Complex AI Models Fail to Outshine Simple Methods in Predicting Gene Expression Changes
Summary
Despite their sophistication, state-of-the-art deep learning models fail to surpass simple baseline methods in accurately predicting gene expression changes caused by genetic perturbations, with higher errors observed for complex models, particularly in cases of double gene perturbations.
Key Points
- Deep learning foundation models do not outperform simple baseline methods in predicting transcriptomic effects of genetic perturbations
- Prediction errors are often higher for complex models, especially with double gene perturbations
- Increased model complexity and computational cost do not guarantee better performance in this context