Investigative study of selective patient genomic profile targeting unseen drugs

Bruce Ho

Bayezene AI

Introduction

Bruce Ho, PhD

Physics background, MIT, Caltech

Harvard Medical School, MGH faculty

Industry career in software development and AI w. Specialty in Healthcare

​

Bayezene AI

A startup with a mission to accelerate clinical trial with AI

Bayesian + Benzene

Problem: High Failure Rate in Clinical Trials

• Only 13.8% of drug candidates successfully navigate all three stages of clinical test
• Reasons include
• Don’t demonstrate efficacy or safety
• Flawed study design
• Shortage of fund
• Participant drop out
• Fail to recruit enough participants
• These failures could result in $B losses

What can Bayezene AI do?

• *Characterize the patients more likely to respond to a drug
• Efficiently mine a large collection of EHR clinical notes in search of the suitable patient candidates (PMSA talk)

Wong, C, etal, Biostatistics 2019

Holy Grail

H_o

Accept

Reject

There is a treasure trove of clinical trial data to be used for training

Recent Breakthrough in Drug Response Prediction - CODE-AE, Xie, et.al. 2022

Out Of Distribution

OOD

Flip the idea on its head

What is Drug Similarity?

• 2D structural fingerprints
• MACCS (for Molecular Access System) structural keys
• codifies the presence or absence of different substructural or pharmacophoric features in each bit position
• Drug drug interaction profile fingerprints (IPFs)
• drugs can be compared on the basis of the similarities between their individual drug interaction profiles
• target (protein) profile fingerprints
• CTET - carrier, target, enzyme, and transport protein
• ADE profile fingerprints
• Drug to adverse effect (abdominal pain)
• 3D pharmacophoric approaches
• identification of atom triplets and the use of a pairwise atom distance distribution approach
• Microbe Drug Interaction
• Mechanism, spectrum, selective toxicity, resistance development

Feed Drug Properties into the Model

Latfollahi, et.al. 2020

basal state

drug properties

Simple MNIST Example

Latfollahi, et.al. 2020

Applied to sc-RNA-seq Data

Pretrained model used 17k drugs

Non-pretrained 188 drugs

Baseline no perturbation input

Hetzel, et.al. 2022

rdkit drug embedding as control signal

Drug Response Classifier

Rampasek, et.al 2018

Control

Perturbed

Semi-supervised learning

Goal of the study

• Train an accurate classifier for predicting whether a patient with certain genomic profile will respond to an unseen drug with only limited pre clinical trial data (ic50)

Steps:

• Find best control signal generated from drug properties to predict the perturbation
• Combine perturbation effect and semi-supervised learning to achieve the highest classification accuracy