10 live pipelines
What we ingest.
| Source | Cadence | What we extract |
|---|---|---|
| FRED (Federal Reserve) | Daily | Macro indicators feeding regime detection (BULL/BEAR/NEUTRAL) |
| ClinicalTrials.gov | Daily | Trial enrollment status, primary endpoint changes, results posting |
| FDA (Drugs@FDA + MedWatch) | Daily | Approval actions, labeling changes, post-market AE signals |
| SEC EDGAR | Daily | Sponsor 10-Q/10-K/8-K filings — pipeline disclosures, milestone payments |
| PubMed | Weekly | New publication ingestion for evidence synthesis |
| Sensor Tower | Weekly | App-store telemetry (used as alt-data for digital-health programs) |
| Reddit API | Weekly | Patient community sentiment (rare-disease pipelines, AE reports) |
| Google Trends | Weekly | Search-interest signal for indication and brand |
| Glassdoor | Monthly | Sponsor employee sentiment (leading indicator for sponsor risk) |
| GitHub | Daily | Atlas Bio's own model commits + open-source biomedical tooling tracker |
Architecture detail: each pipeline runs as a scheduled job (Windows Task Scheduler) writing into PRODUCTION/alpha-predictor/rwe_pipelines/ and ingested by the calibration module on demand. FRED has a live API key; the others use public endpoints.
How RWE feeds the calculator
Trial → Real-world attenuation.
| Metric | Multiplier | Direction | Mechanism |
|---|---|---|---|
| Gr≥3 TRAE | ×1.20 | ↑ ~20% | Broader patient mix (ECOG 2, comorbidities), more comprehensive EMR-based AE capture |
| Fatal TRAE | ×1.40 | ↑ ~40% | Selection bias strongest at fatal end; community management less intensive |
| Discontinuation | ×1.60 | ↑ ~60% | Community oncologists may discontinue rather than dose-modify |
| ORR | ×0.85 | ↓ ~15% | Non-protocolized imaging, mixed RECIST adherence, broader disease at baseline |
| Median PFS | ×0.85 | ↓ ~15% | Compounding effect of lower response + earlier progression detection variability |
| Median OS | ×0.85 | ↓ ~15% | More advanced/refractory disease, more comorbidity-related deaths |
Symmetric attenuation. Safety inflation and efficacy attenuation are roughly symmetric (×1.4 for fatal vs ×0.85 for OS). Both reflect the same selection bias acting in opposite directions. The risk/benefit ratio observed in trials is therefore neither over- nor under-stated; it shifts but the relative shape may be preserved.
Supporting literature
Where the multipliers come from.
- Bilen 2023 (Flatiron real-world RCC, lenva+pembro): mPFS 14.2 mo RW vs 23.9 mo CLEAR (~40% attenuation in unselected pop)
- Hatakeyama 2024 (Japanese RW RCC, lenva+pembro): Gr3 TRAE 88% RW vs 71.6% CLEAR; disc 52% vs 37%
- Adra 2023 (real-world endometrial lenva+pembro): Fatal TRAE 8.2% RW vs 5.7% KN-775 (~45% higher)
- Pinato 2022 (IMbrave150 RWE, HCC atezo+bev): Trial-to-RWE Gr3 ratio ~1.25×; fatal ratio ~1.4× — cross-combo applicability
- Khaki 2021 (Pembrolizumab mono meta-RWE): RW Gr3 irAE ~1.2× trial; pneumonitis ~1.3× higher
Limitations
What RWE multipliers can't do.
- Population-level, not patient-level. They cannot identify which specific patients drive the divergence.
- Risk of double-counting. When patient-level inputs (age ≥75, ECOG 2) already capture much of the fragility, the RWE multiplier risks adding additional attenuation on top of an already-attenuated patient profile.
- Static estimates. The multipliers are calibrated against historical RWE. As community oncology evolves (e.g., more comprehensive EMR capture, better community supportive care), the multipliers should evolve too. v3.0 roadmap includes quarterly recalibration.
Use the trial/RWE toggle to compare both views side-by-side rather than relying on either in isolation. The honest answer for any given patient is probably somewhere between the two.
Want to add a custom RWE pipeline?
The 10 listed pipelines are the operational set. We can add custom EMR cohorts (Flatiron, OPTUM, your own institution's data) to a briefing for patient-level calibration.
Contact Us →