We value your privacy

    We use cookies to understand how you interact with our website to improve your experience. By accepting, you agree to our use of these cookies. You can always change your mind later.

    Back to Insights
    IFLAI Research

    How Physics-Informed AI Accelerates Drug Discovery at AstraZeneca

    Henrik Klein Moberg (CTO)
    February 9, 2026
    3 min read

    The bottleneck in modern drug discovery isn't just finding promising molecules; it's the sheer computational and data-gathering friction required to evaluate them at scale. Recently, Invest in Gothenburg highlighted our ongoing work within the AstraZeneca R&D BioVentureHub, focusing on exactly this problem: how to make AI for drug development fundamentally more efficient.

    You can read the full feature here: How AI accelerates parts of drug development.

    The False Promise of Brute-Force AI in Life Sciences

    There is a pervasive assumption in the AI industry right now: if a model isn't performing well, you just need a bigger model and more data. For large language models (LLMs), this scaling law generally holds true. But in life sciences, particularly in areas closely tied to physical phenomena like microscopy, spectroscopy, and X-ray imaging, this brute-force approach is highly inefficient.

    As Robert Roth at AstraZeneca's BioVentureHub noted in the article, the pharmaceutical industry needs teams that don't just understand foundation models, but deeply understand the underlying research.

    When you apply standard, data-heavy AI to early-stage drug screening, you run into severe scaling issues. Training an AI to industrial scale for phenotypic screening or mass spectrometry typically requires millions of annotated data points. In a lab environment, generating and manually labeling that data is exorbitantly expensive and time-consuming.

    Our Approach: Physics Before Data

    At IFLAI, we reject the brute-force paradigm. Instead of relying solely on massive datasets to teach a neural network how the world works, we encode the laws of physics directly into the AI's architecture.

    As I shared with Invest in Gothenburg:

    "We integrate physics knowledge about how the real world works in our AI models long before they see any data. In this way, they do not need millions of data points and thousands of hours of training, but achieve optimal performance from single data points of training."

    By guaranteeing mathematical properties like rotational equivariance and optical transfer functions at the architectural level, our models already "know" the physics of the imaging modality.

    The AstraZeneca Collaboration: Automated Drug Screening

    Through the BioVentureHub, we've had the opportunity to apply this philosophy to one of the most data-intensive processes in pharma: early-stage drug screening.

    We have jointly developed an automated process for drug screening that achieves the same (or better) accuracy as traditional deep learning pipelines, but built on a significantly smaller amount of data.

    By combining physics-informed priors with agentic active learning, we eliminate the need for massive human annotation. The model learns what matters from a handful of examples because it already understands the underlying physical geometry of the cellular structures it's looking at.

    This isn't just an academic achievement; it's a structural shift in how fast a drug candidate can move from initial screening to preclinical validation. When you remove the data bottleneck, you accelerate the entire pipeline.

    We are proud to be driving this shift alongside the team at AstraZeneca. To learn more about our specific architectures for Medtech, explore our Solutions for Medical Technology.