Discovery
|
Matthew Kunicki, Christoph Eberle, PhD
AI-Integrated Discovery in Flow Cytometry is Within Reach
Medical research technologies are advancing through artificial intelligence (AI), but why has the same impact not yet been seen for a tried-and-true technology like flow cytometry?
While an era of AI-native biology in healthcare and drug development becomes a reality, the quality and precision of AI-outcomes are incumbent upon the scientists and clinicians who lead the charge. Flow cytometry is a cornerstone technique for high-throughput, single-cell analysis. Redundant processes associated with it, such as antibody panel design, literature review, data visualization, and gating, have seen moderate to large improvements for daily research. Still, the larger outcome sought by many in industry is the effective translation of discoveries in medicine and science. With many wondering about AI’s true potential, the flow cytometry industry has occupied a unique position with AI, as trends and hype have focused on healthcare technologies better suited to AI adoption. Although the reasons are many, two attributes of effective AI modeling highlight a key difference between flow cytometry and other technologies that adopt AI: data generation and data interpretation.
The first is data generation. For clarity, large language models (LLMs), such as GPT-4, Claude, or Llama, and large quantitative models (LQMs), such as AlphaFold, SandboxAQ, or AtlasGEN, are AI that use language or numerical data, respectively. Flow cytometry generates numerical data representing multi-parametric single-cell light-scattering and fluorescence intensity.
Researchers annotate these numerical datasets with language to give biological meaning, but current AI in the field has struggled to go beyond improvements in this semantic process. Unlike new technologies however, flow cytometry has been a fundamental tool for medical research for over 50 years, and the root of this struggle with AI adoption is well documented. It is known that variability in flow cytometry datasets is not only a result of inconsistencies between reagents, instruments, and researchers, but is also introduced through assay design. Key decisions in antibody panel design, sample processing, methodology, and instrument configuration have inevitably led to a vast array of flow cytometry data worldwide. Still, the value of these numerical data in discovery and translation is clear, as evidenced by decades of effort from regulators and scientific societies supporting this technology in medicine. Nevertheless, LQMs have failed to leverage the vast amount of flow cytometry data available, which is a stark contrast to their success with proteomics, genomics, or imaging.
The inherent variability in flow cytometry data was prominently shown with some of the earliest machine learning (ML) software applications. T-distributed stochastic neighbor embedding became a standard technique for visualizing single-cell data at exceptional speed, and many improvements and additional algorithms for flow cytometry were developed soon after. Despite their widespread use, these ML tools fail to distinguish patterns in the numerical dataset driven by technical variance rather than biological outcomes. Notably, deviations from experiment to experiment, performed exactly the same way within a single study, called batch effects, would change the numerical values enough for ML tools to treat the same type of single cells as different cell populations. These effects ultimately shed light on:
(1) How sensitive ML (and later with AI) computational workflows were to data patterns, and how sensitive flow cytometry was to real-world technical deviations.
(2) How flow cytometry AI-adoption needs a data-level solution to elevate it into the era of AI-native biology and the LQMs available today.
The second attribute is data interpretation. Putting aside data variability, the semantic challenge of interpretation doesn’t start with the numerical data. It starts with experimental design. Many applications of flow cytometry exist for scientific purposes, such as immunophenotyping, cytokine secretion, molecular signaling, and more. Each imposes constraints on which molecules can be detected simultaneously, which can be somewhat difficult for the researcher, depending on the molecules they require for their scientific hypotheses.
While new discoveries offer researchers additional choices, what is often observed is a tendency for entire laboratories to define cell types using their own selection of molecules. Unfortunately, this raises the question of whether two laboratories are observing the same cell type, which further confounds data variability. Global scientific efforts to standardize cell annotation aim 1,2,3 to unify differences in cell type definitions, recognizing that standardization should still accommodate diverse laboratory interpretations.
Although standardization is a necessary and noble effort to understand ground truth, the choice of molecules deeply affects reported outcomes, such as cell-type abundance and classification by gating. Gating is a pre-processing technique after data collection and is often done before ML or AI computational algorithms are applied. A standardized effort to annotate cell types would inevitably involve cell classification by gating or an alternative algorithmic approach, but, at its core, evidence is needed to bridge one cell type to another when molecular definitions differ from dataset to dataset. Computational tools that take an agnostic approach to the data could reveal significant overlap between cell type definitions, which would be a major win for standardization. In practice, however, efforts are still ongoing.
Ideally, some outcomes in cell-type abundance should align across studies, reflecting common biological characteristics, but even these numerical abundances are subject to the same technical variability in data generation, leading to disparate outcomes. While alignment in biological or medical outcomes between labs supports flow cytometry’s rigor, these numerical variations in the data are problematic for the robust application of AI (LQM) compared to other technologies.
It goes without saying that a widely used and historically proven technology in healthcare, such as flow cytometry, is a desirable target for AI; hence, there has been a growing effort to develop solutions. The sensitivity to technical variability is not just a problem, but a clue to real-world and computational needs that could improve data generation and interpretation. Reputable institutions have paired large multi-omic studies with flow cytometry to increase the number and quality of data, recognizing that the breadth of data collected could create a foundation for improved AI modeling.
Consortia have published hundreds of guidelines, primary articles, and letters examining the influence of experimental design and flow cytometry performance across hundreds of relevant molecules. Companies continue to improve products for data quality given the growing scientific demand for parity among laboratories. There is no doubt that the decades of innovation in flow cytometry are due to its importance in research. This is why even alternative technologies, such as single-cell RNA sequencing, have taken on some of these needs, such as immunophenotyping. However, if AI-adoption and the ubiquity of sequencing tools could replace flow cytometry, why have they not?
The question here is not an allusion to the decline of a once-great technology in biology due to progress in AI. It is quite the opposite. To the expert observer, the opportunity to jump from AI simply improving daily research routines to AI accelerating the effective translation of discoveries into medicine is closer than ever for flow cytometry. New real-world tools within sample collection, processing, and preservation have shown significantly improved sample integrity, helping maintain the detection and propensity of molecules in the cell.
Antibody reagents and instruments are continually tested and improved to ensure consistency. Recent studies show increasing cases where RNA-to-protein equivalence does not hold. Algorithms are reaching levels of sophistication and robustness to help overcome minor batch effects within studies. The process of selecting an instrument, designing an antibody panel, running your experiment, and analyzing the data correctly has become more automated. All these capabilities feed into larger drug development platforms where AI assists with compound prioritization and translational readouts, especially target optimization, biomarker identification, and immune monitoring. The need for more effective, less costly pathways from discovery to medicine remains, and researchers continue to rely on flow cytometry to validate ground truth in an era of AI-native biology.
To the biologists or clinicians interested in what is next, stay tuned for our second article, “Why Universal Comprehensive Discovery is the Next Leap with AI”.
References:
