Rapid evaluation of novel chemistry
by our AI platform
reduces the risk of failure

Heterogeneous data is an asset, not noise

Most scientists and ML practitioners discard subtantial amounts of bioactivity data because absolute values don’t reproduce across assays. We treat that as a modeling problem, not a data problem.

  • Learn rank orders, not absolute values — they reproduce across assays
  • Use scaffold-agnostic molecular representations — generalize beyond a chemical series
  • Result: models that zero-shot generalize to new chemistry

Define the indication first, then design the molecule

Indication-first parallel design beats potency-first sequential discovery

Comparison: traditional potency-first discovery funnel — derivatize known chemistry, find an active hit, retrofit ADME/Tox, retrofit exposure and regimen — versus ArrePath’s indication-first parallel design with a candidate-profile pipeline feeding the AI/ML design engine for synthesis and testing.

Three ways the platform delivers

3×
Actionable hits
vs. unguided screening
>80%
Oral bioavailability
From <10% in lead series
Zero-shot
Generalizes to new chemistry
Where public models fail

~3× more actionable hits from ML-guided screening

We cherry-picked 5,000 compounds from a 400K library using three predicted properties: high antibacterial activity, low cytotoxicity, and dissimilarity to known antibacterials.

  • 2× higher primary hit rate
  • 3× fewer cytotoxicity failures
  • Net: 3× more compounds worth following up on

This work seeded multiple series for our NTM program.

Histogram comparing measured antibacterial inhibition across cherry-picked, non-cherry-picked, and 400K-library compound sets — cherry-picked compounds show a higher fraction reaching high-inhibition values. Histogram of cytotoxicity-window distributions for ML-guided versus unguided compound screens — ML-guided compounds skew toward larger cytotox windows, indicating fewer cytotoxicity failures.

Our models generalize where public models fail

Predict oral bioavailability of our lead-series compounds — our model had no training data on this chemical series.

  • ArrePath model: AUC-ROC ≈ 0.72 (zero-shot on the lead series)
  • Public Random Forest: ≈ 0.56
  • Public GNN: ≈ 0.51 (essentially random)

Cross-scaffold, cross-assay rank-order training is what enables generalization.

Bar chart of zero-shot oral bioavailability classification AUC-ROC on Project 1 (n=70), bootstrap mean ± 1 SD — ArrePath model approximately 0.72, public Random Forest approximately 0.56, public GNN approximately 0.51.

ML-enabled microscopy maps mechanism and flags novelty

Built originally for hit-finding work; remains a complementary capability for any antibacterial program

Mapped antibacterial phenotypic space

Two-dimensional projection of antibacterial compounds colored by mechanism class — visualizing the mapped phenotypic space across known antibacterial mechanisms.
Phenotypic projection highlighting compounds prioritized for mechanistic novelty — separating mechanistically novel hits from compounds resembling known antibiotic classes.

ML models trained on imaging of bacteria treated with known antibacterials predict mechanism of action and flag mechanistically novel compounds. Useful when triaging large hit sets early in a program. Models can be retrained for other therapeutic areas as needed.

See the platform in our pipeline

Project 1 (LpxH inhibitor for outpatient UTI) and Project 2 (NTM lung disease)

Strategy & Pipeline →