Problem
A pharma team designs a therapeutic protein: a symmetric homodimer of an Fc fragment with an agonist of a human receptor. More than 1000 molecules were proposed in silico and 10 went to the first round of experimental production, which takes weeks and costs money per data point. While the wet lab worked, the team had to rank the 10 candidates by an in silico risk. They chose deamidation, where an asparagine (N) turns into aspartate (D) spontaneously and irreversibly, which hurts efficacy, safety and stability. The data are small, imbalanced and noisy, and the industry does not share data because of patents.
Example from the article: the 10 candidates plus one control contained 128 asparagines, and the model flagged 41 of them for deamidation.
Solution in brief
Reproduce the published Random Forest on structural features of each asparagine, try CatBoost against it, then check the predictions in LC-MS/MS. Reproduction matched the paper, and CatBoost gave a gain of about 2% in accuracy. The predictions did not separate the candidates, and the lead molecule was picked on other parameters.
Questions this case answers
- How to rank a few candidate objects by the predicted rate of an irreversible site-level change, when site-level labels are few and strongly imbalanced
- How to decide whether a small accuracy gain over a published baseline changes the choice among candidates that look alike
- How to predict post-translational modifications (deamidation) of therapeutic protein candidates
- How to reproduce a published Random Forest result and check whether CatBoost beats it on a small imbalanced dataset
Tools and methods
- Random Forest — baseline from Delmar et al. (2019), reproduced
- CatBoost — default and optimized runs on the same features
- LC-MS/MS — experimental check of deamidation at one time point
Google or a one-time email code. No password.
