A Deep Learning Approach for Single-Cell Perturbation Prediction Using Small Molecule Chemical Structures

In this study, we develop a deep learning framework aimed at predicting the impacts of chemical perturbations on individual cells, emphasizing the encoding of small molecular chemical structures . Utilizing the LINCS L1000 dataset, the approach incorporates transfer learning with the ChemBERTa model to navigate the challenges of high-dimensional single-cell data analysis and sparse dataset limitations. The framework integrates computational models such as Transformer, DenseNet, and CNN, designed to understand cellular responses to perturbations at the single-cell level. Validated through 5-fold cross-validation, the ensemble model combines the strengths of individual models to improve prediction accuracy and robustness for cellular responses to perturbations. This study proposes a novel encoding scheme for small molecule chemical structures and cell types, integrating various computational models to contribute to the development of predictive models for small molecule perturbations. It offers a step towards enhancing the understanding of complex biological responses through deep learning methodologies.