MBRS is a biology-first system designed to surface differentiation, mispricing, and landscape shifts before clinical outcomes are known.
We quantify how drugs interact with biology, how those interactions translate into disease-relevant effects, and how each program compares within its competitive landscape.
Systematically model drug–biology interactions to identify the underlying signals that drive efficacy and risk.
Evaluate whether those signals restore, conflict with, or amplify disease biology to infer likely outcomes.
Position each asset within its indication landscape to measure competitive density and biological advantage.
Combine signals into standardized outputs that enable cross-program comparison and decision-making.
Start with a processed asset–indication already in MBRS or score a program of your own. MBRS Standard then provides the common broad triage pass; a selected compiled Perspective Landscape adds a specialty- or context-specific higher-conviction lens.
Use the registered universe to inspect a program that already has an MBRS Standard position and supporting views.
Use MBRS AI for program documents and/or Private Compound Submission for account-private compound analysis.
The fixed MBRS Standard is the common triage coordinate across drug–indication programs. It is designed to cover the broad universe and identify what deserves deeper diligence before a more contextual lens is applied.
Each Landscape is independently compiled around a defined evidence or specialty scope. It re-reads the same MBRS universe through that context, allowing higher-conviction selective triage without changing the Pass 2 Standard.
Every Landscape starts from the same frozen MBRS Standard coordinate. A defined contributor scope is then used to form a re-aligned perspective over that same registered universe. The Landscape does not replace or retrain the Standard; it changes the context through which the Standard is interpreted.
Pass 2 stays fixed as MBRS Standard. Every Landscape remains interpretable against that same common coordinate.
A Landscape is compiled from an explicit contributor scope representing a specialty, biological, clinical, indication, or diligence context.
Oncology, Cardiology, Immunology, Aging, and other Landscapes are formed independently. They do not average into one another, so each can express a different conviction lens.
The Landscape validation readout describes how well that completed lens separates known outcomes across the shared labeled cohort. It is a property of the perspective as a decision instrument, not a claim that every individual program inherits the displayed accuracy.
This section retains the detailed MBRS Standard validation for Pass 2. The three-pass workflow above separates program selection/submission from broad Standard triage and from the scoped Landscape precision pass; indication- and specialty-specific Landscapes can occupy Pass 3 while MBRS Standard remains the common Pass 2 reference.
Canonical decisive subset: 60% accuracy on 97% of labeled approval/failure calls. This is the primary MBRS Standard readout because it applies the authority-defined decisive-call rule used by the baseline and Visual Rings. Broad approval/failure alignment has improved (68% balanced accuracy), and canonical baseline OOF AUC is 0.74; this means the current run should still be framed as biological risk triage rather than standalone approval prediction. Class performance is now relatively balanced: Failed precision 68% / recall 63%, Approved precision 68% / recall 73%. This demo set is heavily oncology-enriched, where real-world approval decisions often tolerate higher biological and safety risk than other indications. That makes approval/failure labels a noisy downstream comparator, not the biology MBRS is directly modeling.
Use MBRS as a biological risk-prioritization layer: decisive calls help flag programs that deserve deeper diligence, comparison, or caution.
MBRS is not presented as a standalone regulatory approval predictor; lower precision and calibration scores are shown to keep that boundary explicit.
The current demo set is oncology-heavy, where approval decisions can accept more toxicity and uncertainty than many other indications. That can blur direct approval/failure calibration.
| Decision setting | Reported benchmark | Why it matters for MBRS |
|---|---|---|
| Drug development base rate | ~14% Phase I→approval (2006–2022) | Low base rate → small signal improvements matter |
| Current clinical success environment | 10.8% overall success (2023); Phase I 48%, Phase III 66% | Uncertainty persists late → biology signal remains relevant |
| VC portfolio reality | Majority of VC funds underperform public markets | Power-law outcomes → filtering matters more than prediction |
| BD / M&A execution | Frequent miss on timelines and revenue post-acquisition | Forecast error risk → biology adds downside checks |
| MBRS decisive-call subset | 60% accuracy on 97% of labeled approval/failure calls. | Core signal: authority-defined decisive biological separation |
| Class | Precision | Recall | F1 | Correct / total | Mean margin | Calibration | Readout |
|---|---|---|---|---|---|---|---|
| Failed | 68% | 63% | 65% | 160 / 25595 missed | 0.8691% strong-margin calls | ECE 0.35Brier 0.37 · still calibration-limited | Failure detection is directional; precision is useful, so the signal is useful for risk triage but not a standalone verdict. |
| Approved | 68% | 73% | 70% | 202 / 27876 missed | 0.9395% strong-margin calls | ECE 0.35Brier 0.37 · still calibration-limited | Approval calls are performing well enough to support stronger positive-readout language. |
Category margin shows how separated the model’s top biological call is from the competing class. This is the current advantage: margin filtering reveals the subset where MBRS is most informative, without overstating it as outcome prediction.
Simple readout: the MBRS Standard baseline is summarized by its canonical AUC, decisive-call accuracy, coverage, and calibration. Category margin ≥0.50 remains a separately labeled diagnostic from the same 533-row OOF authority, not a competing baseline definition.
Note: this page is intentionally high-level. Proprietary methods and model details are omitted.