Machine Learning for Small, Heterogeneous, and Incomplete Chemical Data

In many scientific domains, data are often small, heterogeneous, sparsely distributed and sometimes incomplete. Such incompleteness can arise for several reasons, including instrument errors or missing measurements during experiments. Extracting meaningful and knowledgeful patterns from these datasets can open doors for new scientific discoveries while saving both time and cost, because many experimental procedures are very time-consuming and expensive. This is one reason why data-driven machine learning (ML) has become increasingly important across scientific domains, and the chemical sciences are no exception [1]. We were motivated by a similar challenge while working with polymer chemical data, which are useful for understanding materials relevant to drug delivery and other biomedical applications. However, such data come with several challenges: • First, there is a methodological challenge: how should raw and heterogeneous data be curated and transformed into a machine-readable form without distorting the underlying scientific structure? • Second, there is an algorithmic challenge: which class of learning models is most suitable for prediction and data completion under small-sample and mixed-type conditions? • Third, there is an interpretability challenge: how can one explain which input factors are driving the outputs, particularly when the learned relationships may be nonlinear? To address these questions, we evaluate a diverse set of regression models spanning different learning paradigms, including regularized linear methods, instance-based learning, tree ensembles, gradient boosting, and neural networks. These models are applied to prediction tasks on our curated polymer dataset. Since explainability is particularly important in scientific applications, where understanding why a model produces a prediction is often as important as the prediction itself, we further incorporate model-based explainability to identify the contribution of individual features to the predicted outputs. Although several ML studies [2,3] have addressed prediction, explainability, or both, learning from small, heterogeneous, sparse, and multi-output scientific datasets remains challenging. To deal with this problem, we develop a unified analytical framework that combines prediction, missing-data completion, and explainability within the same workflow. Recent studies [4,5] on small polymer datasets have mainly relied on artificial neural network (ANN) models for prediction. This raised an important question for us: whether such complex models are always appropriate for small and heterogeneous datasets. We therefore re-evaluate these settings using comparatively simpler baseline ML models. Our results indicate that conventional ML models can perform competitively, and in several cases better than ANN-based approaches, under limited-data conditions. References [1] Chew, J. W., Andersson, R., Bierweiler, T., Ryttestal, P., & Wik, T. (2026). AI in chemical engineering: From promise to practice. AIChE Journal, 72(7), e70358. [2] Borchani, H., Varando, G., Bielza, C., Larranaga, P., 2015. A survey on multi-output regression. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 5, 216–233. [3] Emmanuel, T., Maupong, T., Mpoeleng, D., Semong, T., Mphago, B., Tabona, O., 2021. A survey on missing data in machine learning. Journal of Big data 8, 140. [4] Shalaby, K.S., Soliman, M.E., Casettari, L., Bonacucina, G., Cespi, M.,Palmieri, G.F., Sammour, O.A., El Shamy, A.A., 2014. Determination of factors controlling the particle size and entrapment efficiency of noscapine in peg/pla nanoparticles using artificial neural networks. International journal of nanomedicine, 4953–4964. [5] Sarajlić, D., Abdel-Ilah, L., Fojnica, A., Osmanović, A., 2017. Prediction of the size of nanoparticles and microspore surface area using artificial neural network. Genetics & Applications 1, 65–70.