Addressing the data scarcity problem in ecotoxicology via small data machine learning methods

Abstract

Machine learning (ML) has been extensively employed in the field of ecotoxicological research, with a notable application in predicting ecotoxicity. Typically, researchers employ chemical substance property data, environmental exposure data, and species information to predict the potential toxicity of pollutants through specific ML models and algorithms. (1) However, it is difficult to gather systematic collected toxicological data as they are costly and time-consuming to generate, while endless combinations of pollutant species with exposure media properties can be made. This leads to unbalanced experimental data, which often results in data sets with varying quality that are unsuitable for modeling. In fact, many modeling studies in this field lack a solid database. (2) In the process of applying ML for toxicity prediction, limited data and sparse learnable features are common challenges. When the available data cannot meet the requirements for model construction, this situation is termed a “small data” problem. This issue can be understood from two perspectives. First, the sample size fails to meet the minimum data requirements for ML modeling; for instance, previous studies suggested such a minimum threshold as 100 samples. (3,4) Second, the sample size fails to satisfy the specific needs of the modeling task. For example, when the ratio of the number of samples to the number of features is too low, the model is prone to overfitting. In such cases, even thousands of samples may still be considered small data for models with hundreds of features. In our previous literature review, most studies that achieve favorable predictions maintain a sample data-to-features ratio above 20. (5,6) Hence, we suggest that a ratio below 20 should be considered a sign of a small data scenario. Such limited data sets not only curtail the predictive performance of ML models but also diminish the representativeness and persuasiveness of the models.

Acknowledging the difficulty of the acquisition of reliable toxicity data, there is an urgent need to develop a specific ML-based modeling process for small data samples, with the aim of improving data utilization and improving the predictive performance of models. A rapid and large-scale prediction of the ecotoxicity of chemicals and/or nanomaterials is crucial for reducing the time and cost of risk assessment, as well as minimizing animal experiments and selecting optimal testing strategies. This is also considered as a key direction in 21st century toxicology. Therefore, we believe that small data ML (SDML) methods are especially valuable when the available research data set is inadequate for applying conventional ML models. The outcomes of SDML models can also identify data gaps and help refine the plan for conducting supplementary ecotoxicity tests, thereby optimizing the allocation of limited resources.

In this Viewpoint, we discuss the implementation principles, modeling processes, and challenges of SDML methods and elucidate the application of SDML methods in ecotoxicology for advancing science and technology.