Physics-Grounded.
AI-Driven.
Drug Discovery.

Request early access Explore TraceDDE

A model can suggest a molecule. It cannot, by itself, tell you the suggestion is physically serious.

[01] Physics filter

We exist so that an AI proposal meets physics before it is treated as evidence.

AI vs Physics Check

[02] Core engine

TRACE is our molecular simulation engine.

High-Throughput Kinetics

[03] Application

TraceDDE™ is how that engine is used to verify candidates in drug discovery.

Candidate Verification Pipeline

From Molecular Hypotheses To Real World Impact

A generative model proposes molecular hypotheses.

Generative and predictive models create and rank diverse, synthesizable candidates conditioned on your target and project goals.

THROUGHPUT: 50,000Candidates generated per campaign
Novelty (to training set): 0.72Tanimoto distance (median)
Diversity: 0.41Median pairwise Tanimoto similarity
Drug-likeness: 0.68QED (median)

TraceDDE verifies the physics.

Using the TRACE engine, TraceDDE evaluates each candidate with adaptive enhanced sampling and applies rigorous, claim-specific scientific gates for pose, kinetics, and affinity.

Pose verification
(Tier 1)

78.6% Pass

Calibrated on blinded set

Pose stability, reproducibility, reference agreement

Kinetic endpoints
(Tier 2)

Δlog k RMSE 0.32 ± 0.05

Chignolin (known-rate recovery)

Commitment analysis, reweighting, uncertainty

Affinity estimates
(Tier 3)

ΔΔG RMSE 0.78 kcal/mol

vs. OpenFE (public series)

Alchemical FEP (OpenFE), cycle closure, uncertainty

Example: Binding pose

TraceDDE
Initial pose

TraceDDE selects, learns, and drives experiments.

TraceDDE uses validated evidence and uncertainty-aware learning to choose the next molecules to simulate or synthesize, closing the loop with experimental results.

GenerateCandidates
Verifywith TraceDDE
DecideNext best experiments
Learnfrom results

Improving decisions over time

Decision utility: 2.3× vs. randomExperimental hit rate @ fixed budget
Cost per useful candidate: 68% lowervs. generator-only selection
Experimental yield (pilot): 27.4% (14 / 51 tested)Biochemical assay hits

TRACE Performance

TRACE is up to 44× faster than OpenMM + PLUMED, with the same level of accuracy, for selected protein systems.

TRACE is highly optimized for NVIDIA GPUs.

See detailed performance benchmark results

We are opening TraceDDE to a small number of design partners.

Request early access

Exclusive Access
Direct Collaboration
Resource Library
Shape the Future