What “uninfected” cells reveal about an infection

A cell classified as uninfected can still show a DNA-damage signal. That is one of the distinctions preserved by MALVINA, a microscopy and machine-learning system that examines bacteria and their human host cells together. In its laboratory experiments, the researchers measured damage-associated fluorescence in cells containing bacteria and separately in cells classified as bacteria-negative.[1]

The label describes what the analysis detects in an image. It does not establish that a cell has never encountered bacteria or their products. Keeping that distinction visible makes it possible to study an infection’s effects beyond the cells where bacteria are found.

Bence Bognár and colleagues reported the method in Nature Communications on 29 September. Their main experimental system combined E. coli with Caco-2 human colorectal epithelial cells grown in laboratory wells. Across the work, they analysed 1.3 million host cells in 85,800 imaging fields from 1,320 wells. These are cellular observations within experiments, rather than independent patients.[1]

Four measurements, with different denominators

MALVINA separates how widely bacteria have spread from how much bacterial material each infected cell contains. It then makes a second separation between DNA-damage signals in infected and uninfected cells. The resulting four measurements answer different biological questions.[1]

Study author Viktória Lázár describes the question behind the split: whether invasion means concentrating bacteria in a few host cells or spreading them across many. A numerical illustration makes the consequence clear. Among 100 cells, ten infected cells carrying ten units of bacterial signal each produce the same overall mean as 100 infected cells carrying one unit each: one unit per host cell. Yet the infected fractions are 10% and 100%, and the burden per infected cell differs tenfold. Those are illustrative fluorescence units, not results from the study.[2]

The DNA readout uses γH2AX, a marker associated with the cellular response to DNA damage. Fluorescence intensity is a proxy for that response; it is not a count of mutations or a measurement of later cancer incidence. Separating the two cell groups therefore makes the experiment more informative without turning an image into a clinical diagnosis.[1]

The AI’s job is to sort the images

The analysis first locates nuclei and expands their outlines to approximate surrounding cell regions. It extracts image texture and fluorescence measurements, then uses a neural-network classifier to distinguish infected cells, uninfected cells and cells with bacteria adhering to them. This last category matters because a nearby bacterial signal can otherwise be mistaken for an intracellular one.[1]

For the colorectal-cell model, 2,373 manually labelled cells supplied the reference examples. The paper reports 86% classification accuracy using cross-validation with a 10% validation fraction. A separate classifier filters problematic images, including overlapping or detached cells. Those classification steps determine which cells enter each denominator, so their errors also matter to the biological measurements.[1]

The initial strain comparisons used three biological replicates with at least ten technical replicates. Measurements from many cells help describe variation within a well; repeating the biological experiment addresses a different question, whether the pattern recurs when the experiment is performed again.[1]

Burden and damage can move separately

LF82 and CCR20, for example, both showed high invasion width and depth. Their damage profiles separated: LF82 was among the two strains with the lowest genotoxicity, while CCR20 produced elevated signals in both infected and uninfected cells. LF82 still produced more direct damage signal than the uninfected control. Similar spread and intracellular accumulation therefore did not mean similar host-cell damage.[1]

The team also screened 23 drugs and measured extracellular bacterial survival. Sixteen compounds, about 70%, changed at least one invasion metric in one or more strains. Measuring survival alongside the cell images allowed changes inside host cells to be considered against changes in the bacteria available to infect them.[1]

In her account of the project, Lázár explains why the investigators went further than comparing treatment averages. They examined the relationship between bacterial signal and DNA-damage signal in the same infected cells. That lets them ask whether damage changes because bacterial burden changes, or whether the relationship between the two shifts as well.[2]

This is a more demanding definition of a useful laboratory result. A smaller bacterial total does not, by itself, describe the condition of the surviving host cells. Equally, an improved average damage score can conceal a different distribution between the infected and uninfected groups. The paired measurements reveal which of these questions the data can actually answer.

The authors adapted MALVINA to a second setting involving lung epithelial cells and Klebsiella pneumoniae, training new infection classifiers for that cell background. Its demonstrated versatility comes from an adaptable measurement method. Its practical contribution is to keep the bacterium’s location, its burden and the host cell’s response connected throughout the analysis.[1]

Study credit: Bognár and colleagues, Nature Communications, CC BY 4.0. This independently written explanation adapts the study’s findings and adds an illustrative calculation.[1][3]

Report an error or suggest a correction