identifier is not a panel hit, which is not the same as being safe: it may still carry Open Targets signals.

Each matched entry carries:

Field	Contents
chembl_id	ChEMBL target identifier — the join key against candidate targets
gene_symbol	HGNC gene symbol
uniprot	UniProt accession
organ_system	Organ system the liability presents in (cardiovascular, hepatic, …)
severity	Severity tier for the liability
clinical_consequence	The observed clinical effect

From a sandbox script (Python/R):

The panel is a lookup, not a file. It ships inside the agent runtime and is never staged onto disk, so there is no path to resolve and nothing to read — a script that goes looking for a safety-panel CSV is hunting for a file that does not exist in any environment.

To use the panel in a script, look the identifiers up first, then persist the result into your working directory and read it back:

import pandas as pd

# Written from the panel lookup you ran before invoking the script.
panel = pd.read_csv("output/safety_panel_hits.csv")
hits = candidates.merge(panel, left_on="target_chembl_id", right_on="chembl_id")

Screen the whole candidate list in one lookup rather than one call per target, and write the result once — the panel is small and fully in memory, so batching costs nothing and keeps the script's input a single reviewable file.

This panel is NOT exhaustive. It is the well-cited tier-1 set, so a clean screen against it is weak evidence of safety and must never be reported as a clear result. For comprehensive liability assessment also pull Open Targets organ-specific safety signals and in-vitro high-throughput screening data (EPA CompTox / ToxCast) using the target- and chemical-lookup tools available to you. The panel short-circuits the easy cases; the live sources cover the long tail.

CYP Metabolism Liability Assessment

When evaluating drug metabolism risk from omics data:

Key CYP Enzymes

CYP	% Drug Metabolism	Key Substrates	Key Inhibitors
CYP3A4	~50%	midazolam, cyclosporine, statins	ketoconazole, ritonavir, grapefruit
CYP2D6	~25%	tamoxifen, codeine, metoprolol	fluoxetine, paroxetine, quinidine
CYP2C9	~15%	warfarin, phenytoin, NSAIDs	fluconazole, amiodarone
CYP2C19	~10%	clopidogrel, omeprazole	omeprazole, fluvoxamine
CYP1A2	~5%	caffeine, theophylline	fluvoxamine, ciprofloxacin

Pharmacogenomic Variant Impact

Look up known gene-drug interactions in PharmGKB, if a PharmGKB source is available to you — there is no network egress, so its web API is not a route. Where no such source resolves, fall back to the CYP table above and say the pharmacogenomic annotation was not consulted. Flag these metabolizer phenotypes:

• Poor metabolizer (PM): reduced enzyme activity → drug accumulation → toxicity risk
• Ultrarapid metabolizer (UM): increased activity → subtherapeutic levels or excess active metabolite (codeine → morphine toxicity in CYP2D6 UM)

CYP Expression in Omics Data

When analyzing transcriptomic data with safety implications:

• Check CYP gene expression levels in treatment vs control
• CYP induction (upregulation) may indicate enzyme induction → drug interaction risk
• CYP inhibition (downregulation or competitive binding) → toxicity risk for co-administered drugs

Structural Alert Analysis

When compound structures are available (SMILES/SDF), screen against PAINS, Brenk, and NIH structural alert catalogs using RDKit's FilterCatalog module.

Flag but do NOT automatically exclude — structural alerts are context-dependent. Report: alert type, affected substructure, and the specific liability concern.

hERG Liability Indicators

No predictive hERG model is available in the sandbox. Instead:

• Flag compounds with ALogP > 3.5 AND basic nitrogen — known hERG liability correlates
• If a ChEMBL lookup is available to you, retrieve existing hERG bioactivity for the compound against target CHEMBL240. There is no network egress, so this only works through a provided data source — do not write an HTTP call to the ChEMBL API, and if no such source resolves, report the gap rather than substituting a guess
• Report: "hERG liability not computationally assessed; experimental patch-clamp assay recommended for compounds entering lead optimization"

Safety Biomarker Panels

When clinical lab data is available alongside omics data:

Organ-Specific Safety Panels

Organ	Biomarkers	Alert Thresholds
Liver	ALT, AST, ALP, TBil, GGT	Hy's Law criteria
Kidney	Creatinine, BUN, eGFR, KIM-1, NGAL	>1.5x baseline
Heart	Troponin I/T, BNP/NT-proBNP, CK-MB	Any troponin >ULN
Bone Marrow	ANC, platelets, hemoglobin, reticulocytes	CTCAE Grade 3+
Pancreas	Amylase, lipase	>3x ULN

Analysis Approach

1. Compute per-patient max toxicity grade across timepoints
2. Correlate safety biomarker levels with omics features (DE genes, pathway scores) — use Spearman correlation with FDR correction
3. Identify molecular signatures predictive of high-grade toxicity
4. Report as safety biomarker associations, NOT causal claims

Drug-Drug Interaction Assessment

When multiple drugs are in scope:

1. Identify CYP metabolism pathways for each drug (PharmGKB if available to you, otherwise the CYP table above)
2. Flag overlapping CYP substrates — co-administration risk
3. Check for known inhibitor-substrate pairs
4. Search FAERS for the drug combination to assess real-world signal

Pharmacovigilance Signal Detection

When analyzing FAERS data for safety signals, use disproportionality analysis to distinguish real signals from background noise.

Disproportionality Measures

Measure	Formula	Interpretation	Threshold
PRR (Proportional Reporting Ratio)	(a/(a+b)) / (c/(c+d))	Frequentist, easy to compute	PRR > 2, chi-squared > 4, N >= 3
ROR (Reporting Odds Ratio)	(a/b) / (c/d)	Odds ratio analog, wider CI	Lower 95% CI > 1
IC (Information Component)	log2((O + 0.5) / (E + 0.5))	Bayesian shrinkage, WHO-UMC method	IC025 > 0 (lower credibility bound)
EBGM (Empirical Bayes Geometric Mean)	Bayesian shrinkage estimate	FDA standard method	EB05 > 2 (5th percentile)

Contingency table: a = drug+event, b = drug+no-event, c = no-drug+event, d = no-drug+no-event.

Signal Detection Workflow

1. Query FAERS adverse-event reports for the drug of interest.
2. For each adverse event, compute PRR and ROR from FAERS counts.
3. Apply signal threshold: PRR > 2 AND N >= 3 AND chi-squared > 4.
4. Rank signals by PRR magnitude and report count.
5. Cross-reference with known drug label adverse events to distinguish known from novel signals.

Temporal Pattern Analysis

When longitudinal FAERS data is available:

• Time-to-onset distribution for each adverse event
• Early signals (< 30 days) suggest direct pharmacological effect
• Late signals (> 90 days) suggest cumulative or immune-mediated toxicity
• Compare temporal patterns with known mechanism of action

Drug-Drug Interaction Signal Detection

When assessing DDI safety:

1. Query FAERS for drug A alone, drug B alone, and combination.
2. Compute interaction signal: is the combination PRR significantly higher than either drug alone?
3. Flag synergistic toxicity signals (combination PRR > max(PRR_A, PRR_B)).
4. Cross-reference with CYP metabolism data from PharmGKB.

See references/pharmacovigilance.md for PRR/ROR computation functions and visualization patterns.

References

Reference	File	Contents
Pharmacovigilance	references/pharmacovigilance.md	PRR, ROR, IC computation from FAERS data, signal detection workflow, DDI signal analysis, temporal pattern visualization

Do NOT

• Claim computational predictions are equivalent to experimental safety assays (hERG patch-clamp, Ames test, in vivo toxicology)
• Use predicted ADMET properties as definitive safety assessments — they are screening-level signals, not regulatory-grade data
• Present structural alerts as toxicity proof — many drugs have alerts and are safe; context matters
• Skip the "experimental validation needed" caveat for any computational safety prediction
• Assign CTCAE grades without actual clinical lab values — grading requires measured values, not predicted values
• Report disproportionality signals as confirmed causal associations — FAERS data has known limitations (underreporting, confounding, notoriety bias)
• Use PRR alone without minimum case count (N >= 3) — low-count signals are statistically unreliable
• Ignore the denominator problem — FAERS reporting rates change over time and across regions