Benchmarking biomedical foundation models

A transparent evaluation and proof of reproducibility, generalization and replicability of algorithms are the bedrock of method development in computational biology. Many benchmarking efforts have been developed for problems ranging from structural biology to translational biomedicine. Rigor is relatively controllable for tasks such as the prediction of patient outcomes or the outcomes of biological assays, but the problem is exacerbated when the aim is to benchmark foundation models. The parameters constituting them are supposed to capture the patterns underlying the data; therefore, the models are parameterized embodiments of the phenomena that gave rise to the data. How can we test the limitations of these models? Here, we discuss the epistemological value of foundation models; whether they can be refuted, verified or evaluated primarily on the basis of utility; what principles should guide their benchmarking; and what role the scientific community should play in that benchmarking process.