Microbial and mammalian cell cultures are widely used as bioreactors in medicine, chemistry, food, and energy industries (Singh et al., 2016). In addition to the genetic engineering of recombinant cells, the development of culture media is essential to improve the productivity of these bioreactors (Packiam et al., 2020; Rodrigues et al., 2014; Zhou et al., 2018). Adjusting the medium compositions, i.e., carbon sources, amino acids, fatty acids, vitamins, trace elements, salts, and growth factors (Bonnet et al., 2020; Price, 2017; Ritacco et al., 2018; van der Valk et al., 2010), are often required to achieve the cultivation purposes, such as cell growth, metabolite productivity, and cost (Fernandes et al., 2020; O'Neill et al., 2021). As a classical methodology of media optimization, the One-Factor-at-Time (OFAT) method is mainly used in laboratories (Combe and Sokolenko, 2021). As this method adjusts the medium components one by one without considering the interactions among the components, it benefits the simple experimental operation but potentially misses the best chemical combinations (Yao and Asayama, 2017). Considering the components' interactions, the response surface method (RSM) has been developed, which uses a quadratic polynomial approximation to represent the relationship between the media and cells (Parampalli et al., 2007). Although RSM is a theoretical methodology, it remains too simple to interpret highly complex systems of media and cells (Singh et al., 2009).

Recently, machine learning (ML)-assisted medium optimization has been proposed, of which the typical workflow includes processing input data, training prediction models, and predicting new data (Camacho et al., 2018). ML is particularly suitable for analyzing large and complex data without knowing the data structure details (Greener et al., 2022), so it can be employed to find the best medium composition for cell culture, regardless of the contributions of medium components to cell culture remaining as a black box. Intensive studies suggested that the medium optimization by ML outperformed RSM and asserted the superiority of RSM over ML (Bankar et al., 2014; Desai et al., 2008; Guo et al., 2010; Nagata and Chu, 2003; Nelofer et al., 2012; Pal et al., 2009; Sampaio et al., 2016; Zafar et al., 2012). For example, the yield of actinomycin V in Streptomyces triostinicus and the lipase activity in Escherichia coli were 36.7% and 11.2% higher in the media developed using ML than those using RSM, respectively (Nelofer et al., 2012; Singh et al., 2009). Despite the advantages of using ML in medium optimization, the applications of ML to medium development and cell culture optimization remain significantly less than those of OFAT and RSM. It might be due to a lack of comprehensive guidelines and manuals on ML-assisted medium optimization. The present review highlights the representative studies in the field and provides an extensive overview of the ML-assisted culture medium development.

We first summarize 41 studies regarding the medium development using ML and the techniques used in the representative studies, such as maximizing productivity and cell concentration. Additionally, the advantages and disadvantages of ML-assisted medium optimization and the perspectives of the methodology are discussed. Finally, guidance on how to choose the medium optimization methods appropriate for the objective is proposed. Note that the definition of medium optimization here refers to finetuning the abundance (i.e., concentration) of the components (e.g., chemical compounds) in the media for improved cell culture.