Thesis work, 30 credits (2 students) Generative Learning in Drug Discovery

AstraZeneca is seeking a Masters thesis student to explore a central question in AI-driven medicinal chemistry: How much information do we truly need to transform an initial hit into a candidate drug You will leverage deep learningbased generative models to accelerate the transition from weak hits to robust candidate drugs quantifying the informational content required along the way. The work builds on prior graduate student foundations and uses AstraZenecas proprietary generative platform ReInvent alongside public benchmarking datasets.

About AstraZeneca:

AstraZeneca is a global science-led patient-centered biopharmaceutical company focusing on discovering developing and commercializing prescription medicines for some of the worlds most serious diseases. But were more than a global leading pharmaceutical company. At AstraZeneca were dedicated to being a Great Place to Work and empowering employees to push the boundaries of science and fuel their entrepreneurial spirit.

About the Opportunity:

As a Thesis Worker at AstraZeneca youll find an environment thats full of unique opportunities and exciting challenges. Here youll have the opportunity to pursue your areas of interest whilst equally developing a broad skillset and knowledge base to get the best out of your experience. Youll be working on meaningful projects to make an impact and deliver real value for our patients and our business.

Thesis work description:

Generative learningrooted in the idea that new knowledge builds on prior knowledgehas rapidly advanced de novo drug design. Reinforcement learning autoencoders and RNNs can generate novel molecules (SMILES) directly from data yet a systematic bridge from modest hit compounds to candidate drugs still lacks standardization. This thesis will:

Data curation:

• Internal datasets (established): Utilize retrospective AstraZeneca project data built on prior code and workflows.
• Public datasets (new): Extend and generalize the methodology to open datasets relevant to hit-to-lead and lead optimization enabling transparent benchmarking and reproducibility.

Generative modeling:

• Work with AstraZenecas proprietary generative AI model ReInvent for hit-to-candidate trajectories.
• Scoring functions:
• Phase 1 (completed): Idealized in-silico scoring to model the optimal scenario.
• Phase 2 (current): Introduce realistic experimentally derived criteria (e.g. lipophilicity solubility biological activity).

Learning strategies:

• Integrate active learning and reinforcement learning to refine structureactivity predictions and decision policies during optimization.
• Policy optimization
• Evaluate optimization algorithms in ReInvent to identify efficient oracle functions for candidate generation.

Publication:

• Prepare a manuscript detailing methodology results and implications with the goal of submission to a peerreviewed journal during or by the end of the placement.

Key Objectives:

• Model benchmarking: Assess the efficiency and accuracy of ReInvent and policy optimization under varying information constraints across proprietary and public datasets.
• Quantify informational content: Determine the minimum and optimal information needed to progress from hit to candidate drug.
• Impact assessment: Deliver actionable recommendations for generative AI use in drug discovery at AstraZeneca and to the wider community via publication.

Impact:
Your findings will directly inform ongoing and future projects using ReInvent accelerating candidate design with improved speed and efficiency. Publication will extend impact to the broader scientific and pharmaceutical communities.

References:

Brown DG. An Analysis of Successful Hit-to-Clinical Candidate Pairs. J Med Chem (2023).

He J et al. Molecular optimization capturing chemists intuition using deep neural networks. J Cheminform (2021).

Gummesson Svensson H et al. Utilizing reinforcement learning for de novo drug design. Mach Learn (2024).

Loeffler HH et al. Reinvent 4: Modern AIdriven generative molecule design. J Cheminform (2024).

Mervin L et al. Property-based scoring and optimizationdetails to be provided during onboarding.

Structure:

• Duration: Spring term 2026
• Credits: 30
• Two Students (apply together)

Essential Requirements:

• Enrolled in a Masters program in a relevant field.
• Programming in python
• Knowledge in Machine learning
• AI knowledge

So whats next

Apply today and take the chance to be part of making a difference making connections and gaining the tools and experience to open doors and fulfil your potential. We cant wait to hear from you!

We welcome your application as soon as possible but ahead of the scheduled closing date 2nd of November 2025. In the event that we identify suitable candidates ahead of the scheduled closing date we reserve the right to withdraw the vacancy earlier than published.

Date Posted

Closing Date