QAL333’s antitumor activity and predictive modeling: integrated transcriptomic-bioinformatic analysis reveals selective cytotoxicity and sensitivity determinants

Abstract

Despite significant advances in chemotherapy, tumor heterogeneity and resistance mechanisms continue to limit clinical efficacy, underscoring the need for novel compounds with mechanistic biomarkers that enable patient stratification. QAL333 is a newly synthesized isoindolin-1-one derivative identified as a promising anticancer candidate. We evaluated its activity across 15 human cancer and three non-malignant cell lines, as well as in zebrafish xenograft models. QAL333 exhibited potent and selective cytotoxicity in colorectal cancer, with IC₅₀ values in the low micromolar range (7.8 µM in SW620), while showing limited efficacy in triple-negative breast cancer (IC₅₀ > 90 µM). In vivo, QAL333 (20 µM) significantly suppressed tumor growth in SW620 xenografts, whereas MDA-MB-231 xenografts showed minimal response. Transcriptomic profiling revealed that QAL333 downregulated cell cycle regulators (CDK1, CDC25A, CCND2, CCNE2, and PCNA) and suppressed PI3K–Akt and p53 signaling in sensitive cells, while resistant cells activated NF-κB–driven stress pathways. By integrating drug-induced transcriptomic data with machine learning, we developed a composite pathway score derived from three KEGG pathways: Metabolic pathways, Proteoglycans in cancer, and Cytoskeleton in muscle cells. This score correlated strongly with IC₅₀ values across 16 lines (Pearson correlation coefficient = 0.737, p = 0.0011), providing a mechanistically interpretable biomarker for predicting sensitivity. These findings support QAL333 as a selective anticancer agent with translational biomarker potential. Secondary observations also indicated that QAL333 interferes with quorum sensing pathways, broadening its pharmacological relevance beyond the central focus of this study.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-36430-4.

Introduction

Despite substantial progress in oncology, tumor heterogeneity and resistance continue to limit the efficacy of conventional chemotherapies. There is a critical need for small molecules with lineage-specific activity and mechanistically interpretable biomarkers that can guide patient stratification. Although numerous transcriptomic and pharmacogenomic resources have contributed substantially to anticancer drug discovery, many current prediction frameworks rely primarily on baseline gene expression^1–4. While informative, such static measurements may provide limited mechanistic resolution and do not always explain lineage-dependent sensitivity or resistance to newly developed compounds. Moreover, baseline-only approaches may not fully account for transcriptional changes induced by drug exposure, which are increasingly recognized as important contributors to selective cytotoxicity.

Isoindolin-1-one derivatives have attracted attention as a versatile chemotype in anticancer drug development. Several studies have reported that compounds in this family exert cytotoxic effects through inhibition of mitotic kinases such as PLK1, blockade of ERK1/2 signaling, and suppression of PARP-1 activity⁵. These mechanisms have been implicated in the treatment of breast, colorectal, and ovarian cancers, reinforcing the therapeutic relevance of this scaffold. Building on this foundation, we synthesized QAL333, a structurally distinct isoindolin-1-one derivative with potential antitumor activity (Supplementary Fig. S1).

Our data show that QAL333 exerts potent and selective cytotoxicity in colorectal cancer cells, with limited efficacy in triple-negative breast cancer, indicating lineage-specific vulnerabilities. Importantly, its efficacy was confirmed in zebrafish xenograft models, which provide advantages such as optical transparency⁶, rapid development⁷, and genetic similarity to humans⁸, enabling real-time monitoring of tumor progression. Mechanistically, transcriptomic profiling revealed that QAL333 suppresses cell cycle regulators and the PI3K–Akt and p53 signaling pathways in sensitive lines.

Given that many current predictive models rely on baseline gene expression from public datasets and focus primarily on predictive performance rather than mechanistic insight^1–4, we developed an integrative framework combining drug-induced transcriptomic profiling with machine learning–based feature selection and pathway-level modeling. This approach facilitated the identification of gene pairs and convergent pathways associated with sensitivity to QAL333, ultimately leading to the development of a composite pathway score that strongly correlates with IC₅₀ values. Such pathway-level biomarkers not only enhance the prediction of drug sensitivity but also provide mechanistic insight into lineage-specific vulnerabilities.

Taken together, this study introduces QAL333 as a novel isoindolin-1-one derivative with lineage-selective antitumor activity and presents a transcriptome-guided biomarker framework for predicting drug sensitivity. Secondary analyses also indicated that QAL333 interferes with quorum sensing–related processes^9–18, a property with potential microbiological implications beyond the primary focus of this work. To the best of our knowledge, this study is among the first to integrate novel compound screening with drug-induced transcriptome-guided prediction and in vivo efficacy validation within a single analytical framework, as summarized and compared with previous studies in Supplementary Table S1^19–25. This combined approach provides not only predictive accuracy but also mechanistic insight, offering a framework toward personalized anticancer therapy based on functional transcriptomic responses.

Results

Differential cytotoxicity of QAL333 in human cancer cell lines

To evaluate the antitumor activity of QAL333, we performed cytotoxicity assays in 15 human cancer cell lines (SW620, SW480, HCT-15, LoVo, DLD-1, Caco-2, MDA-MB-231, MCF7, T-47D, H460, A549, HCT 116, AGS, HeLa, and SK-MEL-2) and three non-malignant lines (HaCaT, IMR-90, and HEK-293). QAL333 induced a dose- and time-dependent decreases in viability in most cancer cell lines, while non-malignant HEK-293 cells were comparatively resistant (Fig. 1A–R). Sensitive colorectal lines such as SW620 (Fig. 1L) exhibited pronounced reductions in viability (IC₅₀ = 7.8 µM), whereas MDA-MB-231 breast cancer cells (Fig. 1G) exhibited relative resistance (IC₅₀ > 90 µM). Non-malignant HaCaT keratinocytes and IMR-90 lung fibroblasts also displayed notable sensitivity, consistent with their relatively low IC₅₀ values (Fig. 1Q, R). These findings indicate that QAL333’s cytotoxicity cannot be explained simply by malignant versus non-malignant status.

Fig. 1.

Dose- and time-dependent effects of QAL333 on cell viability. (A–R) QAL333 decreased viability of 15 cancer and three non-malignant cell lines in a dose- and time-dependent manner. Colorectal cancer cells were most sensitive, while MDA-MB-231 and HEK-293 were resistant. Data are shown as mean ± SD (n = 3). Welch’s t-test was used to compare each treatment with its respective control at 24 h and 48 h. *p < 0.05, **p < 0.01, ***p < 0.001.

Importantly, proliferation rate alone did not correlate with drug sensitivity. HEK-293 cells proliferated robustly yet were resistant, whereas slower-growing colorectal lines such as Caco-2 and DLD-1 were more susceptible (Supplementary Fig. S2A). Instead, lineage-specific features such as basal oxidative stress or vulnerabilities in stress-response pathways may play a central role (Supplementary Fig. S2B). These results suggest that transcriptional programs governing stress responses and metabolism, rather than overall growth capacity, may determine QAL333 responsiveness.

Tumor growth Inhibition in zebrafish xenograft models

The in vivo antitumor effect of QAL333 was assessed using zebrafish xenografts implanted with SW620 (colorectal cancer) or MDA-MB-231 (triple-negative breast cancer) cells. In the xenograft model, tumor growth was quantified as relative tumor size, calculated as the ratio of tumor area at 4 days post-injection (4 dpi) to that at 1 day post-injection (1 dpi). In the SW620 model, QAL333 suppressed tumor growth at all tested concentrations (0.2, 2, and 20 µM; p < 0.05), with the highest dose (20 µM) reducing relative tumor size (4 dpi/1 dpi, %) from 342.6% (95% CI: 280.7–404.6) in the control group to 252.7% (95% CI: 196.2–309.2), corresponding to 73.8% ± 13.3% of the untreated control at 4 dpi following treatment initiation immediately after injection (Fig. 2A). At 20 µM, QAL333 exhibited antitumor activity in the zebrafish model, showing a tumor-suppressive trend in the same direction as that observed with FOLFIRI (0.08 mM Irinotecan + 4.2mM 5-fluorouracil), a standard chemotherapy for colorectal cancer. Representative tumor fluorescence images illustrate these effects (Fig. 3A, B). In contrast, QAL333 displayed limited efficacy in the MDA-MB-231 xenograft model, showing no significant reduction in tumor volume (p > 0.05), while paclitaxel (PTX, 25 nM) significantly suppressed tumor growth (p < 0.01), highlighting that QAL333 activity was minimal in this breast cancer model (Figs. 2B and 3C and D). Taken together, these results indicate that QAL333 exerts potent in vivo antitumor effects in colorectal cancer xenografts but has limited efficacy in triple-negative breast cancer, emphasizing lineage-dependent responsiveness.

Fig. 2.

In vivo efficacy of QAL333 in zebrafish xenograft models. (A, B) SW620 xenografts exhibited a significant reduction in tumor growth following QAL333 treatment (0.2, 2, and 20 µM), showing a tumor-suppressive trend in the same direction as that observed with FOLFIRI (0.08 mM Irinotecan + 4.2mM 5-fluorouracil). (C, D) QAL333 treatment (0.2, 2, and 20 µM) had no significant effect on tumor growth in MDA-MB-231 xenografts, whereas paclitaxel (25 nM) significantly reduced tumor size (**p < 0.01). Bar graphs represent relative change in tumor size at 4 days post-injection (dpi) compared with 1 dpi, expressed as a percentage (4 dpi/1 dpi, %). Treatment was administered immediately after SW620 or MDA-MB-231 cell injection. Data are presented as mean with 95% confidence intervals (n = 5 per group). Welch’s t-test was used to compare each treatment with control. *p < 0.05, **p < 0.01, ***p < 0.001.

Fig. 3.

Representative fluorescence images of zebrafish xenografts. (A) SW620 cells (PVS injection) showed tumor reduction after QAL333 (0.2, 2, and 20 µM) or FOLFIRI (0.08 mM Irinotecan + 4.2 mM 5-fluorouracil) treatment (n = 5 per group). (B) MDA-MB-231 xenografts (yolk sac injection) exhibited modest tumor shrinkage with QAL333 (0.2, 2, and 20 µM), while paclitaxel (25 nM) induced substantial regression (n = 5 per group). Images were acquired at 1 and 4 dpi (days post-injection). Treatment was initiated immediately following injection. Tumor fluorescence is shown in red, and the scale bar located at the bottom left of each panel is shown in green.

Transcriptomic profiling reveals mechanistic basis of QAL333 activity

To explore mechanisms underlying QAL333 activity, we performed RNA-seq in SW620 and MDA-MB-231 cells treated with 15 µM QAL333 for 24 h. In SW620 cells, 512 genes were differentially expressed (|FC| ≥ 1.5, p < 0.05), with significant downregulation of cell cycle regulators including CDK1, CDC25A, CCND2, CCNE2, and PCNA (Fig. 4A; Supplementary Fig. S3). KEGG mapping highlighted suppression of cell cycle (hsa04110), PI3K–Akt (hsa04151), and p53 signaling (hsa04115) pathways (Fig. 4A).

Fig. 4.

QAL333 modulates KEGG pathways in cancer cells. (A) Downregulation of Cell cycle, Metabolic pathways, PI3K–Akt signaling pathway, and p53 signaling pathway in SW620. (B) Modulation of TNF signaling pathway, NF-κB signaling pathway, IL-17 signaling pathway, and Necroptosis in MDA-MB-231. Benjamini-Hochberg procedure was applied to correct multiple testing, and adjusted p-values are shown as bar representing–log10(p-value). Adjusted p-values < 0.05 were considered statistically significant.

In MDA-MB-231 cells, 397 genes were significantly altered (|FC| ≥ 1.5, p < 0.05), with enrichment in TNF (hsa04668), NF-κB (hsa04064), NOD-like receptor (hsa04621), and IL-17 signaling (hsa04657) pathways (Fig. 4B). These profiles suggest that QAL333 triggers inflammatory and stress-response programs in resistant cells, consistent with the observed modest cytotoxicity.

Hierarchical clustering of global transcriptomic changes (genes with |FC|≥ 2, log₂-normalized expression ≥ 3.0, p < 0.05) further emphasized lineage-specific responses: MDA-MB-231 and HEK-293 clustered together, while SW620 and IMR-90 fibroblasts formed distinct branches (Fig. 5A). GO analysis of transcripts with |FC|≥ 2, log₂-normalized expression ≥ 4.0, p < 0.05 exhibited enrichment of oxidative stress-response genes and necrosis-associated regulators in SW620 cells (Fig. 5B, C), whereas apoptosis and autophagy markers showed minimal changes (data not shown). This supports a model in which QAL333 induces non-apoptotic, redox-associated death pathways in sensitive colorectal cells.

Fig. 5.

QAL333-induced transcriptomic changes in cancer vs. non-malignant cells. (A) Hierarchical clustering of DEGs (genes with |FC|≥ 2, log₂-normalized expression ≥ 3.0, p < 0.05): MDA-MB-231 and HEK-293 clustered together, while SW620 and IMR-90 fibroblasts were distinct. (B) Oxidative stress–related genes and (C) necrosis-related genes were altered, consistent with non-apoptotic death in SW620.

Predictive modeling of QAL333 sensitivity using mRNA expression profiles

To connect transcriptional signatures with pharmacologic responses, IC₅₀ values from 15 cell lines (SW620, SW480, HCT-15, LoVo, Caco-2, DLD-1, MCF7, T-47D, HCT 116, MDA-MB-231, HeLa, AGS, SK-MEL-2, HEK-293, and H460) were integrated with their baseline gene expression data obtained either from DepMap or from our experimental dataset.

To define transcriptomic features relevant to drug sensitivity, candidate genes were first selected based on treatment-induced expression changes in two representative cell lines (SW620 and MDA-MB-231), using predefined fold-change, expression, and statistical significance criteria. From this candidate set, genes showing the strongest correlations between baseline expression and IC₅₀ values across a cell line panel were prioritized and used as input features for model training.

Using this feature selection strategy, Spearman correlation analysis between the IC₅₀ values and gene expression identified 100 genes most strongly associated with drug sensitivity. From this gene set, subsequent L1-regularized regression with polynomial feature expansion selected 19 gene pairs as predictive features, suggesting that these genes are closely linked to QAL333 sensitivity (Supplementary Fig. S4). The resulting model achieved robust performance (precision = 0.8571, recall = 0.8571, F1 score = 0.8571, accuracy = 0.8667, AUC = 0.8571), supporting both predictive reliability and biological interpretability. Model performance was evaluated using cross-validation to minimize overfitting and ensure robustness across cell lines. Detailed descriptions of feature selection thresholds, weighting strategies, and validation procedures are provided in the Supplementary Information.

To further identify significantly enriched pathways, we selected the top 300 genes most strongly correlated with drug sensitivity based on Spearman analysis. To enhance the biological interpretability, gene expression for each pathway was summarized into a pathway-level score (Table 1; Supplementary Tables S3A, S3B, and S3C). Briefly, for each pathway, normalized gene expression values were aggregated into pathway scores using weighted summation, enabling dimensionality reduction while preserving biological interpretability.

Table 1.

Correlation of pathway scores with IC₅₀ values across cell lines.

Cell line	IC50 (µM)	Top 100 total score	Top 300 total score	CDS	ODS	Proteoglycans in cancer	Focal adhesion	Metabolic pathways	Cytoskeleton in muscle cells	Apoptosis pathway	Weighted composite pathway score
SW620	7.84	205.64	484.44	14.25	9.63	17.63	15.67	43.55	11.13	7.39	23.13
SW480	8.34	180.39	419.06	15.97	9.82	17.52	14.64	33.56	9.77	8.70	19.62
Caco-2	26.11	179.95	450.23	16.38	11.39	21.21	15.13	44.45	9.21	9.87	23.98
DLD-1	21.96	207.46	495.78	18.69	13.55	20.08	15.31	41.72	8.28	9.66	22.44
LoVo	26.67	191.27	466.04	16.76	12.34	21.55	17.88	41.91	10.41	9.20	23.76
HCT-15	9.29	204.64	485.61	17.56	12.83	19.18	14.99	39.96	8.33	8.73	21.62
MCF7	23.03	177.16	435.20	11.12	9.19	19.27	10.13	43.57	5.16	9.26	21.62
T-47D	12.07	208.58	483.41	11.68	10.38	13.07	8.75	40.04	7.25	8.09	19.12
HCT 116	28.11	199.33	486.02	18.99	14.42	23.08	20.79	42.34	10.58	9.18	24.48
MDA-MB-231	96.41	198.81	496.99	19.02	12.60	25.73	21.86	45.28	12.53	10.03	26.97
HeLa	25.11	185.21	458.23	16.06	13.28	21.33	17.47	44.24	10.90	9.14	24.55
AGS	5.58	193.03	478.23	10.12	7.57	11.73	13.17	40.13	10.20	5.73	19.72
SK-MEL-2	18.35	206.79	497.44	15.22	10.66	24.34	17.90	41.95	11.53	10.11	25.14
HEK-293	56.75	225.41	526.04	17.20	14.57	21.80	17.17	48.74	12.95	9.65	26.78
H460	27.46	177.48	422.57	18.68	11.08	20.06	16.99	43.16	10.69	11.31	23.71
A549	40.83	191.99	462.65	19.67	15.66	24.79	19.95	49.57	11.58	10.54	27.60
PCC	–	0.171	0.328	0.506	0.496	0.663	0.611	0.601	0.510	0.501	0.737
p-value	–	0.526	0.215	0.0456	0.0509	0.00512	0.0119	0.0139	0.0435	0.0482	0.0011

Pathway scores were calculated for the top 100 or 300 genes correlated with IC₅₀ values, for gene sets related to pro-death (Cell death Score, CDS) or oxidative defense (Oxidative defense Score, ODS), and for the top five KEGG pathways enriched among the top 300 genes. IC₅₀ values were obtained from 24 h dose–response assays. Each pathway score was calculated as weighted sums of normalized expression values using ridge regression-derived importance scores. The CDS and ODS represent pathway scores based on pro-death and oxidative defense genes, respectively. Complete gene lists with normalized expression values and the importance-score weights for CDS, ODS, and the top five KEGG pathways are provided in supplementary tables S3A and S3B. The weighted composite pathway score was calculated as a weighted sum of the pathway scores for proteoglycans in cancer, metabolic pathways, and cytoskeleton in muscle cells, as schematically illustrated in supplementary Fig. S6.

Significance of bold values in Table 1:Bold values indicate statistically meaningful correlations, defined as a Pearson correlation coefficient (PCC) greater than 0.5 combined with a p-value less than 0.05

Using this approach, the pathway scores for the Proteoglycans in cancer, Metabolic pathways, and Cytoskeleton in muscle cells were among the top five KEGG pathways associated with QAL333 sensitivity, based on their strength of correlations with IC₅₀ values (Table 1). To capture the coordinated contribution of multiple biological pathways, a weighted combination of pathway score was applied (Metabolic pathways, 0.3; Proteoglycans in cancer, 0.35; Cytoskeleton in muscle cells, 0.35, Σ weights = 1; Supplementary Fig. S6). The resulting weighted composite pathway score exhibited a strong and statistically significant positive correlation with IC₅₀ values (PCC = 0.737, p = 0.0011), with a 95% confidence interval of 0.38–0.90 (Table 1; Fig. 6A and B). Together, these findings suggest that QAL333 sensitivity is determined not by individual genes but by the integrated activity of metabolic, extracellular matrix (ECM), and cytoskeletal networks.

Fig. 6.

Weighted composite pathway score predicts QAL333 sensitivity. (A) Scatter plot showing the correlation between weighted composite pathway scores and IC₅₀ values across 16 cell lines (PCC = 0.737, p = 0.0011). (B) Distribution of weighted composite scores across 16 cell lines, distinguishing sensitive from resistant lines.

Secondary pharmacological effects: quorum sensing inhibition and biofilm disruption

In addition to its antitumor properties, QAL333 displayed quorum sensing inhibitory activity in Porphyromonas gingivalis. At sub-micromolar concentrations, it suppressed biofilm formation by 57.4% ± 4.6% (Supplementary Fig. S5A) and reduced the expression of gingipain proteases (RgpA, RgpB, Kgp), key virulence factors regulated by QS signaling pathways (Supplementary Fig. S5B). Moreover, QAL333 attenuated NLRP3 inflammasome activation in bone marrow–derived macrophages at 1–10 µM (Supplementary Fig. S5C). While these findings broaden QAL333’s pharmacological relevance, they are presented as secondary observations in this work.

Discussion

This study identifies QAL333 as a novel isoindolin-1-one derivative with lineage-selective antitumor activity and provides a mechanistically interpretable biomarker framework for predicting sensitivity. QAL333 exhibited potent cytotoxicity against colorectal cancer cells, with IC₅₀ values in the low micromolar range (e.g., 7.8 µM in SW620) and showed a tumor suppressive effect in the SW620 zebrafish xenograft model that was consistent in direction with that observed for FOLFIRI (0.08 mM Irinotecan + 4.2mM 5-fluorouracil). In contrast, triple-negative breast cancer cells (MDA-MB-231) were markedly resistant in vitro (IC₅₀ > 90 µM) and in the MDA-MB-231 xenograft model, underscoring lineage-dependent efficacy. Importantly, non-malignant cells also exhibited heterogeneous responses: IMR-90 fibroblasts and HaCaT keratinocytes were relatively sensitive (IC₅₀ ≈ 5–6 µM), whereas HEK-293 cells remained resistant despite their rapid proliferation. These findings emphasize that QAL333 responsiveness is not determined by malignant status or growth kinetics, but by intrinsic lineage-specific vulnerabilities.

Transcriptomic profiling suggests that QAL333 acts by suppressing fundamental regulators of proliferation and survival. In SW620 cells, downregulation of CDK1, CDC25A, CCND2, CCNE2, and PCNA was consistent with inhibition of cell cycle, while concurrent suppression of PI3K–Akt and p53 pathways suggested multi-pronged interference with tumor-promoting signaling. GO analysis further implicated oxidative stress and necrosis-related gene programs, consistent with a non-apoptotic mode of death. In MDA-MB-231, by contrast, QAL333 treatment activated TNF and NF-κB signaling as well as NOD-like receptor pathways, highlighting adaptive stress responses that may confer resistance. Together, these lineage-specific patterns indicate that QAL333 may exploits vulnerabilities in redox balance and metabolic regulation, while resistant tumors activate compensatory inflammatory programs. The isoindolin-1-one scaffold of QAL333 resembles structures of known mitotic kinase inhibitors, supporting its potential development as a mechanistically relevant anticancer agent.

In line with the cell viability data, hierarchical clustering also exhibited minimal overlap between the two non-malignant cell lines, HEK-293 and IMR90, which displayed varying degree of sensitivity to QAL333, with negligible changes in cell cycle, oxidative stress, or necrosis pathways after the treatment. This subdued transcriptomic response is consistent with the resistance of HEK-293 cells even at 40 µM QAL333, underscoring lineage-specific selectivity.

Our predictive modeling approach advances biomarker development by integrating drug-induced transcriptomic responses with pathway-level analysis. To facilitate interpretation, we hypothesized that QAL333 sensitivity could be predicted more accurately by aggregating gene-level signals into biologically meaningful pathways. Specifically, the top 300 genes, ranked by their absolute Spearman correlation coefficients with IC₅₀ values were assigned to KEGG pathways. For each pathway, a pathway score was calculated by multiplying log₂-normalized expression values by the corresponding ridge regression-derived importance scores, thereby reflecting the relative contribution of each gene. Table 1 and Supplementary Tables S2 and S3 summarize the pathway assignments, coefficients, expression values, and score contributions. Unlike conventional single-gene or baseline-expression markers, this framework identified three convergent pathways—Metabolic pathways, Proteoglycans in cancer, and Cytoskeleton in muscle cells—as dominant determinants of QAL333 sensitivity. The weighted composite pathway score derived from these networks correlated strongly with IC₅₀ values across 16 lines (PCC = 0.737, p = 0.0011), outperforming single-pathway approaches (Table 1). These results indicate that QAL333 sensitivity is best explained not by individual genes but by the integrated activity of coordinated transcriptional programs. In resistant cell lines, enrichment of the three pathways indicates that these biological processes are functionally perturbed under baseline conditions, potentially contributing to QAL333 resistance. Conversely, their relative instability or lower baseline activity in sensitive lines implies that QAL333 may exert its cytotoxic effects, at least in part, by perturbing these same pathways. For instance, alterations of Proteoglycan in cancer pathway in resistant cell lines such as MDA-MB-231 aligns with previous reports that ECM remodeling promotes drug resistance by modulating cell-matrix interactions and growth factor signaling. Since these findings are based on correlation analyses, further studies are warranted to determine whether these pathways are activated or suppressed through drug perturbation or functional validation approaches. Therefore, this composite pathway score framework may provide both mechanistic insight and predictive power, advancing beyond single-gene biomarkers and supporting rational combination strategies to overcome resistance.

Notably, the top-ranked gene pairs identified by the Lasso model (Supplementary Fig. S4, Supplementary Table S3D) highlight biological themes that reinforce our transcriptomic findings observed in our experiments. For example, ECHDC2–CTSL and GGH–MAP3K13 pairs link metabolic stress to apoptotic or necrotic regulators^26–30, whereas SARAF–MAP3K13 suggests crosstalk between ER–Ca²⁺ homeostasis and stress-activated kinases^29–31. The BLOC1S1–DNAJB11 pair reflects ER stress adaptation^32,33, and EXOC6–CEP70 connects vesicle trafficking with spindle dynamics^34,35, consistent with the downregulation of cell cycle pathway in SW620 cells. In addition, multiple APOBEC3-related pairs (APOBEC3B–ZNF594, ZNF429–APOBEC3F, APOBEC3F–NUDT4B) imply mutagenesis and immune signaling^36–40, in line with NF-κB/NOD-like pathway activation observed in resistant MDA-MB-231 cells. Collectively, these predictive features reinforce the mechanistic themes identified in our cytotoxicity and transcriptomic analyses—namely, that QAL333 sensitivity is shaped by lineage-specific vulnerabilities in redox balance, cytoskeletal and ECM integrity, and adaptive transcriptional programs.

Clinically, these findings suggest that QAL333 may be most effective against tumors characterized by high basal oxidative stress, disrupted metabolic homeostasis, and destabilized cytoskeletal or ECM architecture. Such molecular signatures are frequently observed in subsets of colorectal cancer and could be captured using patient-derived organoids or clinical transcriptomic datasets. Future work should validate the composite pathway score in these models and establish clinically relevant thresholds to stratify patients into responder and non-responder groups. In parallel, combination strategies targeting ECM remodeling or NF-κB signaling may overcome resistance observed in breast and other tumor types.

While the anticancer properties of QAL333 constitute the primary focus of this study, we additionally report exploratory quorum-sensing (QS)-related observations that do not form the basis of the study’s primary antitumor conclusions. Specifically, QAL333 exhibited QS inhibitory activity, reducing Porphyromonas gingivalis biofilm formation by ~ 57% and attenuating gingipain expression at sub-micromolar concentrations. In addition, QAL333 suppressed NLRP3 inflammasome activation in macrophages at 1–10 µM.

Quorum sensing (QS) is a bacterial communication mechanism that enables microorganisms to coordinate collective behaviors by signaling molecules in response to cell density changes⁹. Emerging evidence has implicated QS-associated pathways in modulating inflammation and immune evasion within the tumor microenvironment¹³. Quorum sensing inhibitors (QSIs) have therefore gained attention as potential therapeutics for both infectious diseases and cancer.

In the present study, however, these QS-related effects should be interpreted as hypothesis-generating observations, rather than as mechanistic contributors to QAL333’s lineage-specific antitumor activity. Accordingly, although these findings suggest that QAL333 may possess broader biological activities beyond direct tumor cell targeting, no causal relationship between quorum sensing modulation and antitumor efficacy is implied. Further investigation using dedicated microbiome-tumor interaction models will be required to determine whether such effects have functional relevance in cancer-associated settings.

This study has several limitations that should be acknowledged. First, in vivo validation was limited to zebrafish xenograft models, which enable rapid lineage-specific assessment but do not fully reflect the complexity of mammalian tumor microenvironments. Second, transcriptomic profiling was performed at a single post-treatment time point, potentially overlooking early or late transcriptional changes relevant to drug sensitivity. Third, although the multi-lineage panel encompassed colorectal, breast, and non-malignant cell lines, inclusion of additional cancer subtypes will be necessary to generalize the pathway-level determinants identified in this study. Finally, model performance was assessed through cross-validation within the same dataset, as no independent QAL333 dataset was available. Validation using external datasets or orthogonal experimental systems will therefore be required to further substantiate and refine the proposed pathway-weighted framework.

Future investigations may extend these findings in several directions. Validation of QAL333 efficacy in murine xenograft or patient-derived models would help clarify its translational relevance. Time-resolved transcriptomic profiling could uncover dynamic regulatory programs underlying early stress responses and cell death commitment. In addition, CRISPR-based perturbation of key pathways, including oxidative defense, cytoskeletal remodeling, and proteoglycan signaling, may enable direct assessment of their causal roles in sensitivity. Finally, systematic exploration of combination therapies guided by the composite sensitivity score could identify synergistic partners capable of overcoming intrinsic resistance.

In conclusion, these findings suggest that QAL333 selectively disrupts cell cycle progression, survival signaling, and oxidative defense in vulnerable tumor lineages. Its strong in vivo efficacy in colorectal xenografts, combined with the development of a mechanistically interpretable predictive framework, positions QAL333 as a promising candidate for translational development. Further validation in organoids, advanced animal models, and patient cohorts will be essential to confirm its therapeutic window and establish biomarker-guided clinical application.

Materials and methods

Cell culture and materials

SW620, SW480, HCT-15, LoVo, DLD-1, Caco-2, MDA-MB-231, MCF7, T-47D, H460, A549, HCT 116, AGS, HeLa, SK-MEL-2, HaCaT, IMR-90 and HEK-293 cell lines were obtained from the Korean Cell Line Bank (KCLB, Seoul, Korea). Cells were maintained in either high-glucose DMEM or RPMI-1640 medium (Cytiva, Marlborough, MA, USA) supplemented with 10% fetal bovine serum (FBS) and antibiotics (Gibco, Waltham, MA, USA) at 37 °C in a humidified incubator with 5% CO₂. IMR-90 cells were cultured in MEM medium (Cytiva, Marlborough, MA, USA) under the same incubation. All cell lines were obtained from authenticated sources (KCLB), tested for mycoplasma contamination prior to use, and handled according to standard cell culture quality-control procedures, with experiments performed within a limited passage range (e.g., passages 5–20).

Reagents and compounds

QAL333 was synthesized in-house and fully characterized by ¹H NMR, IR spectroscopy, LC-MS, X-ray powder diffraction, and melting point analysis, confirming its structural identity and high chemical purity. LC-MS analysis revealed a single dominant chromatographic peak corresponding to the expected molecular ion, with no detectable major impurities. The compound exhibited a sharp melting point (187.9–190.6 °C), low water content (0.05%, Karl Fischer titration), and minimal residual palladium (8 ppm, ICP-OES), supporting its suitability as an analytical-grade research compound. A 10 mM stock solution in DMSO, aliquoted, and stored at − 20 °C to ensure stability and avoid repeated freeze-thaw cycles. DPD (4,5-dihydroxy-2,3-pentanedione) was also synthesized in-house and used for quorum sensing assays. Paclitaxel (PHR1803, Sigma-Aldrich) and the FOLFIRI regimen—consisting of 0.8 mM irinotecan (I1406, Sigma-Aldrich) and 4.2 mM 5-fluorouracil (5-FU, F6625, Sigma-Aldrich)—were used as positive controls.

Cell viability assay

Cells were seeded at 1 × 10⁴ cells per well in 96-well plates. Cell viability was measured using the CellTiter 96^® AQueous One Solution Cell Proliferation Assay (Promega). After 24–48 h of QAL333 treatment, 10 µL of reagent was added to each well, and cells were incubated at 37 °C for 3 h. Absorbance was measured at 490 nm using a SpectraMax 190 microplate reader (Molecular Devices, Sunnyvale, CA, USA). IC₅₀ values were determined using GraphPad Prism 10 software (GraphPad Software, San Diego, CA, USA).

Zebrafish xenograft assay

Wild-type AB strain zebrafish (Danio rerio) were obtained from Zefit Inc. (Daegu, Korea) and maintained under standard laboratory conditions. Embryos were obtained by natural spawning and raised in E3 medium (5 mM NaCl, 0.17 mM KCl, 0.33 mM CaCl₂, 0.33 mM MgSO₄; pH 7.0–7.2) at 28.5 °C under a 14 h light/10 h dark photoperiod. Zebrafish embryos at 48 h post-fertilization (hpf) were dechorionated and anesthetized in 0.003% tricaine. SW620 and MDA-MB-231 cells were labeled using CellTracker™ CM-DiI (Invitrogen, Thermo Fisher Scientific, USA) prior to injection. Approximately 200 DiI-labeled SW620 cells were microinjected into the perivitelline space (PVS), a standard site for zebrafish xenografts, whereas the more invasive MDA-MB-231 cells were injected into the yolk sac of each zebrafish embryo to ensure more stable and reproducible engraftment⁴¹. Based on standard zebrafish screening practice (10 µM as a reference working concentration)⁴², a dose range of 0.2–20 µM was tested. Injected embryos were incubated at 34 °C in 6-well plates and treated immediately after injection with QAL333 (0.2, 2, or 20 µM) or vehicle (0.16% DMSO). In zebrafish xenograft assays, embryos are commonly maintained at 32–35 °C to support human tumor cell proliferation; accordingly, 34 °C was used to balance tumor growth and embryo viability⁴³. Tumor fluorescence was quantified at 4 days post-injection (dpi) using ImageJ software based on red channel intensity. Immediately after the final imaging session, zebrafish were anesthetized and transferred to ice-cold E3 medium for more than 20 min, and death was confirmed by the absence of cardiac activity under a stereomicroscope. No significant developmental abnormalities or malformations were observed upon monitoring at 24-hour intervals. All animal experiments were conducted in accordance with the institutional guidelines for animal research of Zefit Inc. (Daegu, Korea). Experimental protocols involving zebrafish were approved by the Institutional Animal Care and Use Committee (IACUC) of Zefit Inc. (Approval No. ZEFIT-IACUC-23030601-0001). All procedures complied with the relevant institutional and national guidelines and regulations, and the study was designed, performed, and reported in accordance with the ARRIVE guidelines.

RNA extraction and mRNA sequencing

SW620, MDA-MB-231, IMR-90, and HEK-293 cells were treated with 15 µM QAL333 or vehicle (DMSO) for 24 h, a concentration chosen to induce a measurable transcriptional response without extensive cytotoxicity. Total RNA was extracted using the Trizol reagent (Invitrogen), and RNA integrity was evaluated with the Agilent TapeStation 4000 system (Agilent Technologies, Amstelveen, The Netherlands). RNA concentrations were determined using a ND-2000 spectrophotometer (Thermo Scientific, DE, USA). For library construction, the QuantSeq 3’ mRNA-Seq Library Prep Kit (Lexogen, Austria) was utilized in accordance with the manufacturer’s instructions. In this process, total RNA was hybridized with an oligo-dT primer containing a partial Illumina adapter sequence at the 5′ end, which served to initiate first-strand synthesis via reverse transcription. Sequencing was carried out on the Illumina NextSeq 550 platform using single-end 75 bp reads. QuantSeq 3′ mRNA-Seq reads were mapped to reference sequences using Bowtie2⁴⁴. Index files were generated based on either full genome assemblies or representative transcript sets, allowing for alignment to both genomic and transcriptomic references. The alignment outputs were subsequently utilized for transcript assembly, quantification of gene expression levels, and identification of differentially expressed genes (DEGs). To determine DEGs, both uniquely and multiply aligned reads were counted, and coverage data were assessed using Bedtools⁴⁵. Raw read counts were normalized employing the trimmed mean of M values in combination with counts per million transformation via the EdgeR package in the R statistical environment⁴⁶. Functional annotation and gene categorization were carried out through the DAVID Bioinformatics Resource (https://david.ncifcrf.gov) and NCBI Medline (https://www.ncbi.nlm.nih.gov). Data analysis and expression profiling were conducted using ExDEGA software (Ebiogen Inc., Korea). Pathways and Gene Ontology (GO) enrichment analyses were conducted using DAVID and KEGG Mapper⁴⁷.

Bioinformatic modeling of drug sensitivity

We developed a machine learning classifier to predict drug sensitivity in previously untested cell lines, using baseline transcriptomic profiles and the IC₅₀ values from 15 cell lines. Transcriptomic data for 14 cell lines (SW620, SW480, HCT-15, LoVo, Caco-2, DLD-1, MCF7, T-47D, HCT 116, MDA-MB-231, HeLa, AGS, SK-MEL-2, H460) were obtained from DepMap (https://depmap.org/portal), while those for HEK-293 cells, which were not available in DepMap, were experimentally generated as described in “RNA Extraction and mRNA Sequencing” section above. For modeling, drug sensitivity was defined by dichotomizing the IC₅₀ values at the median: cell lines with IC₅₀ values below the median were classified as sensitive, and those above the median as resistant. To identify transcriptomic features relevant to drug sensitivity, we first defined 1,010 candidate genes showing an absolute fold change (FC) ≥ 1.2, p < 0.05, and log₂-normalized expression ≥ 2.0 after QAL333 treatment in two test lines, SW620 and MDA-MB-231 cells. From this set, the top 100 genes with the strongest Spearman correlations with IC₅₀ values across the cell line panel were selected as input features to train a baseline-expression–based classifier (14 DepMap lines and HEK-293; total n = 15).

To gain pathway-level insight, the 300 genes most strongly correlated with IC₅₀ values were analyzed using KEGG Mapper for pathway enrichment⁴⁷. For each enriched pathway, a pathway score was computed by summing the products of log₂-normalized expression values and their corresponding ridge regression–derived importance scores (Table 1; Supplementary Tables S3A and S3B), reflecting the weighted contribution of individual genes to pathway activity. Multiple combinations of these pathway scores were then evaluated by fitting regression models to IC₅₀ values, allowing pathway-level weights to be re-estimated for each combination. Among all tested combinations, the integration of Metabolic pathways, Proteoglycans in cancer, and Cytoskeleton in muscle cells achieved the highest Pearson correlation coefficient (PCC) with IC₅₀ values and was therefore selected. Pathway-level weights derived from this model were subsequently applied to compute the final weighted composite pathway score for each cell line. The schematic illustration of the workflow for evaluating combinations to maximize the PCC and for optimizing weighting strategies to enhance correlation with IC₅₀ values is described in Supplementary Fig. S6. The inclusion of one in-house cell line was evaluated for consistency with DepMap-derived samples using pathway-wise z-score standardization, leave-one-out analyses, and rank-based comparisons, which showed no evidence of global bias or disproportionate influence.

Given the limited sample size (n = 15), we employed a logistic regression model with second-order polynomial feature expansion and L1 regularization. This approach balances model complexity and interpretability while minimizing overfitting. Model performance was evaluated using leave-one-out cross-validation (LOOCV), which showed robust predictive ability with an accuracy of 0.87 and an area under the ROC curve (AUC) of 0.86.

Model implementation details

The transcriptome-based predictor was implemented using logistic regression with L1 regularization and a second-order polynomial feature expansion. Given the limited sample size (n = 15), this approach was selected to balance model expressiveness and interpretability while controlling overfitting. Feature selection was performed using Spearman correlation–based ranking, and all feature-selection steps were conducted within each training fold only. Model performance was evaluated using leave-one-out cross-validation (LOOCV). The logistic regression model was trained using the liblinear solver with a maximum of 1000 iterations and a fixed random seed (42). Prior to modeling, gene expression values were log2-normalized; missing values were imputed using the mean; polynomial features were generated; and all features were standardized using z-score normalization. All analyses were implemented in Python using the scikit-learn library and executed in a Google Colaboratory environment.

Quorum sensing and biofilm assays

Porphyromonas gingivalis (ATCC 33277) was cultured anaerobically (10% H₂, 10% CO₂, and 80% N₂) at 37 °C in brain heart infusion (BHI) broth supplemented with vitamin K (0.2 mg/mL) and hemin (10 mg/mL). For biofilm assays, bacterial suspensions were seeded into 24-well plates containing sterile round glass coverslips (12 mm radius) in the presence or absence of QAL333 (0.002–64 µM) with 10 µM DPD (4,5-dihydroxy-2,3-pentanedione) as a universal precursor for autoinducer molecules for 48 h. After incubation, biofilms formed on the coverslips were gently washed three times with phosphate-buffered saline (PBS), stained with 1% crystal violet for 15 min, and destained with an acetone–ethanol solution (20:80, v/v). The absorbance of the destained solution was measured at 590 nm using a microplate reader (Epoch2, Bio-Tek, USA) to quantify biofilm biomass.

For gingipain gene expression analysis, P. gingivalis cultures were treated with 10 µM DPD, with or without 2 µM QAL333, for 48 h. Total RNA was extracted using the easy-BLUE™ RNA Extraction Kit (iNtRON Biotechnology), and cDNA was synthesized using AccuPower^® CycleScript RT PreMix dN6 (Bioneer, Korea). RT-qPCR was performed using Power SYBR™ Green PCR Master Mix (Applied Biosystems) to quantify gingipain expression.

For inflammasome activation and cytokine quantification, bone marrow–derived macrophages (BMDMs) were isolated from mouse femurs and tibias as previously described⁴⁸. All animal experimental protocols were approved by the Institutional Animal Care and Use Committee (IACUC) of the Chonnam National University (Approval No. CNU IACUC-YB-2021-05) and conducted in compliance with that institution’s guidelines for animal research. Cells were maintained in Iscove’s Modified Dulbecco’s Medium (IMDM; Gibco, 12440061) supplemented with 1% penicillin/streptomycin (Gibco, 15140122) and cultured without FBS for QAL333 treatment. BMDMs were seeded at 1 × 10⁶ cells/mL in 48-well plates and treated with QAL333 (1–10 µM) under serum-free conditions. For NLRP3 inflammasome activation, cells were primed with lipopolysaccharide (LPS, 100 ng/mL; InvivoGen) for 6 h, followed by QAL333 treatment for 30 min and ATP stimulation (1 mM; Sigma- Aldrich, A2383) for an additional 30 min. Culture supernatants were collected, and the levels of IL-1β and TNF-α were measured using ELISA kits (R&D Systems, USA) according to the manufacturer’s protocols.

Statistical analysis

All experiments were independently repeated at least three times. Statistical comparisons for cell viability and zebrafish xenograft assays were performed using Welch’s t-test. Biofilm and IL-1β assays were analyzed by one-way ANOVA followed by Dunnett’s multiple comparison test. A p -value < 0.05 was considered statistically significant.

Supplementary Information

Below is the link to the electronic supplementary material. References 49–69 are cited exclusively in Supplementary Tables S3C and S3D, which summarize prior studies used for pathway annotation and gene set curation. 

Author contributions

Conceptualization, S.P. and J.S.; methodology, S.P., J.K., Y.J., E.H., S.Y.J., J.Su., E.J., M.T.R., E.-J.R., and Y.-J.K.; software, J.K., Y.J and H.H.K.; validation, S.P. E.J., M.T.R., S.I. and H.H.K.; formal analysis, H.H.K., and S.I.; investigation, J.K., Y.J., E.H., Y.-J.K., S.Y.J., and J.Su.; resources, E.-J.R., J.S. and Y.-J.K.; data curation, J.K., Y.J., J.Su. E.-J.R. and E.H.; visualization, J.K., Y.J., S.Y.J., and E.H.; writing—original draft preparation, S.P. and H.H.K.; writing—review and editing, S.P. E.J. M.T.R. S.I. and J.S.; supervision, S.P.; project administration, S.P.; funding acquisition, S.P and J.S.Hyon Hee Kim and Suji Im contributed equally to this work as co–first authors. Seyeon Park and Jaehyun Shim are co–corresponding authors. All Authors have read and agreed to the published version of the manuscript.

Funding

This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2022R1A2C1004297, RS-2023–00264573, and RS-2025-07072968).

Data availability

>All source data used for figure generation, including IC₅₀ values, gene expression matrices, and model predictions, are available in the Supplementary Information and at the following repository: 10.5281/zenodo.18012549. The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request for non-commercial academic purposes. The mRNAseq data supporting the findings of this study have been deposited in the NCBI Sequence Read Archive (SRA) under BioProject ID PRJNA1281798. The repository additionally contains annotated analysis scripts that reproduce the key computational results of this study, including (i) transcriptomic feature selection, (ii) leave-one-out cross-validation–based performance metrics (accuracy, AUC, precision, recall, and F1 score), (iii) identification of important gene pairs, and (iv) generation of the predictive modeling figures and tables. The main analysis pipeline is provided in the notebook “QAL333.ipynb,” and execution details are described in the repository README file.

Declarations

Competing interests

The authors declare no competing interests.