Platform Architecture
Built for the polypharmacy complexity that standard disproportionality analysis wasn't.
Not a dashboarded version of ROR. A graph-based detection model that treats the FAERS corpus as a network of drug-AE co-occurrences — not a table of pairwise counts.
Detection Architecture
Three-layer analysis engine
Each layer adds signal specificity the previous one cannot achieve alone.
Graph construction from co-prescription and AE co-occurrence
TrialVyx builds a heterogeneous graph where nodes represent individual drugs and adverse event preferred terms (MedDRA PT). Edges encode co-occurrence frequency across the full FAERS corpus — not just pairwise drug-AE associations, but cross-drug co-prescription weights drawn from reported drug lists in each ICSR. The result is a network that captures the polypharmacy reality of 40% of post-market adverse event reports.
Bayesian network propagation across drug nodes
A Bayesian propagation model flows probability mass through the graph from newly observed co-occurrences toward connected drug nodes. When a new adverse event report adds evidence to one drug pair, the model updates the posterior probability of associated drug combinations — allowing indirect interaction signals to emerge before enough direct co-occurrence data exists to trigger conventional statistical thresholds.
Temporal signal scoring with recency weighting
Raw detection probability is adjusted by a temporal decay function that upweights recent adverse event submissions while preserving historical signal context. Reports filed in the most recent 180-day window receive a recency multiplier — capturing emerging signals before accumulation reaches conventional thresholds. The combined score drives prioritization: Monitor, Investigate, or Escalate.
Data Sources
Three regulatory ICSR databases, unified in one detection corpus
| Source | Coverage | Update Frequency | Volume |
|---|---|---|---|
| FDA FAERS | US adverse event reports since 1990. Consumer, healthcare professional, and manufacturer submissions. | Quarterly (FDA), 180-day refresh | 20M+ reports |
| EudraVigilance | EU/EEA suspected adverse reactions. Includes suspected ADR reports for human medicines authorized in the EEA. | Continuous (EMA) | 30M+ reports |
| VigiBase / WHO UMC | Global ICSR database, 130 countries via national pharmacovigilance centers. | Monthly (WHO UMC) | 30M+ ICSRs |
| ClinTrace Network | Proprietary de-identified ICSR feed from participating clinical data networks. Supplemental co-prescription data. | Weekly | Supplemental |
Signal Output
Signal brief output: structured for PSMF review and GVP IX documentation
Each signal brief contains everything a PV scientist needs to make an initial triage decision — drug combination, AE PT code, IC score, contributing ICSRs, and recommended action tier.
-
01
Signal brief
Drug pair(s), adverse event PT code, IC score, trend direction, detection date.
-
02
Supporting evidence
Top 10 contributing ICSRs with FAERS report numbers, reporter type, and seriousness flags.
-
03
Regulatory context
MedDRA SMQ mapping, relevant REMS programs, WHO-ART coding for cross-reference.
-
04
Recommended action tier
Monitor / Investigate / Escalate — determined by IC score, temporal trend, and clinical severity.
Validation methodology: retrospective FAERS signal corpus, 2019–2022
TrialVyx's detection model was validated against 312 historically confirmed drug interaction signals from the 2019–2022 FAERS record. For each signal, we identified the date at which our model would have generated an Investigate or Escalate tier alert — then compared that date against the date the signal was confirmed through conventional post-hoc disproportionality analysis.
This gives a reproducible, honest lead time measurement. No synthetic benchmarks, no modeled projections. If your PV team maintains an internal corpus of historically confirmed signals from your compound portfolio, we can run the same retrospective validation against your compounds before you commit to a contract.
Numbers derived from internal retrospective analysis. Results may vary for specific compound portfolios or therapeutic areas. Not a prospective clinical claim.
Data Security
Security by design
No PHI ingested
FAERS and EudraVigilance data is de-identified at source by the regulatory agencies. TrialVyx does not receive, store, or process individually identifiable patient data at any point in the pipeline.
SOC 2 controls in design
Access controls, audit logging, and data handling procedures are designed with SOC 2 Type II criteria in mind. Certification process ongoing — we do not claim current certification.
Immutable audit trail
Every signal access, triage decision, and suppression is logged with timestamp, user ID, and action detail. Exportable in FDA/EMA inspection-compatible format. Full security overview.
Walk through the detection model with your compound portfolio.
Request a methodology briefing and we'll show how the N-drug graph analysis applies to your specific therapeutic area — including a retrospective run against your historical FAERS data if you have a confirmed signal corpus.
Talk to the team — [email protected]