Hubert Misztela is a director of data science at Novartis working on generative molecular design, scientific reasoning and enterprise AI. He has helped create influential research tools for drug discovery while developing approaches that combine language models, specialized software and human expertise.
From cryptography to computational chemistry
Misztela began in computer science and cryptography, coauthoring research on private information retrieval using trusted hardware. His later research brought machine learning into pharmaceutical discovery, where he has led AI researchers working on drug design.
In 2021, he coauthored FS-Mol, a molecular dataset and benchmark for few-shot learning in drug discovery, designed around a central pharmaceutical constraint: experimental measurements for individual biological targets are often scarce. His subsequent work included molecular-graph reconstruction with variational autoencoders and PREFER, a predictive-modeling framework for molecular discovery.
He also coauthored Chimera, a retrosynthesis-prediction framework that combines chemically complementary models through learned ensembling. The project addresses a practical limitation of computational drug design: a promising molecule is valuable only when researchers can identify plausible ways to make it.
- Scientific reasoning around retrieval. Misztela argues that conventional retrieval-augmented generation struggles when answers depend on concepts scattered across different disciplines. His approach applies reasoning to both questions and evidence before retrieval, then uses structured, causal or algorithmic methods to interpret the results. As he describes in his own writing, specialized tools can handle particular forms of inference more reliably than a language model operating alone.
- Retrospective discovery evaluation. Working with medicinal chemist Derek Lowe, Misztela explored whether AI could reconstruct insights underlying RNA interference from scientific literature published before the discovery. The evaluation framework tests whether systems can recover overlooked relationships, generate grounded hypotheses and approach mechanistic explanations without relying on later knowledge.
- Workflow-aware enterprise agents. Misztela maintains that useful organizational agents require planning, tools, persistent memory and detailed operational context. His enterprise-agent framework maps practical employee roles—including connectors, multipliers and knowledge hubs—and progresses from individual assistants to team agents and collaborative multi-agent systems. He emphasizes employee participation, human judgment and ethical accountability as autonomy increases.
