Extraordinary Biomed Seminar: From miRNA Binding to Biological Function: Benchmarks and Autonomous Machine Learning

Predicting molecular interactions does not necessarily mean predicting their biological consequences. MicroRNA target prediction illustrates this challenge particularly well: identifying potential binding sites is fundamentally different from predicting the effects of microRNA regulation on gene expression. In this seminar, I will present our work on developing more rigorous benchmarks for microRNA–target interactions, including miRBench, and predictive models such as miRBind2. I will discuss challenges in dataset construction, negative sampling, model generalization, and evaluation. I will also introduce ongoing work extending these benchmarks towards functional prediction using transcriptomic measurements from microRNA perturbation experiments.

Building on these challenges, I will present Agentomics, an autonomous system based on large language model agents that develops machine learning pipelines for various biological tasks. I will describe its architecture, experimental evaluation, and the role of iterative feedback in improving model development. Finally, I will discuss how robust benchmarking and autonomous machine learning can be brought together to support increasingly automated biomedical research, while maintaining reproducibility, meaningful evaluation, and scientific oversight.