Tech & AI LIVE: Gen AI – AI & Drug Development, AstraZeneca

December 06, 2024

undefined mins

Nikolay Burlutskiy, Director of ML & AI at AstraZeneca discusses how Gen AI is revolutionising drug development

At Tech & AI LIVE: Gen AI, Nikolay Burlutskiy, Director of ML & AI at AstraZeneca, sheds light on how generative AI is revolutionising drug development. Highlighting its potential to accelerate processes and enhance precision, Nikolay discussed AI’s transformative role in preclinical and clinical stages, particularly in pathology.

The challenges of traditional drug development

Nikolay began by outlining the complexities of drug development, describing it as a multi-stage, decade-long process costing billions.

“Drugs often fail at every stage, from discovery to regulatory approval,” he notes.

For pathologists, critical tasks include determining drug efficacy, identifying toxicity and predicting therapeutic outcomes. Traditionally, this involves laborious manual analysis of tissue samples under a microscope—a time-consuming and error-prone approach.

Nikolay Burlutskiy, Director of ML & AI at AstraZeneca

Empowering pathologists with AI tools

AI tools are now enabling pathologists to process digitised tissue samples more efficiently. High-resolution microscopy images, often comprising billions of pixels, are analysed using computer vision and machine learning models.

These AI systems assist in:

Segmenting malignant areas: Identifying and quantifying tumours or inflammation.
Staging diseases: Classifying cancer stages or severity levels.
Predicting outcomes: Estimating patient survival rates or optimal treatments.
Assessing drug toxicity: Ensuring drugs are both effective and tolerable.

“These tools excel in detecting minute malignant areas that might be overlooked manually, significantly enhancing accuracy,” Nikolay explains.

Nikolay Burlutskiy, Director of ML & AI at AstraZeneca

Generative AI applications in pathology

Nikolay details three primary generative AI methodologies transforming pathology:

1. Transformers

Generative pre-trained transformers (GPTs), initially associated with text, now enable self-supervised learning for image analysis. Nikolay shares a case where transformers generated heatmaps to highlight inflammation in biopsy images.

“The redder the region, the likelier the presence of inflammation,” he says, emphasising how these heatmaps guided pathologists to focus on critical areas without manual annotation.

2. Variational Autoencoders (VAEs)

VAEs compress large datasets into compact representations while retaining essential features. Applied to pathology, these models encode complex cancer features, enabling downstream tasks such as drug response prediction.

VAEs help us understand tumour characteristics and anticipate adverse events,” Nikolay adds.

3. Generative Adversarial Networks (GANs)

GANs improve the generalisability of AI models by addressing domain shifts. Nikolay highlights a use case where images from one scanner were re-styled to resemble training data from another. This approach increased model performance by 16%, enhancing its adaptability across different lab environments.

Visual language models: A step forward

Visual language models (VLMs), combining images and text, represent the next frontier. These models integrate visual and textual data to enhance interpretability and precision.

“While GPTs focus on text, VLMs extend their utility to image-based analysis,” Nikolay explains, emphasising the potential of VLMs to provide context-aware insights for complex datasets.

Nikolay Burlutskiy, Director of ML & AI at AstraZeneca

Driving innovation with generative AI

Generative AI has already shown immense potential in optimising workflows, reducing human error and delivering faster insights in pathology and drug development. By automating laborious tasks and providing actionable insights, AI allows pathologists and researchers to focus on higher-order problem-solving.

Nikolay concluded with a clear message: generative AI is not just a tool but a catalyst for innovation in pharmaceutical research.

Its applications, from identifying malignant regions to predicting patient outcomes, are streamlining processes, improving accuracy, and ultimately saving lives. As the technology continues to evolve, its role in transforming drug development will only grow stronger.