*In silico* modeling of transcriptional responses to perturbations is crucial for advancing our understanding of cellular processes and disease mechanisms. We present PertEval-scFM, a standardized framework designed to evaluate models for perturbation effect prediction. We apply PertEval-scFM to benchmark zero-shot single-cell foundation model (scFM) embeddings against baseline models to assess whether these contextualized representations enhance perturbation effect prediction. Our results show that scFM embeddings offer limited improvement over simple baseline models in the zero-shot setting, particularly under distribution shift. Overall, this study provides a systematic evaluation of zero-shot scFM embeddings for perturbation effect prediction, highlighting the challenges of this task and the limitations of current-generation scFMs. Our findings underscore the need for specialized models and high-quality datasets that capture a broader range of cellular states. Source code and documentation can be found at: https://github.com/aaronwtr/PertEval.

MOTIVATION: Anti-cancer drug response prediction (DRP) using cancer cell lines (CLs) is crucial in stratified medicine and drug discovery. Recently, new deep learning models for DRP have improved performance over their predecessors. However, different models use different input data types and architectures making it hard to find the source of these improvements. Here we consider published DRP models that report state-of-the-art performance predicting continuous response values. These models take chemical structures of drugs and omics profiles of CLs as input. RESULTS: By experimenting with these models and comparing with our simple baselines, we show that no performance comes from drug features, instead, performance is due to the transcriptomics CL profiles. Furthermore, we show that, depending on the testing type, much of the current reported performance is a property of the training target values. We address these limitations by creating BinaryET and BinaryCB that predict binary drug response values, guided by the hypothesis that this reduces the noise in the drug efficacy data. Thus, better aligning them with biochemistry that can be learnt from the input data. BinaryCB leverages a chemical foundation model, while BinaryET is trained from scratch using a transformer-type architecture. We show that these models learn useful chemical drug features, which is the first time this has been demonstrated for multiple testing types to our knowledge. We further show binarizing the drug response values causes the models to learn useful chemical drug features. We also show that BinaryET improves performance over BinaryCB, and the published models that report state-of-the-art performance. AVAILABILITY AND IMPLEMENTATION: Code is available from https://github.com/Nik-BB/Understanding_DRP_models.