Szalay, K. Z. | bioRxiv Science

bioRxiv2024

Benchmarking a foundational cell model for post-perturbation RNAseq prediction

Accurately predicting cellular responses to perturbations is essential for understanding cell behaviour in both healthy and diseased states. While perturbation data is ideal for building such predictive models, it is considerably sparser than baseline (non-perturbed) cellular data. To address this limitation, several foundational cell models have been developed using large-scale single-cell gene expression data. These models are fine-tuned after pre-training for specific tasks, such as predicting post-perturbation gene expression profiles, and are considered state-of-the-art for these problems. However, proper benchmarking of these models remains an unsolved challenge. In this study, we benchmarked a recently published foundational model, scGPT, against baseline models. Surprisingly, we found that even the simplest baseline model - taking the mean of training examples - outperformed scGPT. Furthermore, machine learning models that incorporate biologically meaningful features outperformed scGPT by a large margin. Additionally, we identified that the current Perturb-Seq benchmark datasets exhibit low perturbation-specific variance, making them suboptimal for evaluating such models. Our results highlight important limitations in current benchmarking approaches and provide insights into more effectively evaluating post-perturbation gene expression prediction models.

systems biology↗

bioRxiv2023

The EFFECT benchmark suite: measuring cancer sensitivity prediction performance - without the bias

1.Creating computational biology models applicable to industry is much more difficult than it appears. There is a major gap between a model that looks good on paper and a model that performs well in the drug discovery process. We are trying to shrink this gap by introducing the Evaluation Framework For predicting Efficiency of Cancer Treatment (EFFECT) benchmark suite based on the DepMap and GDSC data sets to facilitate the creation of well-applicable machine learning models capable of predicting gene essentiality and/or drug sensitivity on in vitro cancer cell lines. We show that standard evaluation metrics like Pearson correlation are misleading due to inherent biases in the data. Thus, to assess the performance of models properly, we propose the use of cell line/perturbation exclusive data splits, perturbation-wise evaluation, and the application of our Bias Detector framework, which can identify model predictions not explicable by data bias alone. Testing the EFFECT suite on a few popular machine learning (ML) models showed that while library-standard non-linear models have measurable performance in splits representing precision medicine and target identification tasks, the actual corrected correlations are rather low, showing that even simple knock-out (KO)/drug sensitivity prediction is a yet unsolved task. For this reason, we aim our proposed framework to be a unified test and evaluation pipeline for ML models predicting cancer sensitivity data, facilitating unbiased benchmarking to support teams to improve on the state of the art.

systems biology↗