Human Intelligence combined with
Multimodal AI/ML results
in Accelerated Drug Discovery

ArrePath’s ML models enrich for progressible compounds prior to screening

Cherry-pick diverse screening compounds using ML models to select compounds that are antibacterial and not cytotoxic while excluding structures similar to known antibiotics.

Arrepath-graph-predictedEcoli-activity 1

ML guided selection yields 3x as many progressible compounds as screening using traditional medicinal chemistry guidelines

Compounds were:

  • 2x more likely to be active
  • 3x less likely to be cytotoxic
  • 1.5x more likely to be active against WT E. coli
Arrepath-large_chem_space

ArrePath’s ML-enabled microscopy rapidly
predicts mechanism & mechanistic novelty

ML models are currently trained on known antibacterials and can
be trained for other therapeutic areas as needed

Mapped antibacterial phenotypic space

Arrepath-Mapped-antibacterial-phenotypic-space 2
Arrepath-Novelty-prioritized-1

We predict >6 properties for every molecule in our lead program

Molecules are prioritized based on the following parameters:

Predicted MIC or other potency measure

Predicted MIC or other potency measure

Predicted microsomal clearance

Predicted microsomal clearance

Predicted logD

Predicted logD

Predicted cytotoxicity

Predicted cytotoxicity

Docking to target structure

Docking to target structure

Compounds meeting minimal criteria are sub-selected for Pareto-optimality

Arrepath-pareto_plot 1

Rigorous usage of ML models doubled success rate

Using AI / ML to drive hit to candidate activities

Arrepath-BarGraph 1

Rigorous prioritization of AP-001
compounds for synthesis using a predictive
ML model of WT E. coli MIC doubled the
fraction of molecules with target MICs

Model was used in conjunction
with predictive models of
clearance, toxicity, and multiple
other ADMET properties to select
compounds for synthesis

Arrepath-ML-success-rate 1

ArrePath couples virtual and experimental compound
testing to accelerate compound optimization

Using AI / ML to drive hit to candidate activities

Arrepath-ML-Models-Chart 2

Our Philosophy

Rigorously evaluate models in real-world use

Test multiple models on novel compounds as they are synthesized

Promote the most predictive models to decision-making

Automatically retrain models as new data is acquired

Use all available data

Learning-to-rank models allow ingestion of multiple heterogeneous data sets (on different strains, different endpoints, etc.).

Multiple real-world tests show the resulting models outperform models trained on smaller homogeneous data sets

Integrate technologies

Leverage physics-based models such as docking, molecular dynamics, and quantum mechanics when these outperform or augment ML models

Leverage human knowledge and understanding to integrate medicinal chemistry principles not captured in our ML models

Reduce cycle time

Self-serve tools allow rapid exploration of new designs with ML models.

Parallel cloud deployment reduces compute time

Technology to alleviates bottlenecks in compound design and ordering

ArrePath’s lead program
targets an unrealized MOA

AP-001: a clinically novel MoA Gram
negative antibacterial

  • Discovery of novel target/inhibitor using ArrePath’s
    platform: < 18 months to LO
  • Development of AP-001 is accelerated by ArrePath’s
    proprietary AI/ML models
Arrepath-InitialHit-Graphic 1

Initial hit identified by ArrePath’s platform
Structure of AP-001 target with docked inhibitor

Status

Lead optimization

Unrealized Gram-negative target

Unrealized Gram-negative target, highly conserved and no human homolog

Target validated in vitro and in vivo

Predicted MIC or other potency measure

Potent activity vs. Enterobacterales

Promising ADMET profile, including oral bioavailability 

Promising ADMET
profile, including oral
bioavailability

Strong Intellectual Property estate

Strong Intellectual Property estate

Next key decision point

Development Candidate nomination