The compounds 5,7,8-trihydroxyflavone, 5-deoxykaempferol, acacetin, apigenin, apigetrin, β-escin, chrysoeriol, chrysophanol, coumarin, cyanidin, delphinidin, diosmetin, diosmin, ellagic acid, eriocitrin, eriodictyol, eriodictyol-7-O-glucoside, eupatorin, ferulic acid, fisetin, flavanone, galangin, gallic acid, genkwanin, gentisic acid, guaiazulene, herniarin, hesperidin, homoeriodictyol, homoorientin, homovanillic acid, hyoscyamine, isorhamnetin, isorhamnetin-3-O-glucoside, isorhamnetin-3-O-rutinoside, isorhoifolin, juglone, kaempferide, kaempferol, kaempferol-3-O-rutinoside, kaempferol-7-O-glucoside, kaempferol-7-O-neohesperidoside, linarin, liquiritigenin, lutein, luteolin, luteolin tetramethylether, luteolin-3′,7-di-O-glucoside, luteolin-4′-O-glucoside, luteolin-7-O-glucoside, malvidin, mangiferin, maritimein, myricetin, myricitrin, myrtillin, naringenin-7-O-glucoside, narirutin, oleuropein, orientin, pelargonidin, pelargonin, phloroglucinol, pyrogallol, quercetin-3-O-(-6-acetylglucoside), quercetagetin, quercetin-3,3′,4′,7-tetramethylether, quercetin-3,4′-dimethylether, quercetin-3-methylether, quercetin-3-O-glucuronide, rhoifolin, robinin, saponarin, scopolamine, sennoside A, sennoside B, sulfuretin, tectochrysin, tiliroside, verbascoside and vitexin were purchased from Extrasynthese (Genay, France). The molecules (-)-norepinephrine, (+/-)-dihydrokaempferol, 1,4-naphtoquinone, 18α-glycyrrhetinic acid, 3,4-dihydroxybenzoic acid, 3,4-dimethoxycinnamic acid, 3-hydroxybenzoic acid, 4-hydroxybenzoic acid, 5-methoxypsoralen, ajmalicine, berberine, betanin, betulin, boldine, catechol, cholesta-3,5-diene, chrysin, cynarin, daidzein, emodin, galanthamine, genistein, guaiaverin, hesperetin, hydroquinone, lupeol, myristic acid, naringenin, naringin, p-coumaric acid, phloridzin, pinocembrin, plumbagin, quercetin, quercetin-3-O-β-D-glucoside, quercitrin, rosmarinic acid, rutin and silibinin were acquired from Sigma-Aldrich (St. Louis, MO, United States). 5,8-Dihydroxy-1,4-naphthoquinone and caffeine were from Fluka (Buchs, Switzerland), vanillin was supplied by Vaz Pereira (Santarém, Portugal), vicenin-2 was purchased from Honeywell (Charlotte, NC, United States). Chlorogenic acid was from PhytoLab (Vestenbergsgreuth, Germany). Cinnamic acid was obtained through Biopurify (Chengdug, China). Swertiamarin was from ChemFaces (Wuhan, China).

Minimum Essential Medium (MEM), Dulbecco’s Modified Eagle Medium/Nutrient Mixture F-12 (DMEM/F-12), fetal bovine serum, penicillin/streptomycin solution (penicillin 5,000 units/mL and streptomycin 5,000 μg/ml), trypsin-EDTA (0.25%), Qubit^® RNA IQ assay kit and Qubit^® RNA HS assay kit, the SuperScriptTM IV VILOTM MasterMix, the Qubit^® Protein Assay Kit, the anti-BiP mouse monoclonal antibody (Catalog # MA5-15619, lot WI3375321) and the anti-mouse secondary antibody (Catalog # 61-6,520, lot WB319554) were obtained from Invitrogen (Grand Island, NE, United States). Dimethyl sulfoxide (DMSO) was acquired from Fisher Chemical (Loughborough, United Kingdom). Isopropanol was obtained from Merck (Darmstadt, Germany). 3-(4,5-Dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT), calcium ionophore A23187, thapsigargin, RNAzol^®, chloroform, isopropanol, diethyl pyrocarbonate (DEPC), KAPA SYBR^® FAST qPCR Kit Master Mix (2X) Universal, Aβ_25-35, 2-(4-amidinophenyl)-6-indolecarbamidine dihydrochloride (DAPI), protease inhibitor cocktail, phosphatase inhibitor cocktail, acrylamide/bis-acrylamide 30% solution, bromophenol blue, glycerol, sodium dodecylsulfate (SDS), glycine, Trizma^® base, Trizma^® hydrochloride, sodium deoxycholate, sodium chloride, potassium phosphate monobasic, potassium chloride, sodium phosphate dibasic, sodium bicarbonate, D-glucose and Triton X-100 were purchased from Sigma-Aldrich (St. Louis, MO, United States), while formaldehyde was from Bio-Optica (Milan, Italy). The fluorescent probe Fura-2/AM, as well as the anti-GAPDH (Catalog # ab181602) rabbit monoclonal antibody and the anti-rabbit secondary antibody (Catalog # ab6721, lot GR172025-6) were obtained from Abcam (Cambridge, United Kingdom). Thioflavin T was from Alfa Aesar (Kandel, Germany). The WesternBright^® ECL HRP was supplied by Advansta (Menlo Park, CA, United States). Trans-Blot Turbo Mini 0.2 µm Nitrocellulose Transfer Packs were acquired from Bio-Rad (Hercules, CA, United States).

Since it is the most detailed and up to date database of annotated natural products (Sorokina et al., 2021), we started from the COCONUT database (https://coconut.naturalproducts.net) and created a data frame comprising all the molecules under study, where each row corresponded to a different molecule and columns to features. Among all the original features, we have decided to keep only 48, as the remaining ones did not contain relevant information for this work (SMILE notation, fragments, among others). Supplementary Table S3 contains a list of all variables kept and their description. All data was cleaned and explored using Python 3.9. For preprocessing data, we have used Simple Imputer, Standard Scaler, Ordinal Encoder and Label Encoder from Scikit-learn (Pedregosa et al., 2011). Dimensionality reduction was conducted using Scikit-learn (Pedregosa et al., 2011) library (Principal Component Analysis [PCA], Multidimensional Scaling [MDS], t-distributed Stochastic Neighbor Embedding [TSNE] and Uniform Manifold Approximation and Projection (UMAP) used the UMAP library (McInnes, 2018). For t-SNE we have used number of neighbors of 5, 13 and 45 and minimum distance of 0.01, 0.1 and 0.5. In the case of UMAP we have iterated through perplexity of 5, 20, 50 and 100 and number of iterations of 300, 900, 1,500 and 2,000. Unless otherwise specified, specifications of each algorithm were set for default. Graphics in Figures 1, 7, 8 and Supplementary Figure S3 were created using Seaborn or MatPlotLib and the remaining ones in GraphPad Prism 8.

MRC-5 fibroblasts (European Collection of Authenticated Cell Cultures, Porton Down Salisbury, United Kingdom) were cultured in MEM and SH-SY5Y cells (American Type Culture Collection, LGC Standards S.L.U., Spain) were cultured in DMEM/F-12. Both mediums contained 10% FBS and 1% penicillin/streptomycin and both cell lines were maintained at 37°C with 5% CO₂.

For viability assessment, MRC-5 fibroblasts were seeded at a density of 2 × 10⁴ cells/well. After 24 h the cells were incubated with the compounds of interest. 24 h later, the wells are aspirated, the medium was replaced by MTT at 0.5 mg/ml and incubated for 2 h. At the end of this period, the solution was discarded and the formazan crystals in the wells dissolved in 200 µL of a 3:1 DMSO:isopropanol solution. The absorbance at 560 nm was read in a Thermo Scientific™ Multiskan™ GO microplate reader.

Results are presented as the percentage of control and correspond to the mean ± SEM of at least three independent experiments, each of them performed in triplicate.

MRC-5 fibroblasts were plated on black bottomed-96-well plates at the density of 2 × 10⁴ cells/well. After 24 h, the fluorescent probe Fura-2/AM was added at 5 µM and incubated for 1 h. Then these were incubated with the selected nontoxic compounds in HBSS. The calcium ionophore A23187 was added at a final concentration of 5 µM after 2 h. Finally, 2 h later the fluorescence was read at 340/505 and 380/505 nm on a Cytation™ 3 (BioTek (Vermont, MA, United States)) multifunctional microplate reader. Data analysis was performed considering the ratio F_340/505/F_380/505. Results are presented as fold decrease vs. positive control and represent the mean ± SEM of at least three independent experiments, each of them performed in triplicate.

Cells were seeded at 1.6 × 10⁵ cells/well in 12-well plates, left at 37°C for 24 h and then incubated with the compounds that displayed positive results in previous assays. After 2 h, thapsigargin (Tg) was used as a positive control and added at a final concentration of 3 μM, and a period of 16 h of incubation ensued. The cells were lysed by aspirating the culture medium and adding 500 µL of PureZOL RNA isolation reagent, pipetting up and down multiple times. The lysate was kept at room temperature for 5 min and then RNA extraction ensued.

The extraction was performed by phase separation, by adding 100 µL of chloroform, shaking, incubating for 5 min and centrifuging at 12,000 g for 15 min at 4°C. The aqueous phase was collected and 250 µL of isopropyl alcohol were added, following 5 min of incubation and new centrifugation at 12,000 g for 10 min, at 4°C. The supernatant was then discarded, and the pellet was washed with 75% ethanol, vortexed, centrifuged at 7,500 g for 5 min at 4°C, air dried and resuspended in DEPC-treated water. The RNA in the sample was quantified resorting to the Qubit^® RNA HS assay kit. Sample integrity was then evaluated with the Qubit^® RNA IQ assay kit, and, if in appropriate conditions, 1 µg was converted to cDNA using the SuperScriptTM IV VILO™ MasterMix.

RT-qPCR reaction was performed in the following thermal cycling conditions: 3 min at 95°C, 40 cycles of 95°C for 3 s (denaturation), gene-specific temperature for 20 s (annealing temperatures listed on Table 1) and 20 s at 72°C (extension). The employed mastermix was KAPA SYBR^® FAST qPCR Kit Master Mix (2X) Universal. The reaction was conducted on a qTOWER3 G (Analytik Jena AG, Germany), and the data were analyzed on the software supplied along with the equipment (qPCRsoft 4.0). Primers for target genes were designed on Primer BLAST (NCBI, Bethesda, MD, United States) and synthesized by Thermo Fisher (Waltham, MA, United States). The respective nucleotide sequences are presented on Table 1. Product specificity was checked with melting curves. GAPDH was selected as reference genes for normalization of expression. All the RT-qPCR reactions were performed in duplicate, and the experiments were repeated, at least, four times. Results are presented as mean ± SEM.

On black bottomed-96-well plates, MRC-5 and SH-SY5Y cells were seeded at the densities of 2 × 10⁴ and 3 × 10⁴ cells/well, respectively. In the following day, they were incubated with the selected molecules in the presence of Aβ_25-35 at 10 µM. After an incubation period of 24 h, cells were washed with HBSS and incubated for 30 min with thioflavin T at 5 µM (prepared in HBSS). At this point, fluorescence was read at 450/482 nm on a CytationTM 3 (BioTek (Vermont, MA, United States)) multifunctional microplate reader. Results represent the fold decrease on fluorescence compared to the positive control. We present the mean ± SEM of, at least, three independent assays, individually performed in triplicate.

To image protein aggregates, cells were seeded on clear bottomed-96-well plates at the aforementioned density and treated with the molecules of interest for 24 h in the presence of Aβ_25-35 at 10 µM. Cells were fixed on 4% formaldehyde at room temperature for 15 min and incubated with 10 µM thioflavin T for 60 min. The counter staining was performed with DAPI at 0.25 μg/ml, for 30 min. Finally, three washing steps with HBSS for 5 min were carried out and cells were imaged under an Eclipse Ts2R-FL (Nikon) fluorescence microscope equipped with a Retiga R1 camera. A FITC filter was employed to observe thioflavin T and DAPI staining was imaged using a DAPI filter. Images were analyzed resorting to the CellProfiler software version 4.2.1 (Broad Institute of MIT and Harvard, Cambridge, MA).

MRC-5 and SH-SY5Y cells were seeded in 6-well plates at the densities of 3.2 × 10⁵ and 4.8 × 10⁵ cells/well. After 24 h, they were incubated with compounds selected as promising in previous experiments. After 2 h, Tg at 3 µM was added, and the plates were left in the incubator for another 22 h. By the end of this time frame, cell samples were lysed in RIPA buffer containing 1% protease inhibitor cocktail and 1% phosphatase inhibitor cocktail for 15 min at 4°C, and subsequently centrifuged at 14,000 g for 15 min to remove cell debris. The supernatant was then collected, and the protein content was determined resorting to the Qubit^® Protein Assay Kit.

The following step was to submit 30 µg of each sample, previously denatured at 76°C for 10 min, to 10% SDS-PAGE electrophoresis, in a gel containing 10% acrylamide/bisacrylamide. The proteins were then transferred to a nitrocellulose membrane on a Trans-Blot^® Turbo (Bio-Rad; Hercules, CA, United States). The membrane was blocked in a PBS solution containing 5% non-fat milk and 0.1% Triton X-100 for 1 h with orbital agitation. Hereupon, the membrane was incubated with the primary antibody overnight, at 4°C, with agitation. The anti-BiP mouse monoclonal antibody was used at a working dilution of 1:1,250, while the anti-GAPDH rabbit monoclonal antibody was used at a dilution of 1:10,000. A washing step was then performed with PBS 0.1% Triton X-100 for 30 min. Then, the secondary antibody was incubated for 2 h at room temperature. The anti-mouse secondary antibody was used at 1:2,000, and the anti-rabbit secondary antibody at 1:3,000. At this point, two more washing steps were performed with PBS 0.1% Triton X-100 for 15 min and then with PBS, again twice for 15 min. Finally, bands were detected by addition of a chemiluminescent substrate, WesternBright ECL HRP. The resulting bands were visualized on a ChemiDoc™ Imaging System (Bio-Rad, Hercules, CA, United States). Their relative optical density was calculated by densitometry resorting to the Fiji/ImageJ software version 1.51 (Fiji, Madison, WI, United States) and normalized against the loading control (GAPDH). Results are presented as mean ± SEM of, at least, three independent experiments.

The statistical analysis was performed using GraphPad Prism 8 software. We resorted to unpaired Student’s t-tests to compare every single individual treatment to the control group. Values of p < 0.05 were considered statistically significant. Outliers were identified and excluded by the Grubbs’ test.

We compiled a library of 134 molecules of natural origin (Supplementary Table S2). The average molecular weight was 358 g/mol, the inferior limit being 110.1 g/mol (catechol and p-hydroquinone) and the upper limit 1,131.3 (β-escin) (Figure 1). The computed natural product-likeness score, a parameter that tries to quantify how much a given molecule is “natural product-like” (Vanii Jayaseelan et al., 2012) had an average of 1.29 (min: 0.4, max: 2.72, Figure 1). Over 50% of the molecules displayed no violations of Lipinski’s Rule of 5 with over 30% displaying 2 or more (Figure 1), a relevant trait for candidate drugs.

For chemical ontology, we have used the computed chemical classes generated by ClassyFire (Djoumbou Feunang et al., 2016). From a chemical point of view, the distribution of the molecules by chemical superclass can be found in Figure 1. Phenylpropanoids and polyketides corresponded to the majority of the entries (65.9%), followed by benzenoids (14.4%) and lipids or lipid-like molecules (6.1%). Within phenylpropanoids, flavonoids were the most prevalent molecules (Figure 1), followed by cinnamic acids and their derivatives. In the case of benzenoids, simple phenols were predominant, while in the case of lipids, prenol lipids were the most representative (Figure 1). The chemical structures of all of the molecules are represented in Supplementary Figure S1, separated according to the herein mentioned chemical superclasses.

Given the high number of molecules involved, all compounds were tested at a single concentration (50 µM). Statistically significant decreases in MTT reduction were interpreted as negative impact on cellular viability and the corresponding molecules excluded from subsequent experiments. As evidenced by Figure 2, 37 molecules (depicted in Supplementary Table S2) were dropped from the pipeline for their toxicity.

The remaining 97 compounds were screened for potential protective capacity against ER stress. The legend for the heatmap presented in Figure 2 can be consulted in Supplementary Table S2, while Supplementary Figure S2 contains the individual results for each molecule represented in the bar charts.

The next step involved the screening of the nontoxic molecules for potential protection against the onset of ER stress. Towards this goal, and considering the high number of molecules still in the pipeline, we selected changes on calcium levels as an indicator of compromised ER homeostasis. The maintenance of the high concentration of calcium in the ER lumen creates an appropriate environment for protein folding, and thus changes on its levels heavily impact ER function (Kraus and Michalak, 2007; Carreras-Sureda et al., 2018). This organelle is able to maintain its calcium levels due to the abundance of resident calcium-buffering proteins, like calnexin and calreticulin (Coe and Michalak, 2009). Under stress conditions, the ER may release calcium through channels, such as IP₃R and RyR. Eventually, this may lead to increased calcium in mitochondria, indirectly inducing the generation of ROS and triggering the dissipation of the mitochondrial membrane potential, ultimately leading to the release of resident proapoptotic factors (Bhattarai et al., 2021).

To evaluate the potential protective effect of the molecules under study, we conducted experiments with the fluorescent probe Fura-2/AM. Calcium ions were detected employing the divalent cation ionophore A23187 (5 µM) as a positive control for calcium leakage into the cytosol and molecules that could inhibit or counter its effect were searched. Calcium ionophore A23187, or calcimycin, is a calcium-binding molecule that increases the cellular permeability of ionic calcium (Verma et al., 2011). There are multiple examples in the literature of the use of this molecule to increase cytosolic calcium levels, inducing ER stress, UPR activation and eventual apoptosis or autophagy (Castillero et al., 2015; Jung et al., 2015; Shinde et al., 2016).

Under these experimental conditions, we were able to identify several molecules that significantly prevented or ameliorated the calcium overload caused by A23187, as displayed in Figure 3. Such molecules included six polyphenols (5-deoxykaempferol, delphinidin, fisetin, naringenin, naringin, quercetin-3-O-β-D-glucoside), one anthraquinone (sennoside B), one fatty acid (myristic acid) and one biogenic polyamine (spermine). The full list of molecules tested in this assay can be consulted in Supplementary Table S3, while Supplementary Figure S3 represents these data in the bar charts.

The non-toxic molecules that successfully lowered Ca²⁺ levels were considered potential candidates for countering ER stress. As so, they were evaluated for their effect upon UPR signaling pathways.

Thapsigargin, a natural sesquiterpene lactone that is widely used as UPR inducer, was the positive control. This molecule acts as an irreversible SERCA pump inhibitor, leading to an impaired calcium homeostasis and subsequent onset of organized cell death mechanisms (Pereira et al., 2015). Accordingly, as evidenced by Figure 4, we can see that thapsigargin (3 µM) significantly increased the expression of all the genes of interest, namely the transcription factors DDIT3 and ATF4, the chaperones HSPA5 and HSP90Β1 (BiP and GRP94 proteins, respectively), and EDEM1, a gene coding for the ER degradation-enhancing alpha-mannosidase-like protein 1, an enzyme involved in ER-associated degradation.

5-Deoxykaempferol at 50 µM inhibited the expression of DDIT3, HSP90Β1, EDEM1 and HSPA5, even though it did not significantly decrease ATF4 expression. Fisetin was the only molecule to prevent the increase of all target genes in a significant manner. Delphinidin negatively impacted the expression of ATF4, HSP90Β1 and EDEM1. Myristic acid and quercetin-3-O-β-D-glucoside displayed a similar pattern in the protection against UPR activation, by modulating HSP90Β1 and EDEM1 expressions. Sennoside B acted upon ATF4 and EDEM1, while naringenin, naringin and spermine failed to inhibit any of the selected genes.

As discussed earlier, activation of the UPR affects the capacity of cells to fold proteins properly, which frequently results in the formation of protein aggregates and may lead to proteinopathies (Doyle et al., 2011; Ogen-Shtern et al., 2016). Conversely, several diseases, notably those of neurodegenerative nature, are associated with these aggregates (Kim et al., 2016). As so, new molecules capable of countering ER stress are expected to be able to prevent or lower the amounts of protein aggregates in a biological context. To confirm this, we proceeded to an experimental model of ER stress-triggered protein aggregation using Aβ_25-35 as a protein aggregation inducer. Aβ_25-35 is an active fragment of Aβ₄₂ (Zhou et al., 2016), a peptide that is associated to the pathogenesis of Alzheimer’s disease (AD) and has been characterized in experimental models, where it can lead to the formation of protein aggregates in vitro and AD in vivo (Liu et al., 2017).

As shown in Figure 5A, incubation of MRC-5 cells with Aβ_25-35 elicits protein aggregation, as assessed with thioflavin T, a benzothiazole that presents high fluorescence when bound to β-sheet-rich structures common in protein aggregates, both in a microplate-based (Figure 5A) and an imaging method (Figure 5B). In the same conditions, delphinidin, fisetin, myristic acid, sennoside B and quercetin-3-O-β-D-glucoside significantly inhibited this effect.

Our Western blotting results show that the exposure of Tg-insulted MRC-5 cells to delphinidin and fisetin clearly reduced BiP expression, and, consequently, we presume that the UPR activation in the presence of these molecules is mitigated.

For this assay, the concentrations of the molecules under study had to be adjusted to the highest nontoxic concentration towards this cell line. Fisetin was tested at 25 μM, while delphinidin was tested at the concentration used in MRC-5 cells (50 µM). The effect of fisetin was not replicated on SH-SY5Y cells. We hypothesize that this may be due to the reduction of the concentration that was necessary in order not to impact cell viability in this model. On the other hand, the effect of delphinidin translates to these cells, virtually nullifying the advent of protein aggregates, as evidenced by Figures 6A,B.

Considering all the information generated during this work, together with the chemical, structural and topographical information available for the molecules under study, we were interested in understanding if the physico-chemical properties of the library could, per se, be a good indicator for its activity. To this end, we took all the molecules used and computed multiple pairs of their properties, sorting them as active or inactive according to the results of the aggregation assays depicted in Figure 5. While there seemed to be a tendency of active molecules to be apart from inactive molecules in some cases (Supplementary Figure S4), it soon became apparent that a pairwise comparison could leave behind important features of the molecules, as it reduces high dimensional data to just two dimensions. As so, we have decided to use other methodologies that could take into account all the features of the library, namely dimensionality reduction algorithms, which allows the combination of different features into principal components that can be more easily graphed and understood. We first conducted a correlation study in order to identify molecular features that were highly correlated, which could negatively impact the performance of the statistical analysis given the redundancy it would add. Figure 7 shows the original features, where it is clear that many are highly correlated, for example, the features “total atom number” and “number of carbons” (0.98), as one directly affects the other. Such features were dropped, and the remaining ones were used. We also decided to assess the robustness of the methods in identifying the chemical space occupied by the molecule that was shown to be active in both cell lines, namely delphinidin.

We initially used the classical algorithm for principal component analysis (PCA), which is widely used both for chemometric and biological data (Meglen, 1992; Boileau et al., 2020). As shown, PCA was unable to identify a chemical space occupied solely by the molecule that was shown to be active in all the assays used (Figure 8A), the same being true for multidimensional scaling (MDS, Figure 8B). In light of this, we applied a suite of other algorithms, both individually and in tandem with PCA, namely uniform manifold approximation and projection (UMAP) and t-distributed stochastic neighbor embedding (t-SNE).

Differently from PCA and MSD, the combined use of PCA + UMAP led to 2-principal component-based space, in which delphinidin clearly occupies an outlier space. We have tried distinct combinations of parameters, as stated in the Materials section, the number of neighbors of 45 giving the best performance, both for minimum distance of 0.01 and 0.1. In order to confirm this result, we have employed an additional algorithm, namely PCA + t-SNE. Again, it was shown that delphinidin, together with an outlier, occupies a space that is distinct from over 98% of the library.