The fragrance industry has seen increased demand for ultimately biodegradable ingredients (e.g., &gt;60% mineralization), yet in silico QSAR models to identify candidates in an R&D pipeline have largely remained unchanged with an emphasis on predicting ready biodegradability (e.g., 60% mineralization in 28 days). From an R&D perspective, this approach restricts design as the ready biodegradability data that feed these models are notoriously prone to false negatives. This creates a prediction model that is overly stringent causing potential for “missed opportunities”. In this study, we report an in silico model to predict ultimate biodegradability of fragrance ingredients. The data for 876 chemicals were collected from IFF in-house and external databases and the chemicals were classified as ‘biodegradable’ if they met the criterion of 60% ThCO2/THOD production in 60 days in an OECD 301/310 screening test, a commonly accepted regulatory and industry threshold for ultimate biodegradability. The model was developed using Biovia Pipeline Pilot and Optibrium StarDrop software, trained and tested on two 80/20 dataset splits: one with random selection into the hold-out test set and another focusing on chemical diversity. Multiple combinations of physicochemical molecular descriptors, molecular fingerprints and machine learning algorithms were examined to find the best combination. The ensemble tree model (XGBoost), was identified as the top performing model with overall prediction accuracy of 0.75 – 0.83 and Cohen’s Kappa of 0.45 – 0.61 for randomly selected and structurally diverse independent test sets. These results are comparable to previously reported QSAR models that were trained to predict ready biodegradability. The advantage of the new model is that it expands the chemical space and allows for more data to be utilized. Most of the model’s mispredictions can be traced to low support in the training set, potentially exacerbated by improper classification of ingredient biodegradability due to limited experimental trials.

Practical machine learning model for predicting biodegradability of fragrance materials

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.