Representing descriptors derived from multiple conformations as uncertain features for machine learning.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Chemical descriptors often lack 3D information, limiting predictive model accuracy.
• Machine learning methods like random forests are powerful tools for chemical data analysis.

Purpose of the Study:

• To investigate the impact of incorporating 3D conformational uncertainty into chemical descriptors on random forest predictive performance.
• To evaluate different strategies for handling uncertainty within random forest models.

Main Methods:

• Conformational analysis was used to introduce uncertainty into chemical descriptors across 11 datasets.
• Random forests were employed for binary classification tasks.
• Various strategies for handling uncertainty, including uniform and normal distributions, were assessed.

Main Results:

• Uniform probability distributions with fractional compound distribution outperformed normal distributions and sampling methods when incorporating 3D uncertainty.
• Random forest models using only expected values from uncertain distributions achieved nearly equivalent accuracy to models using the full distributions.
• Similar model performance was achieved using 3D descriptor information from single conformations.

Conclusions:

• The inclusion of 3D information is beneficial, but complex uncertainty handling in random forests offers minimal advantage over using single-point descriptor values.
• Simpler methods of incorporating 3D chemical descriptor information are often sufficient for achieving high predictive performance with random forests.