The nuclear morphology can be used as a biomarker for cellular senescence.
We sampled the BNN ensemble or deep ensemble in order to estimate their uncertainty (Extended Data Figure). 3a, b). The correct predictions are more likely to be in the lower or higher ranges of the output. This indicates greater certainty regarding the samples’ states. Incorrect predictions, however, tend towards the threshold of 0.5. By removing predictions from the middle of the model, we can increase the confidence in its predictions. We tested a variety of thresholds on several models (Extended data Fig.). The accuracy of the models is significantly improved by removing the ambiguous samples. The same approach was used for other models including the IR- and RS-models (Extended Data Fig.). This approach (Fig. 3g, h) increases accuracy by 10-15% while reducing the number of cells taken into consideration.
We analyzed human cells that were induced to senescence with 10 Gy IR, and imaged on days 10, 17, 24, and 31. The predictor can identify senescence in all four time points, with a probability that increases between days 10 and 17 but decreases by day 31, (Extended Data Figure). 4a). It is interesting to note that, when examining the distribution of the predictor, it becomes apparent that the number of nonsenescent peaks increases after day 17. This suggests that only a few cells are able escape the induction of senescence and ultimately overgrow senescent peaks (Extended Data Figure). 4b). When we examine markers of proliferation over time, PCNA decreases until day 17 after which expression begins to return (Extended data Fig. 4c). The stain intensity of p21Cip1 shows an opposite pattern, initially increasing and then decreasing slightly by day 31, (Extended Data Figure). 4D). We also observed a decrease in DAPI intensities for days 10-17, indicating senescence. However, the intensity returned to normal by day 31. 4e). We evaluated if markers for senescence and proliferation correlated with predicted senescence to confirm that the predictor accurately determines senescence, even 31 days following IR. Accordingly, cells that were predicted to be senescent had higher p21Cip1 and lower PCNA levels and lower DAPI intensity and vice versa. 4f-h). For predicted senescence morphologically, the area and aspect is higher, while convexity (Extended Data Figure) is lower. 4i-k). Finaly, a simple count of nuclei confirms the growth following IR (Extended Data Fig. 4l). Overall, the senescence prediction predictor captures state during development and is in agreement with multiple markers.
The appearance of persistent nuclei of DNA damage markers 53BP1 and gH2AX (refs. 31,32). Using high-content microscopes, we examined our base data set, which included control, RS, and IR lines, for damage foci. We found that the mean number of controls was below 1, whereas RS showed 4.0 gH2AX foci, 2.0 53BP1 and IR showed 3.4 gH2AX foci, and 3.053BP1 (Fig. The data set included control, RS and IR lines. We examined the damage foci using high-content microscopy. We found that controls had a mean count of less than 1 for each marker. RS had 4.0 gH2AX and 2.0 53BP1 foci while IR had 3.4 gH2AX with 3.0 53BP1 foci (Fig. 5a). We calculated Pearson correlations between predicted aging and the counts of gH2AX foci and 53BP1 foci and found that there was a moderately high correlation (around 0.5) across all conditions. 4c). The same association can be seen when plotting the foci count and senescence predictions, where predicted senescence is shown to flip from low (low) to high (high), along with shifts of foci counts. 5b). The feature reduction masks internal nuclear structure. However, it is noteworthy that senescence predictions correlate with the number of foci. We also compared predicted senescence with area. The correlation was around 0.50. There is a significant correlation between the number of foci and senescence.
Source:
https://www.nature.com/articles/s43587-022-00263-3
