Immunomodulation of
Tumor Micro-environment

Only 1 in 4 patients respond to Anti-PD1 therapies like Keytruda. We need to find treatments that shift the patient's phenotype from 'Non-Responsive' to 'Responsive'.

Generating Patient Digital Twins

We generated digital twins for 10 unseen cancer patients—5 sensitive and 5 resistant to Anti-PD1 therapy.

We then tasked TRAILBLAZER to simulate Anti-PD1 therapy to see if it can accurately reconstruct patients' responses.

Study Design

5

Sensitive Patients

Responsive to Anti-PD1 baseline

5

Resistant Patients

Non-responsive to Anti-PD1 baseline

TASK: Virtual Clinical Trials to predict responses at molecular level and phenotypic responses of real patients to virtual anti-PD1 treatment.

Per-Patient Responsiveness to Anti-PD1

TRAILBLAZER reconstructed the responsiveness distribution of 10 unseen patients. Responsive patients (P1–P5) score consistently above 0.5; non-responsive patients (P6–P10) score below.

0.00.20.40.60.81.0ResponsivenessDecision boundary (0.5)P1P2P3P4P5P6P7P8P9P10Patient
Responsive (P1–P5)
Non-Responsive (P6–P10)
Decision boundary
93%
ROC AUC

Combinatorial Anti-PD1 Treatment Predictions

TRAILBLAZER performs a zero-shot virtual drug screen by ranking treatments most likely to lead to desirable α-PD-1. Accuracy measures corroboration of top predictions against published clinical literature, evaluated on unseen breast cancer patients.

TRAILBLAZER83%CELLFLOW33%

Predictions validated against published clinical literature on treatments augmenting immune checkpoint inhibitor response. CELLFLOW is the leading neural network benchmark trained on the same data foundation.

2.5×
vs state-of-the-art

Patient-Specific Treatment Rankings

Starting from each patient's non-responsive phenotype, TRAILBLAZER independently ranks candidate treatments by their predicted ability to shift the immune landscape toward α-PD1 responsiveness. The model re-discovers the clinically validated α-PD-1/IL-15 combination at the correct patient-specific rank — and proposes 10 novel first-in-class candidates (ANU-01–10) now in active validation. Hover any cell to highlight all patients sharing that treatment.

α-PD-1 NON-RESPONSIVE PATIENTSPatient 1Patient 2Patient 3Patient 4Patient 5Patient 6Patient 7#1#2#3#4#5ANU-02ANU-02ANU-02ANU-02ANU-03ANU-03ANU-03ANU-01ANU-01ANU-01ANU-03ANU-04ANU-04ANU-04α-PD-1/IL-15ANU-04ANU-03ANU-01ANU-07α-PD-1/IL-15α-PD-1/IL-15ANU-06α-PD-1/IL-15ANU-08ANU-08α-PD-1/IL-15ANU-07ANU-07ANU-04ANU-06α-PD-1/IL-15α-PD-1/IL-15ANU-09ANU-02ANU-02
Corroborated in human trials — α-PD-1/IL-15
Novel first-in-class (ANU-01–10)
10
Novel treatments in validation
93%
ROC AUC

For simulating individual patient responses

83%
Prediction Accuracy

Match rate with clinical outcomes

2.5x
Zero-shot recall

Compared to state-of-the-art

10
Novel Combinations

Identified for further study

What This Means for Patients

By accurately predicting which combination therapies can convert non-responders into responders, TRAILBLAZER enables personalized treatment strategies that maximize patient outcomes.

This approach reduces trial-and-error in clinical settings and accelerates the path to effective immunotherapy combinations.

$350M

Saved per approved drug for every 10% improvement in predictive positive value of clinical trials.

5–15%

Current in vitro and in vivo pre-clinical predictive positive value — the benchmark TRAILBLAZER is built to surpass.