Inspiration
Drug development is costly (≈\$2.6B per approval) with <10% success. Analysts are buried in fragmented registries, papers, filings, and market signals. We built a physics‑informed, multimodal AI to replace hunches with transparent probabilities, rankings, and timelines that decision‑makers can trust.
What it does
NexJen Bio turns disparate clinical, scientific, and market signals into actionable intelligence:
- Insight Hub (web): a 360° dashboard with live/historical trial data, Probability‑of‑Success (PoS), enrollment‑risk forecasts, design red‑flag detection, competitive‑density heatmaps, and integrated BD/IPO outlooks.
- Risk‑Scoring API: JSON endpoints that deliver PoS plus rationale and evidence links for CRO/LIMS or portfolio tools.
- Paper & Signal Feed (mobile): a “PubMed/bioRxiv‑meets‑ClinicalTrials” recommender with on‑the‑go alerts for new papers, model updates, and points of interest.
- Core engines: RAG literature synthesizer (MOA, safety, KOL sentiment), clinical‑trial tracker (phase‑transition curves, enrollment forecasts), and a BD & IPO scoring module that cross‑references pipeline quality, runway, and comparable exits.
How we built it
We fuse multimodal data from ClinicalTrials.gov/EudraCT, PubMed/bioRxiv, SEC filings, and market feeds in a cloud‑native pipeline.
Models
- GBDT PoS estimator: trained on 3,5000 historical phase transitions (Phase I to III).
- GNN: reads target–disease–MOA graphs to capture biological structure.
- Temporal survival + transformer ensemble: blends time‑to‑event dynamics with text/context understanding.
- IPO forecaster: stacked LSTMs + XGBoost trained on 1000 biotech IPOs.
Basic Math
- Phase-advance probability (at horizon) via survival:
$$ S(t) = \exp!\left(-\int_{0}^{t} h(u)\,du\right), \qquad \hat{p}_{\text{advance}} = 1 - S(t^{\ast}). $$
- Ensemble for PoS:
$$ \hat{y} \;=\; \alpha\,\hat{y}{\mathrm{GBDT}} \;+\; \beta\,\hat{y}{\mathrm{Survival}} \;+\; \gamma\,\hat{y}_{\mathrm{Transformer}}, \qquad \alpha + \beta + \gamma = 1. $$
- Every score includes feature attribution and evidence-trace explanations to keep results auditable and explainable.
Challenges we ran into
- Data wrangling: de‑dup across registries, harmonize endpoints/indications, map MOA taxonomies, and handle missingness without label leakage.
- Class imbalance & calibration: rare successes required careful resampling, hierarchical calibration by indication/phase, and drift monitoring.
- Trust & explainability: showing why a score moved—linking features, literature snippets, and comparables—was as important as raw accuracy.
- Privacy & compliance: GDPR/HIPAA‑aware data flows and safe retrieval for literature and filings.
Accomplishments that we’re proud of
- A single pane of glass unifying clinical, scientific, and financial signals.
- A first‑of‑its‑kind pairing of BD scoring and IPO timing/valuation alongside clinical PoS.
- Real‑time literature RAG producing concise, source‑linked MOA/safety/KOL digests.
- An integrated demo spanning Insight Hub, Risk‑Scoring API, and Paper & Signal Feed.
What we learned
- Blending mechanistic intuition with data‑driven models outperforms either alone.
- In this domain, explainability = usability; scores must ship with evidence trails.
- Value compounds when science ↔ clinical ops ↔ finance live in one workflow.
- Calibration and survival modeling are critical for timelines and enrollment.
What’s next for NexJen Bio
- Finalize product: harden web, API, and mobile; expand CRO/LIMS integrations and data vendors.
- Partnerships & pilots: strategic CRO/LIMS partnerships and enterprise pilots; deeper competitive‑intelligence analytics.
- Milestones: broaden indication coverage, enrich BD/IPO comparables, and target **profitability within 3 years via subscription growth and retention.