Sterols are lipids produced by eukaryotic organisms and are widely used as biomarkers for reconstructing past phytoplankton communities. However, individual sterols often show poor taxonomic specificity because of overlapping biological sources. To overcome this limitation, this study employed machine learning to evaluate whether sterol assemblages can predict phytoplankton communities better.
Overall, we obtained the relative compositions of 15 sterols and the corresponding 18S rDNA-derived phytoplankton community data from 261 modern suspended particulate matter samples collected from the East China Sea shelf, Qiongdong coastal waters, the South China Sea basin, and the western tropical Pacific. Based on the sequencing results, seven phytoplankton groups were initially evaluated as potential response variables: Bacillariophyta, Dinoflagellata, Chlorophyta, Haptophyta, Cryptophyta, Ochrophyta, and Rhodophyta. Among these groups only Bacillariophyta, Dinoflagellata, and Chlorophyta were abundant enough to yield higher test-set coefficient of determination (R²) values, and thereby used for model optimization and sediment-core reconstruction. The dataset was randomly divided into training and test sets at a ratio of 9:1, and principal component analysis showed that the test samples were generally covered by the training-set sterol composition space. Seven multi-output regression models were compared, covering tree-based ensemble, support vector, regularized linear, partial least squares, and neural network methods. Hyperparameters were optimized using grid or randomized search with five-fold cross-validation. Model performance was assessed using R², root mean square error (RMSE), and mean absolute error (MAE). Our results show extreme gradient boosting regression (XGBR) and random forest regression (RFR) performed best. The optimized models were applied to the NH07 sediment core from the northern South China Sea. Based on conventional individual sterol indicators, the community structure of phytoplankton did not show an expected trend responding adequately to millennial marine and climate changes since the last glaciation. In contrast, machine learning reconstructed reasonably the relative variations of Bacillariophyta, Dinoflagellata, and Chlorophyta, providing a more detailed view of community changes since the last deglaciation under the influence of Kuroshio intrusion, upper water-column stratification, and nutrient availability. This study provides a novel method to fully elucidate the ecological information contained in phytoplankton sterol profiles and to accurately reconstruct paleo community structure change in marine environments.