Effect of missing data on multitask prediction methods.

Area of Science:

• Chemoinformatics
• Computational Chemistry
• Machine Learning

Background:

• Multitask prediction is crucial in chemoinformatics for profiling compound activities.
• Deep neural networks (DNNs) are increasingly used for multitask prediction.
• Multitarget datasets are often sparse, posing challenges for model training.

Purpose of the Study:

• To investigate the impact of data sparsity on multitask prediction performance.
• To compare the effects of missing data on DNNs and Bayesian probabilistic matrix factorization (Macau).

Main Methods:

• Simulated data sparsity by removing data from complete datasets.
• Trained DNNs and Macau models on these sparse datasets.
• Compared model performance across varying levels of data removal.

Main Results:

• Both DNNs and Macau showed similar performance degradation with increasing data sparsity.
• Performance decrease was gradual initially, then accelerated significantly after substantial data removal.
• Identified a critical threshold for data requirements in multitask prediction.

Conclusions:

• Data sparsity significantly affects multitask prediction model performance.
• Understanding data requirements is essential for reliable biological activity profiling.
• This study offers insights into data sufficiency for effective chemoinformatics models.