Cell culture underpins biomanufacturing across sectors ranging from biologics and advanced therapies to sustainable food and materials. However, the ability to translate scientific breakthroughs into commercially viable products can be constrained by one fundamental component: the culture medium. Cells rely on culture media to provide the right environment to optimize cell growth, health, productivity, and product quality, but no single formulation is optimal for all applications. This is particularly relevant as biomanufacturers adopt more diverse cell types and develop increasingly complex products, such as antibodies, and cell and gene therapies. Ultimately, the right formulation has the potential to improve the reproducibility, scalability, and process control that commercial manufacturing of cell products demands.

Despite this clear association, many companies continue to rely on generic, one-size-fits-all formulations that are not designed for their specific cells, products, or manufacturing objectives. This can potentially lead to inconsistent results, batch-to-batch differences, and unpredictable process performance. Further variability comes from undefined or animal-derived components – such as fetal bovine serum (FBS) and human platelet lysate – which can also introduce safety, regulatory, and supply chain concerns, leaving manufacturers vulnerable to shortages and price fluctuations. These problems are especially significant in commercial manufacturing, where suboptimal media can reduce yield and product quality, complicate scale-up, increase production costs, and make processes harder to control.

‍

Addressing cell behavior from a data science perspective

Customizing media could address many of these limitations, but conventional development methods are resource intensive, expensive to apply routinely, and prohibitively slow, typically taking from two to four years. For many companies, this delay is not commercially viable, forcing them to buy generic off-the-shelf formulations that are perceived as ‘good enough’ but are not best suited to their cell lines or aims. At the heart of the challenge is the fact that the effects of changing an enormous number of possible ingredients, concentrations, and interactions are exceptionally difficult to predict. Metabolic models provide valuable insights but cannot capture every interaction, making media optimization as much a data science issue as a biological one

Changing the focus in this way acknowledges that it is not always necessary to understand every biological mechanism before making a change to a formulation. Instead, developers should take an empirical approach, systematically mapping the relationships between media inputs and measurable cellular outputs. However, the number of potential permutations of a formulation containing, for example, 60 ingredients is immense.

‍

How AI accelerates media optimization

AI contributes to data science-based media optimization by applying an engineering cycle of design, build, test, and learn to a highly complex biological problem. Large language models help to analyze scientific literature and identify ingredients and variables that may be relevant to a particular cell line. Other tools detect patterns in existing knowledge, explore complex patent filings, and translate intellectual property restrictions into practical rules, helping to ensure freedom to operate from the start. These insights are then combined with human expertise to define the initial formulation space and the desired objective, whether that’s increasing cell growth or product titer, improving product quality, reducing cost, or all three.

AI also expands the range of measurable parameters that can inform the process. For example, AI-powered image analysis can convert subjective assessments of cell morphology into quantitative metrics, which, combined with measurements such as growth, viability and monoclonal antibody titer, provide a comprehensive assessment of each formulation.

Crucially, it is relatively quick and simple for an AI algorithm to assess the performance of multiple media formulations. An active learning AI approach significantly shortens the time required for development by observing how cells respond to different formulations and using those results to guide the next experiments. Data is analyzed using Bayesian optimization, which identifies promising experiments and those with high levels of uncertainty, helping scientists to quickly reduce the number of formulations to take forward to the next round.

Unlike conventional machine learning – which analyzes existing datasets – active learning continuously updates as new data are generated, selecting the next experiments to maximize learning. This feedback loop makes better use of time, materials, and laboratory capacity, reducing media optimization from around two years to approximately six months.

‍

AI depends on generating high quality experimental data

The key to AI-driven optimization is reliable experimental data that identifies meaningful patterns, together with a modeling process that can learn efficiently from every result. Results can be combined with scientific expertise and customer knowledge to improve ingredient selection and provide a stronger starting point. Further high quality experimental data can then guide optimization within that reduced search space, making it possible to efficiently tailor formulations. Realistically, only a minute fraction of the possible combinations can be tested within commercially viable timelines, so each experiment must be selected carefully. Inaccurate or inconsistent measurements risk sending optimization in the wrong direction, wasting precious time and resources.

‍

Changing attitudes towards AI in biological engineering

The development of this approach has coincided with a change in industry attitudes towards AI, as well as a growth in data volumes that is making traditional media development harder to manage. Scientists and manufacturers now recognize the potential of active learning and Bayesian optimization in this space, but adoption remains limited by the lack of high quality data and specialized AI models. Effective media optimization requires several AI tools that complement robotics, analytics, and human expertise to give manufacturers verified outcomes, including faster development and the right balance between cost, performance, and quality.

Success in media development no longer depends on running more experiments, but on generating better data and learning more quickly. Combining engineering principles, automation, AI, and large-scale data generation is creating a fundamentally new approach to media development, augmenting scientific expertise rather than replacing it. Organizations that use these technologies to build richer datasets and faster learning cycles will define the future of media design, helping industries from food to pharma to achieve breakthroughs sooner.

‍

Talk to one of our scientists about media designed for the cell types you're working with.