Powerhouses and Foodstuff

Submitted by Powerhouse Biology.
Originally published in Life Sciences Insights Magazine, August 2025

There are few things that solidify a reputation better than a sticky nickname. In 1957, Philip Siekevitz observed that the mitochondrion is “a small body which appears to play a central role in the oxidation of foodstuff,” and as such, he dubbed the mitochondrion the powerhouse of the cell. In the 68 years since that report, we have learned that these small bodies are more than just foodstuff oxidizers. They are also highly dynamic organelles that dictate cellular fates downstream of a wide array of nutrient, energy, or stress signals. Mitochondria are essential, but mitochondrial efficiency and regulation are especially critical in tissues with high metabolic demands such as muscle, heart, and brain. As we age, those energy-demanding tissues accumulate decades of small, unresolved stresses that can tip our mitochondria towards dysfunction. While we can’t go back in time, we believe we can use AI to mitigate the effects of aging and age-related pathologies by directly targeting mitochondrial dysfunction.

Mitochondria continuously react to their environments by rapidly modifying how they look, where they go, when they get there, and how much energy they need to get the job done (sidenote: this feels like an analogy to parenthood). With all of these critical functions, it isn’t surprising that a little dysregulation or lapse in maintenance can devolve into pathological conditions, especially when amplified over time. It is also easy to imagine the scale of dynamic cellular responses that correlate with these changes, and in that realization, we have to acknowledge that the dimensionality of traditional cell biological assays is insufficient to resolve the mitochondrial behaviors that drive age-related diseases.

For example, single endpoint assays, such as those that measure viability or cell death, are designed to assess acute, non-sustainable cellular states. To assess biological phenomena such as prolonged mitochondrial dysfunction, we instead need assays with high-dimensional readouts that are representative of persistent states of cellular health. We have therefore built our mito-centric platform around a live fluorescence imaging assay that captures cellular compartments and features, and coupled that to temporally linked omics data. Importantly, these data types are target-agnostic, so we can use machine learning to train multi-modal fingerprints of mitochondrial behavior without biases towards specific pathways or functions.

Much like the crime drama CSI (only without the drama), we will match these AI-generated mitochondrial fingerprints to generalized profiles from a large cohort of patient sample data. This will allow us to identify in vitro cellular states associated with mitochondrial dysfunction that closely resemble human diseases. Using the experimental conditions that are most representative of real world human health conditions, we will then test a series of in vitro treatments that mimic real world human drug responses.

In CSI, this looks something like feeding data into a computer, saying “enhance” a few times, and getting our answers. In reality, however, training AI models that generate accurate outputs (and not word salad) hinges on our ability to (1) generate data that are optimized for AI model training and representative of human physiology, and (2) analyze those data using AI while adhering to the scientific method.

In classical cell-based assays, we can designate positive and negative controls at either end of a spectrum and plot test conditions along that spectrum to look for drug-dependent shifts. This becomes more complicated as we add layers (or dimensions) of data, and so to generate AI-ready data, we must be thoughtful in how we design our experiments and analytical workflows. To this effort, there should be control conditions and quality metrics implemented that can be used to identify batch effects, experimental variability, or potential confounding factors.

These are critical for understanding whether downstream model performance is driven by biologically relevant effects rather than plate, date, or donor dependent ones. And while we need to minimize experimental variability, we should simultaneously maximize human variability and generate a sufficient amount of high quality data to encourage models to differentiate conditions by generalizable and biologically relevant features. Thus, biological controls that allow us to implement normalization strategies across experiments, plates, and donors should be included in experiments so that trained models more accurately represent the underlying biology.

Bearing the above in mind, it is important to remember that not all data are suited for model training, and not all of their subsequent outputs tell us what we think they do. For example, models trained to identify healthy or diseased states can get confused when encountering cells that are very different from everything in the training data. In biology, this can happen frequently with compound treatments or cell perturbations that shift cells to unknown states. If we don’t have a means, such as anomaly detection or off-basis measurements, to remove such data, then we should expect model outputs to be populated with a mixture of accurate predictions and total misses.

Ultimately, AI can be an incredibly powerful tool to build profiles of complex biological states in an automated and unbiased way, but only if we can generate data that are optimized for AI analysis. AI (at least in its current form) is incapable of hypothesis generation and testing, so to maintain scientific rigor and progress, AI outputs need to be interrogated and held to the same standards as all scientific analyses. We therefore must encourage close collaborations between scientists and engineers for experimental design, analytical workflows, and model validation.

Mitochondrial dysfunction has been causally linked to aging and age-related diseases, yet it has historically been difficult to establish preclinical human-relevant systems that accurately capture sustained, pathological mitochondrial behavior. Recent progress in AI now enables us to generate profiles of functional cellular states from high-dimensional datasets that are representative of human health. By following the guidelines outlined above, we can apply these AI advances to develop effective mitochondrial therapeutics that unlock the power of our little powerhouses and maximize our health as we age.

Get the latest thought leadership from California’s life sciences sector in the quarterly Life Sciences Insights magazine. Share your own news and insights with CLS Member Voice.