D-MPNN today. Causal mechanism-of-action by 2027.
For computational biology and target-discovery teams. Rosenbound's D-MPNN molecular property prediction stack is validated on the standard MoleculeNet scaffold splits (BACE, BBBP, HIV) and is available today as a benchmark-locked capability. Active causal-MoA productization — treating drug-target effects as causal estimands with explicit sensitivity bounds — is deliberately deferred to 2027 while we focus the platform on pharmacovigilance and real-world evidence.
Sequencing is everything in a solo-founder platform.
Drug discovery is a richer technical surface than PV or RWE. Productizing it well means doing it after the methodology platform has design-partner traction, not before.
The Rosenbound platform's underlying causal-inference engine works as well on drug-target effects as on adverse-event signals — structurally, the math is the same. But the team is different (comp-bio group vs. drug-safety reviewer), the data is different (assay panels and SMILES strings vs. ICSRs), the workflow is different (target validation and lead optimization vs. signal triage), and the regulatory framing is different (no 21 CFR Part 11 in early-stage discovery).
Building all four verticals in parallel would dilute each of them. Pharmacovigilance and real-world evidence come first and go into Design Partner hands; full attention turns to drug discovery in 2027.
What that means for you today: the D-MPNN benchmark capability is real and citable, and the platform technology is portable. The active vertical is on the 2027 roadmap.
D-MPNN benchmark capability, production-validated.
In-house PyTorch implementation of the Directed Message-Passing Neural Network architecture (Chemprop v2 parity). Validated on the three standard MoleculeNet scaffold-split benchmarks. Ready to extend to your in-house assay data as a Design Partner co-authorship.
BACE benchmark
β-secretase inhibition (Alzheimer's target). Rosenbound AUROC 0.8861 ± 0.001 vs Chemprop v2 reference 0.859 ± 0.024. Within published σ-overlap. Scaffold split, no ensembling.
BBBP benchmark
Blood-brain barrier penetration. AUROC 0.9144 ± 0.0113 vs Chemprop v2 reference 0.897 ± 0.012. Within published σ-overlap.
HIV benchmark
HIV replication inhibition. AUROC 0.7937 ± 0.0149 vs Chemprop v2 reference 0.776 ± 0.020. Within published σ-overlap.
Native PyTorch rewrite, ~500 lines
No Chemprop runtime dependency. 5-test gradient-check suite green. Bemis-Murcko scaffold splits enforced. Published-baseline parity under stricter conditions than the reference paper.
Causal-substrate interoperability
The same causal-inference engine that powers PV and RWE consumes molecular-property scores as treatment-effect inputs. The technical substrate is ready; the discovery-specific UX and the causal-MoA estimator are the 2027 work.
Co-authorship on assay-specific benchmark
If your team has an in-house assay panel you'd like benchmarked, a Design Partnership includes Rosenbound running the D-MPNN stack against your data and co-authoring the resulting benchmark publication (under your IP terms).
What gets built when drug-discovery productization starts.
Deferred — not abandoned.
The drug-discovery build activates once the pharmacovigilance and RWE verticals have each completed their first Design Partnerships. Estimated activation: Q3 2027.
Planned scope: causal mechanism-of-action estimator integrating the D-MPNN molecular feature space with the platform's neural counterfactual estimator; assay-panel ingestion UX; target-validation workflow with sensitivity-bounded effect estimates; explicit per-target Γ-bound on every MoA claim.
Early-access program will open Q1 2027 to comp-bio teams interested in shaping the discovery-specific UX. Email harsh@rosenbound.com to be on the early-access list.
Watch the platform walkthrough at rosenbound.ai — three moments: the Cognitive Validation Report refusing incoherent data, the live Γ-bound sensitivity visualization, and the reproducibility certificate generated on every study. The full platform is open to Design Partners.
pip install rosenbound
— Official Python SDK for programmatic access: cohort upload, sensitivity-bounded study runs, and reproducibility certificate retrieval. Apache 2.0; Pydantic v2 typed; py.typed for IDE autocomplete + mypy. Platform access gated by Bearer token + RBAC + tenant scoping — the SDK is open, the audit substrate is not.
Interested in shaping the 2027 discovery roadmap?
Comp-bio teams who want input on the causal-MoA UX, the assay ingestion path, or the target-validation workflow are welcome to apply for early-access status. No commitment, no cost — just a structured feedback channel with the founder during the design phase.