A live-cell, high-content imaging survey of 206 endogenous factors across five stress conditions reveals context-dependent survival effects in mouse primary beta cells

We set out to compare endogenous soluble factors that may directly promote or inhibit survival under five controlled stress conditions in dispersed mouse islet cells. Mouse islet cells were chosen because of their reproducibility and low baseline rates of apoptosis, relative to human islet cells where the in vitro rates of cell death are typically much higher and more variable. The cells were concurrently treated with a library of 206 recombinant factors compiled using our previous bioinformatics analysis of islet cell ligands and receptors [7], along with candidates from the literature (Fig. 1a). Row and column variability in cell seeding density and cell death was negligible (ESM Fig. 1a,b). Cells cultured in 0.1% FBS and in the absence of serum displayed the same levels of death (ESM Fig. 1c). Analysis of dispersed mouse insulin-1 promoter-green fluorescent protein (MIP-GFP) islet cells suggested that beta cells were more sensitive to these stress conditions compared with the non-beta cell population within the same cultures (ESM Fig. 1c,d).

High-content, image-based analysis can simultaneously assess multiple parameters for internal validation, providing a level of redundancy that increases confidence in the results. We measured the loss of Hoechst-positive cells and the accumulation of PI-positive cells as indices of cell death. This captures multiple forms of cell death, including the ‘partial apoptosis’ that we recently demonstrated is the predominant mode of death in cultured primary beta cells [10]. The results of these studies revealed both pro-survival and pro-death factors within our library of endogenous ‘biologic’ factors, consisting mostly of recombinant full-length proteins (Figs 2, 3, 4, 5 and 6). Heatmaps showed the relative agreement between the replicate experiments (Panel c in Figs 1, 2, 3, 4, 5 and 6, ESM Table 2). As expected, cells treated with our positive control (10% FBS, presumably containing high concentrations of many islet survival factors) displayed the lowest PI levels. The agreement between the measurements of cell loss and PI incorporation can be seen in the distribution of the top and bottom parts of Panel a in Figs 2, 3, 4, 5 and 6, and is especially evident in stress conditions with larger effects (i.e. cytokines, thapsigargin, palmitate). Divergence is expected if specific factors modified the adhesion of dead or dying cells. The number of early apoptotic, AnnexinV-positive and PI-negative cells was also analysed, but as we have recently reported, these are relatively rare and their analysis is less informative [10].

First, we compared effects of each factor across all of the tested stresses to identify factors that might be generally protective or generally deleterious. We present the data here, and elsewhere, in multiple ways. The heat map in Fig. 1b represents the simple average rankings of PI⁺ cell count and cell loss, at two time blocks (0–24 h, 24–48 h), across all conditions (see ESM Methods). This analysis is meant to be exploratory and does not provide statistical information on the extent of protection or relative differences in protection among factors. Nevertheless, several known anti-apoptotic factors, including insulin [4, 14, 15] and adiponectin [16‐19], were highlighted, as well as unexpected factors. Pan-protective factors included those that were ranked amongst the top ten most potent survival factors under one or more conditions, along with those displaying moderate but consistent effects across all conditions. A sortable list of these results can be found in ESM Table 2. We validated the survival effects of olfactomedin 1 (OLFM1), across multiple conditions, and other factors on MIP-GFP beta cells (ESM Fig. 2-5). We confirmed that OLFM1 dose-dependently promotes human islet cell survival under lipotoxic conditions (ESM Fig. 6). Mechanistically, OLFM1 significantly downregulated p38 MAPK and caspase-3-dependent apoptotic signalling (p < 0.05; ESM Fig. 7).

We measured the consistency between replicates using rank product testing. A normalised robust ‘cell death score’ was log₂ scaled and then divided by the MAD of scores for the plate. Results for each well were then aggregated for day 1, day 2 and day 3 as the median cell death score for the day, providing a measure of the durability of protection. Statistical significance was measured as the reproducibility of cell death score ranks across replicate plates. As the factors were pre-selected, with expected and non-independent effects, we reasoned that naive multiple testing correction would be too conservative. Factors that were nominally significant (i.e. p < 0.05) on any day are illustrated with a triple triangle, with the significant days illustrated by filled triangles in PCA plots (Fig. 1c, ESM Fig. 8-12). The protective effect of 10% serum, our positive control, was globally statistically significant (p < 0.05) over all 3 days of testing, but is not shown in PCA plots so that the test factors can be more easily viewed. The complete analysis, with p values for each factor can be found in ESM Table 3. Semaphorin 4A (SEMA4A) held up against the most conservative of multiple testing correction, and 41 nominally statistically significant factors were identified as being pro- or anti-survival over all 3 days tested. Each will require further testing to rule out multiple testing false positives. Using this approach of nominal statistical significance, adrenocorticotropic hormone (ACTH) and neurotrophin 4 were consistently significantly protective over all 3 days (p < 0.05), whereas OLFM1, somatostatin, chemokine (C-X-C motif) ligand 11 (CXCL11), fibroblast growth factor (FGF)1, IGF1, bone morphogenic protein 15, ephrin 4A, xenin 8 and FGF12 showed significant time-dependency of protection. The results differ from the raw average of rankings for several reasons, including the omission of cell count data (which was informative in the qualitative analysis, but difficult to combine into the ‘cell death score’), the adoption of raw data normalisation and the inclusion of the 3rd time block (49–60 h). Collectively, results of both qualitative and statistical analyses suggest that factors exist that may protect islet cells under a wide range of stress conditions.

Comparing the effects of our chosen factors among stresses revealed that each stress had a relatively distinct complement of pro-survival factors and that there was little overlap between the highly ranked and nominally significant pro-survival or anti-survival factors found for each of the five stress conditions (Figs 1, 2, 3, 4, 5 and 6). For example, oncostatin M, from the cytokine family that includes leukaemia inhibitory factor, granulocyte-colony stimulating factor and IL-6 [20], was protective in context of both thapsigargin and high glucose, but no such evidence could be found in the context of low glucose serum withdrawal and the toxic cytokine cocktail (Fig. 6, ESM Table 2). Members of the neuroligin family, including NLGN1 and NLGN4, showed bias towards protection in endoplasmic reticulum (ER) stress and lipotoxicity, which we and others have shown are mechanistically linked [13, 21], and these findings were confirmed with secondary validation (ESM Fig. 4 and 5).

Comparing different stress conditions also revealed factors that were pro-survival under one condition, but pro-death under others. Most strikingly, semaphorin 4A, known for its roles in axon guidance, morphogenesis, carcinogenesis and immunomodulation [22], was the 2nd highest ranked (and most statistically consistently highly ranked) protective factor in the context of palmitate lipotoxicity, but it promoted cell death under all other conditions tested, including the baseline serum-free condition and the high glucose serum-free condition where it was the most consistently toxic factor (p < 0.05; Figs 1, 2, 3, 4, 5 and 6).

Some of our stress conditions were more resistant to protection than others (Figs 1, 2, 3, 4, 5 and 6). For example, only 37 factors provided any numerical protection of islet cells from palmitate above the negative PBS control, with the rest of the factors being apparently neutral or exacerbating lipotoxic cell death (Fig. 5). On the other hand, more than half of the factors provided some protection in the context of ER stress induced by a moderate dose of thapsigargin (Fig. 4). The shallow shape of the curve in Fig. 2a suggests that serum withdrawal alone was not a very severe stress in our studies, most likely because the islet cells can supply themselves with autocrine/paracrine survival factors.

We compared the concentration-dependent effects of each factor on cell survival under 5 mmol/l serum-free conditions (ESM Fig. 13). Some factors showed classical concentration-dependent effects. However, other factors exhibited bell-shaped dose–response curves, a phenomenon we have consistently observed with insulin [4, 14]. Temporal analysis revealed that some factors showed protective effects throughout the entire time course, while others were only protective early. It remains to be determined whether these transiently effective factors were rapidly degraded or taken up following treatment or whether the cells displayed receptor desensitisation.

When survival effects were analysed by pooling rank data from both the 0.1 nmol/l and 10 nmol/l concentrations, some factors that were not originally identified as protective when given in moderate concentrations, were shown to have efficacy at the lower concentrations (ESM Fig. 14). For example, cryptic family 1, γ-melanocyte stimulating hormone, somatostatin and ephrin-B2 were ranked with higher protective effects when the low and moderate concentrations were taken into consideration. Rank product testing of the entire 0.1 nmol/l data set (i.e. all factors across all tested conditions) identified hepatocyte growth factor, vascular endothelial growth factor-C, sonic hedgehog and serum amyloid-A1 as factors with statistically significant consistent protection across all 3 days at this dose (p < 0.05; ESM Fig. 15). Identifying factors with pro-survival effects under lower concentrations is widely considered to be important for therapeutic development because low effective concentrations may help reduce off-target effects.

Qualitative analysis of the canonical signalling pathways stimulated by the protective factors revealed that pro-survival signalling could be mediated by a number of known pathways, including the janus kinase (JAK)–signal transducer and activator of transcription (STAT) cytokine receptors, G-protein-coupled receptors, tyrosine kinase receptors, serine/threonine kinase receptors and axon guidance receptors (Fig. 7a, b). These analyses revealed that specific signalling pathways were more important in the context of certain stresses, relative to others. In the baseline serum withdrawal condition, factors that stimulate phospholipase C and/or the activation of adenylyl cyclase tended to be protective, whereas factors that inhibit adenylyl cyclase were less effective. In the context of thapsigargin or palmitate, activation of adenylyl cyclase was identified as the predominant pro-survival G-protein-mediated pathway.

We also assessed insulin accumulation in the media. We identified factors that promoted both survival and insulin accumulation in the media, including angiopoietin 2 in the 5 mmol/l glucose serum-free condition (Fig. 2b). Our crude data do not allow us to distinguish whether the survival effects were due to insulin independent survival signalling or pro-survival autocrine insulin signalling [4, 15]. Other factors were protective while inhibiting insulin accumulation in the media. It is known that inhibition of calcium flux can protect beta cells under specific conditions [23, 24], while these same manipulations block insulin secretion [25]. We also detected factors, such as semaphorin 4A, which triggered cell death and increased insulin in media, likely secondary to the loss of cell integrity (Figs 2b and 3b). The effects of soluble factors on insulin release were context-dependent. Analysis of insulin release was not a primary endpoint in the present study, although these data nonetheless provide a starting point for additional detailed studies.