Learning curves for drug response prediction in cancer cell lines

Abstract Background Motivated by the size and availability of cell line drug sensitivity data, researchers have been developing machine learning (ML) models for predicting response to advance cancer treatment. As studies continue generating a common question is whether generalization performance existing prediction can be further improved with more training data. Methods We utilize empirical curves evaluating comparing data scaling properties two neural networks (NNs) gradient boosting decision tree (GBDT) trained on four screening datasets. The are accurately fitted power law model, providing framework assessing behavior these models. Results demonstrate that no single model dominates in terms across all datasets sizes, thus suggesting actual shape depends unique pair an ML dataset. multi-input NN (mNN), which gene expressions cells molecular descriptors input into separate subnetworks, outperforms single-input (sNN), where features concatenated layer. In contrast, GBDT hyperparameter tuning exhibits superior as compared both NNs at lower range set sizes tested datasets, whereas mNN consistently performs better higher sizes. Moreover, trajectory suggests increasing sample expected improve scores NNs. These observations benefit using evaluate models, broader perspective overall characteristics. Conclusions A curve provides forward-looking metric analyzing serve co-design tool guide experimental biologists computational scientists design future experiments prospective research studies.