3 questions for...
Dr. Lea Eckhart
Group member
I earned my Ph.D. in 2025 from Saarland University on the topic of trustworthy machine learning methods for personalized cancer treatment. Since then, I have been working as a postdoc at the “AI and Multi-Omics Data” Chair, which is part of the Institute for Medical Bioinformatics at the University Medical Center Göttingen. My research continues to focus on the development of trustworthy machine learning models for medical applications. In order for us to trust the predictions of artificial intelligence in research, drug development, and patient treatment, criteria such as reliability, interpretability, and robustness play a central role alongside predictive accuracy. Unfortunately, these factors are often neglected or pitted against one another during model development. Therefore, I am developing new methods that treat “trustworthiness” as a multifaceted concept and fulfill as many of the above-mentioned criteria as possible. In addition to oncological questions, I also focus on predicting antibiotic resistance based on routine clinical data from bacterial and fungal infections.
1.
They are researching how AI can predict a patient's response to cancer drugs. What opportunities does this approach open up for more personalized cancer treatment?
Our models learn to predict, based on multi-omics data from cancer cells (e.g., gene expression, DNA mutations), whether a drug is suitable for treatment. In addition to making predictions, they can provide insights into which characteristics of cancer cells and drugs positively or negatively influence efficacy. This knowledge allows us to define increasingly specific cancer subtypes with corresponding treatment strategies. Furthermore, our models can help in the targeted development of drugs with improved efficacy for specific subtypes. A ultimate goal is the development of clinically applicable models that provide personalized treatment recommendations based on tumor and patient data, with the aim of making therapy more efficient and better tolerated.
2.
How can researchers ensure that an AI system's recommendations remain transparent and understandable to doctors?
Comprehensible AI systems can only be developed through close multidisciplinary collaboration among researchers, physicians, and others. As researchers, we must—among other things—ensure that our AI models provide clear justifications for their predictions. It must be clear whether a prediction is based solely on clinically recognized biomarkers or whether other (complex) factors play a role. Furthermore, it must be clear which treatment-relevant information is not taken into account by the model. Cases in which the AI cannot provide a clear prediction must be clearly flagged. To ensure competence in working with AI systems, physicians must receive targeted training and continuing education.
3.
What opportunities does the combination of bioinformatics and artificial intelligence offer for gaining new insights into complex diseases?
Conventional bioinformatics develops problem-specific algorithms that operate according to predefined rules and often draw on biological background knowledge to solve problems. In contrast, machine learning can automatically learn relationships from a set of example data points—in principle, without any prior knowledge. In the case of complex diseases such as cancer, many mechanisms—for example, those related to disease onset or drug efficacy—are not yet understood, as they are often dictated by a complex interplay of cellular processes. By combining conventional bioinformatics and artificial intelligence, we can ensure that models respect known biological realities, which often also improves model interpretability. At the same time, the model can discover new relationships that are not covered by the current state of knowledge.
