How does deep learning reveal drug-driven changes in bacterial virulence?

Two virulence traits measured in the same cell

The paper's starting problem is measurement. Invasion — bacteria entering and persisting inside host cells — and genotoxicity — bacterial toxins damaging host DNA — are usually assayed separately and reported as population averages. Antibiotic protection assays, for instance, lyse host cells and count internalised bacteria as colony-forming units, which hides cell-to-cell heterogeneity. MALVINA instead keeps the co-culture intact, stains host nuclei, DNA-damage foci and bacterial signals, and uses a deep-learning classifier to keep only genuine, artefact-free infected cells before quantifying thousands of cells per well.

The output is a per-strain 'virulence scorecard' with separate metrics. Invasion width is the proportion of infected cells, i.e. how widely bacteria disseminate across the cell layer. Invasion depth is bacterial load per infected cell. Genotoxicity is split into direct damage in infected cells and population-level damage that also appears in uninfected neighbours. Keeping these apart matters: a strain can spread widely while carrying little intracellular load, and a drug can change one metric while leaving the other untouched.

Strain profiles diverge in width and depth

Across the E. coli isolates tested on human colorectal epithelial cells, the study found substantial strain-specific variation in both invasion width and invasion depth. Some isolates scored high on only one of the two dimensions, which the authors read as evidence of mechanistically different invasion strategies rather than a single virulence gradient. The genotoxic strains ATCC and CCR20 and the hyper-invasive LF82 recur throughout the experiments as reference points, alongside the typically non-invasive laboratory strain BW.

This is a descriptive result about the isolates in this panel, not a ranking of clinical risk. Its value is that the two invasion metrics are reported separately, so a strain that disseminates broadly is not automatically treated as equivalent to one that accumulates heavily inside fewer cells.

Co-infection reshapes invasion through competition and cooperation

Mixing strains changed invasion in ways neither strain showed alone. The non-invasive BW gained invasion width during co-infection with the hyper-invasive LF82 and CCR20, most prominently with LF82, which also increased BW's invasion depth — a pattern the authors describe as cooperative invasion or bystander uptake. Other pairings went the other way: co-infection with ATCC or CCR20 reduced BW's invasion depth, and ATCC with BW produced mutual antagonism, with both strains losing depth, consistent with competition for intracellular niches.

The genotoxic ATCC isolate behaved asymmetrically with LF82 and CCR20: it modestly increased its own dissemination while suppressing invasion of both hyper-invasive partners. The authors note that ATCC uniquely carries toxin- and host-modulation-associated determinants, including the autotransporter toxins Sat and Pic, hemolysin-associated loci and the immune-modulatory factor TcpC, plus an inactivated rpoS allele, but they explicitly state that the precise mechanisms remain to be determined. The interaction pattern is measured; the genetic explanation is a hypothesis.

Colibactin links competitor suppression to DNA damage

Genotoxicity was quantified relative to each labelled strain's own single-strain baseline. ATCC showed moderate population-level and direct genotoxicity alone, and this rose markedly when mixed with itself, more than for CCR20, which the authors attribute to density-dependent colibactin production. More strikingly, co-infection with the non-genotoxic strains LF82 or BW increased both direct and population-level genotoxicity of the colibactin producers, particularly ATCC — non-genotoxic partners potentiated colibactin-associated DNA damage.

A direct test supports colibactin as the link. Fluorescent LF82 co-infected with wild-type ATCC was suppressed more than LF82 co-infected with an isogenic mutant lacking the terminal step of colibactin biosynthesis; the mutant was markedly less effective at suppressing LF82 invasion. That connects intact colibactin to both host-cell genotoxicity and intracellular dominance, and supports the authors' interpretation of colibactin-mediated intracellular exclusion. The proposed downstream route — increased mutational burden or phage activation — is offered as a possibility, not demonstrated here.

Drug screening separates killing from virulence modulation

The team screened 23 compounds — antibiotics, biocides, antimicrobial peptides and chemotherapeutics — at approximately 1× and 2× the minimum inhibitory concentration, on cells infected with CCR20 or LF82. Crucially, extracellular bacterial counts were measured in parallel by colony-forming unit assays, so a drop in invasion or genotoxicity could be checked against simple bacterial killing rather than assumed to be virulence modulation.

That control changes how several results should be read. Gepotidacin, cisplatin and 5-fluorouracil produced near-zero genotoxicity shifts in CCR20 co-cultures, consistent with strong bactericidal activity at the tested doses: the reduced DNA damage mainly reflected bacterial clearance and lower extracellular toxin levels, not direct inhibition of pathogen-mediated genotoxicity. By contrast, omadacycline, doxycycline and tridecaptin markedly reduced direct genotoxicity while increasing CCR20 invasion depth, and the effect was specific to the genotoxic strain — pointing to interference with colibactin production or activity rather than killing.

Bacteria can amplify drug toxicity inside host cells

The paper also reports the reverse direction: infection altered host responses to drugs. The genotoxicity shift metric captures this by fitting, for each drug–strain combination, the log–log relationship between intracellular bacterial signal and DNA-damage signal in infected cells, then measuring how much the slope angle moves relative to the untreated condition. A non-shifting drug leaves bacterial-load-dependent genotoxicity unchanged; a shifting drug changes that relationship.

Shifts that depend on intracellular bacterial load, as seen with polymyxin B, are interpreted by the authors as bacterial modification of drug toxicity or of host-cell sensitivity. This is a distinct claim from virulence-targeting: here the bacterium changes what the drug does to the host, rather than the drug changing what the bacterium does. Both are read from the same single-cell measurements, which is what allows the two directions to be separated at all.

Where the workflow stops

MALVINA requires fluorescent tagging of bacteria, which may not be feasible for all strains or clinical isolates, and it relies on gentamicin to remove extracellular bacteria, limiting use with gentamicin-resistant strains. It provides a static snapshot rather than live tracking, and the classifier must be retrained for different host cell lines because morphology and staining vary. The platform is optimised for adherent human cell lines, so it does not recapitulate tissue complexity or immune interactions; the authors point to organoids as a future extension.

These constraints bound the interpretation of every result above: the findings describe E. coli isolates interacting with colorectal epithelial monolayers in a controlled co-culture, with drug effects measured at two doses around the minimum inhibitory concentration. They do not establish what happens in polymicrobial infection in tissue, nor do they identify the molecular route by which colibactin suppresses competitors.

Why the killing control changes the reading

The most decision-relevant distinction in this paper is between a compound that lowers genotoxicity because it clears bacteria and one that lowers genotoxicity while bacteria persist or even invade more deeply. Only the second group — omadacycline, doxycycline and tridecaptin against CCR20 — supports the authors' claim of virulence-targeting activity beyond bacterial killing. Grouping all genotoxicity reductions together would erase exactly the difference the workflow was built to detect.

For readers evaluating virulence-aware drug profiling, the practical priority follows from that split: pair any invasion or genotoxicity readout with an independent measure of extracellular bacterial viability, and check whether an effect is strain-specific. The paper's own example is that the same screening design yields bacterial clearance, virulence suppression and bacteria-dependent drug toxicity as three separate outcomes — and only the single-cell, multiparametric measurement keeps them apart.

Sources used in this answer