Branco Weiss Fellow Since
2025
Research Category
Computer science, statistics, bioinformatics
Research Location
Department of Mathematics, ETH Zürich, Switzerland
Background
Scientific questions often revolve around uncovering the drivers and mechanisms behind observed phenomena and thus tend to be causal in nature. Yet, despite the scientific mantra that “correlation does not imply causation”, most machine learning (ML) and AI tools used to gain insights from observational or experimental data rely solely on pattern recognition and are therefore purely associational. The field of causal inference studies how to go beyond correlation and is concerned with developing formal statistical methods to extract causal structure from data, quantify the strength of cause-effect relationships, and predict the outcomes of new interventions.
Traditionally, these methods assume that the variables of interest (e.g., a treatment and health outcome) are directly observed. However, modern data sources instead often contain thousands of low-level proxy measurements (e.g., pixels, words, EEG signals, RNA or protein counts), while the abstract high-level quantities among which causal relations are thought to operate (e.g., objects, topics, disorders, gene programs) remain hidden in the raw observations. This limits the applicability of traditional causal inference methods to complex and unstructured data.
Details of Research
Dr. Julius von Kügelgen’s research falls within the nascent subfield of causal representation learning, which integrates causal inference with machine learning tools to extract high-level features from low-level data. Specifically, Dr. von Kügelgen aims to develop principled new methods for gaining causal insights from high-dimensional and heterogeneous data, with a particular attention to approaches informed by practical needs and applications in other scientific disciplines.
Due to the increasing availability of large-scale experimental omics data, the field of bioinformatics offers unique opportunities for such applications. At the same time, many problems in biology (and beyond) involve measurements of tens or hundreds of thousands of variables, necessitating the envisioned methodological advances. Problems Dr. von Kügelgen aims to address include, for example, inferring gene pathways and their interactions from RNA transcription data or predicting the effects of unseen combinations of drugs or gene perturbations.
