Skip to main content
NCBI home page
RSC Advances logoLink to RSC Advances
. 2026 Sep 28. Online ahead of print. doi: 10.1039/d6ra05448j

Lipid tail length dictates membrane disruption by carmofur analogs and reveals noncanonical modes of action

Chancie Chou a, Arin Kasliwal a, Lavernie Chen a, Kathleen Hsu a, Aarav Anand a,b, Uday Srivastava a, Sanika Vaidya a, Felicia Leung a, Eileen Yuyama a, Ieva Chepurna a, Mansha Gupta a, Lucas Quan a, Nolan Tsai a, Kayla Huang a, Olivia Kwok a, Jenny Zhang a, Audrey Tseng a, Viet-Tran Gemma Do a, Evelyn Du a, Tapash Jay Sarkar c, Akira Yamamoto a,c, Joseph Laurienzo b, Joseph Pazzi a,c,†,✉, Edward Njoo a,†,✉
PMCID: PMC13618345  PMID: 42807973

Abstract

Carmofur, a lipidated urethane derivative and prodrug of 5-fluorouracil, has an established clinical application as an antineoplastic agent for the treatment of colorectal cancer, primarily through inhibition of acid ceramidase and subsequent intracellular drug release. Beyond this oncological application, recent drug repurposing efforts have identified covalent inhibition of the SARS-CoV-2 main protease as an additional antiviral mechanism. The pronounced amphiphilic character of carmofur, arising from a polar fluoropyrimidine headgroup appended to a hydrophobic aliphatic tail, raises the possibility of an additional membrane-directed mechanism orthogonal to canonical target engagement. Driven by observations that certain lipidated carmofur analogs induce potent disruption in giant unilamellar vesicles (GUVs), which serve as model-membrane systems, we evaluated nine analogs with differing alkyl tail lengths and saturation to characterize the capacity of such lipidated compounds to directly engage with biological membranes. To efficiently quantify the membrane-disrupting capabilities of these analogs, we benchmarked the detection accuracy of a transformer-based model and a CNN-based vision model, where YOLOv11 matched RT-DETRv3's detection accuracy while running at a fraction of its computational cost. Through both cell-free and in vitro experiments, we demonstrate that lipidated analogs of carmofur exhibit a chain-length-dependent membrane disruption mechanism, with maximal cell-free disruption peaking at the decyl and dodecyl tail lengths. By performing critical micelle concentration and dynamic light scattering analyses, we found that hydrophobic interactions drive self-assembly, with alkyl chain length influencing micelle formation and size distribution. Collectively, these properties may underlie their biological activity, revealing the significance of a previously underappreciated mode of action of lipidated anticancer agents.

Enabled by an optimized computer vision platform for large-scale image analysis, we demonstrate that lipidated carmofur analogs of varying alkyl length exhibit potent membrane-disrupting activity on giant unilamellar vesicles (GUVs).graphic file with name d6ra05448j-ga.webp

Introduction

Lipidation of small-molecule therapeutics has emerged as a powerful strategy for improving pharmacokinetic properties, including drug tolerability, membrane permeability, and bioavailability.1–5 Beyond these advantages, lipidated bioactive compounds can directly interact with biological membranes, thereby eliciting mechanisms of action distinct from canonical target engagement. Carmofur, a hexyl carbamoyl lipidated derivative and prodrug of 5-fluorouracil (5-FU), has long been utilized clinically as a potent antineoplastic agent for the treatment of colorectal cancer.6–9 Subsequent studies revealed that its anticancer efficacy could be ascribed to acid ceramidase inhibition10–14 while large-scale drug repurposing efforts also identified carmofur as a covalent inhibitor of the SARS-CoV-2 main protease (Mpro).15–20 Distinct from its established enzymatic targets, our laboratory has recently demonstrated that select carmofur analogs are able to induce rapid and potent disruption of model membrane systems21 (Fig. 1A).

Fig. 1. Carmofur has demonstrated diverse clinical and biophysical functions. (A) A brief history of carmofur's development, clinical use, and applications. Protein-ligand structures were rendered from PDB IDs 6MHM and 7BUY using Maestro (Schrödinger, LLC, 2026). (B) Chemical structures of synthesized carmofur urethane analogs with varying alkyl tail lengths and saturation (1–9).

Fig. 1

Open in a new tab

Structurally, carmofur and its analogs comprise a hydrophilic nucleobase headgroup appended to a hydrophobic aliphatic tail, imparting prominent amphiphilic character that may facilitate lipid membrane association. The inhibitory activity of carmofur against acid ceramidase is putatively attributed to its structural mimicry of long-chain lipophilic ceramides, which constitute endogenous substrates of the enzyme.10,11 Our subsequent observation that lipidated carmofur analogs exhibited cytotoxicity in cancer cells at the limit of their aqueous solubility suggested that their surfactant-like biophysical properties may be associated with their biological mode of action. This is consistent with broad cellular toxicity induced by amphiphilic small molecules.22–24 We further interrogated this biophysical phenomenon through quantification of their membrane-disrupting capabilities in Giant Unilamellar Vesicles (GUVs).25 This artificial biomimetic membrane system resembles the phospholipid bilayer of cellular membranes while avoiding the complex and possibly confounding system of membrane proteins, carbohydrates, and extracellular matrix.26–28 From a compound library of nine carmofur analogs consisting of varying alkyl, aryl, and heteroalkyl side chains, the lipidated urethane analogs bearing dodecyl and octadecyl side chains exhibited exceptionally potent membrane-rupturing activity.21 These findings underscored the need for further systematic investigation of how alkyl tail length governs biophysical interactions with membranes through associated changes in physicochemical properties.

Herein, we synthesized a library of carmofur urethane analogs bearing alkyl chains ranging from six to eighteen carbons (compounds 1–8), along with a mono-unsaturated oleyl lipid analog (9) (Fig. 1B). With the aim of providing insight into the role of incremental 5-FU lipidation in driving membrane disruption, we evaluated the potency of each carmofur analog in both artificial membrane systems and in vitro assays.

Given the inefficiency of manual counting for large-scale GUV populations, we recently reported the applications of You Only Look Once,25,29–34 version 11 (YOLOv11) to perform high-throughput and real-time quantification of membrane disruption. This enables a streamlined approach to characterizing small-molecule GUV-disrupting agents, including the aforementioned lipidated carmofur analogs, utilizing canonical fluorescence microscopy. With the need to maintain accuracy when capturing both subtle and significant differences between the membrane-rupturing activity of carmofur analogs, we trained YOLOv11 on increasingly challenging large-scale GUV imaging datasets, including those heavily populated with overlapping GUVs and other lipid vesicle species. In parallel, we trained Real-Time Detection Transformer, version 3 (RT-DETRv3), an alternate computer vision algorithm, which matched but did not exceed YOLOv11's detection accuracy on our GUV dataset. With a platform to quantify GUV-disruption activity of this compound panel, we deployed our YOLOv11 model to study the dynamics of GUV disruption.

The self-assembly and biophysical properties of our carmofur analogs were further investigated through critical micelle concentration (CMC) and dynamic light scattering (DLS) experiments. Collectively, this study demonstrates a pronounced dependence of membrane interactions and biophysical properties on the alkyl chain length of carmofur analogs, revealing that analogs bearing chain lengths of ten and twelve carbon atoms exhibit the most potent membrane-disrupting activity in both cell-free and in vitro experiments. These findings support a model where lipidation modulates biophysical activity through direct membrane association and perturbation.

Results

For the preparation of a systematic library of lipidated carmofur analogs, we employed standard isocyanate addition reactions to 5-FU,8 which afforded the corresponding aliphatic N-alkyl 5-fluorouracil urethanes 1–9 (Scheme 1). In instances where the corresponding lipidated isocyanates were not commercially available, we prepared the corresponding alkyl isocyanate in situ through elimination of a triphosgene-activated primary aliphatic amine which could be directly used in subsequent nucleophilic addition of 5-FU without the need for intermediary isolation, furnishing the decyl and oleyl urethanes (4, 9) in serviceable yields of 6.7% and 23% respectively. The structures of compounds 1–9 were characterized through 1H and 13C nuclear magnetic resonance (NMR), liquid chromatography-mass spectrometry (LC-MS), and Fourier-transform infrared (FT-IR) spectroscopy. Full synthetic protocols and characterization can be found in the SI document.

Scheme 1. General synthetic workflow for the synthesis of lipidated carmofur analogs through standard isocyanate addition reactions. All isolated yields are shown above excluding carmofur (1), which was purchased commercially.

Scheme 1

Open in a new tab

The increasing usage of deep learning-based approaches has transformed biological imaging.29 Convolutional architectures such as the YOLO variants have demonstrated remarkable speed-accuracy trade-offs across a wide range of scientific tasks.30 More recently, however, vision transformers have achieved notable performance on benchmark detection datasets.31 The RT-DETR family,32,35 subsequently refined to RT-DETRv3,36 achieves transformer-grade accuracy while still maintaining reasonable inference speeds, making it appealing for high-throughput GUV image analysis. Here, we report a systematic evaluation and comparison of optimized YOLOv11 Nano33 and RT-DETRv3-R50 36 models, both trained on an identical annotated dataset under controlled conditions.29–34,36

Development of a computer-vision detection method for GUVs

All raw GUV imaging was performed with a Nikon Ti2-E microscope equipped with a 5× air objective, a Lumencor SOLA light engine, and a sCMOS camera (4912× 3264 pixels). A dataset of fluorescence images, including a broad range of vesicle densities and sizes, was developed for model training and validation (Fig. 2). All details regarding data retrieval and augmentation are reported in SI-Db. For model training, each full GUV image was converted to grayscale and split into a 16 by 16 grid of 256 equally sized, non-overlapping tiles. From each image, a subset of tiles was selected, and every visible object within them was manually annotated with axis-aligned bounding boxes. This resulted in 264 total labeled tiles used for training. For five-fold cross validation, each micrograph's tiles were separated into five vertical bands. For each training run, one band was held out for validation and the remaining bands were used for an 80 : 20 training-validation split. While such vertical separation reduces training-validation leakage, the multiple runs account for bias.37 Each model was trained on a fixed schedule (YOLOv11 for 100 epochs, RT-DETRv3 for 300 epochs) with the best-validation checkpoint selected. For inference, we adopted slicing aided hyper-inference (SAHI), where per-tile detection is merged through a cross-tile de-duplication step, a procedure that removes redundant overlapping detections (SI-Dd). Detections were then passed through intensity-based and geometric post-process filtering to ensure the biophysical plausibility of each object, accounting for and mitigating model hallucinations. Specifically, an eccentricity upper bound, bounding-box area lower bound, and an intensity variance test threshold were placed on all GUV detections. Filter threshold logic is reported in depth in SI-De. After running inference and post-processing, annotated images with bounding circles were outputted, along with the pixel surface area of each accepted object. Pixel surface areas were converted to molar lipid volume (SI-Df), and all data was reported as percent membrane disruption based on molar lipid volume.

Fig. 2. An overview of the computer vision-enabled workflow for GUV quantification. RGB fluorescence micrographs are converted to grayscale, split into 256 tiles, then used to train object-detection models. Feature extraction is utilized for post-processing before processed images and pixel surface areas are outputted. The formula shown is used to convert the pixel surface area to molar lipid content.

Fig. 2

Open in a new tab

On average across five folds, RT-DETRv3 did not exceed YOLOv11 in detection accuracy. The precision–recall curves compare the various sizes of YOLOv11 against those of RT-DETRv3 (Fig. 3A), revealing closely matched precision at low and mid recall, but a growing gap at higher recall. Across five folds, the YOLOv11 nano variant achieved the largest maximum-F1 score on average of 0.841, slightly elevated from YOLOv11 s′s 0.838, RT-DETRv3-R50's 0.815, and RT-DETRv3-R18's 0.807, with the curves diverging only at high recall, where YOLOv11n retained a modest but non-significant edge. During training, YOLOv11n′s validation mAP@50 plateaued by ∼50–90 epochs, while RT-DETRv3-R50 continued improving until ∼150–170 epochs (Fig. 3B), consistent with the longer schedules typically required by vision transformers. Among the averaged statistics across five folds, YOLOv11n exceeded RT-DETRv3-R50 in mAP@50 (0.864 vs. 0.853; Fig. 3C), recall (0.824 vs. 0.803; Fig. 3D), and mAP@50–95 (0.430 vs. 0.422; Fig. 3E). These margins were small and not statistically significant: the 95% confidence intervals overlapped for every metric, and paired across-fold differences did not reach significance (ΔF1 = +0.020 ± 0.028, mean ± 95% CI). YOLOv11n held a small, consistent edge in every mean without a statistically resolvable gap. Automated GUV counts closely agreed with manual segmentation in both models (Fig. 3F), resulting in near-unity coefficients of determination; YOLOv11n achieved an R2 of 0.972, and RT-DETRv3-R50 achieved an R2 0.965. It can thus be concluded that both structures train successfully.

Fig. 3. Detection performance of YOLOv11 (blue) and RT-DETRv3 (red) on the GUV dataset under 5-fold cross-validation. (A) Precision–recall curves on the validation are set at an IoU threshold of 0.5 for model variants; the maximum-F1 operating point is reported in the legend. (B) Validation mAP@50 is reported at every tenth training epoch for YOLOv11n and RT-DETRv3-R50, and (C) per-fold mAP@50 is shown on the right. (D) Per-fold recall and (E) per-fold mean average precision over IoU 0.50–0.95 (mAP@50–95) are shown. In the per-fold panels (C–E), bars denote the mean and error bars give the 95% confidence interval. (F) Agreement between model predictions and ground-truth GUV counts for each image are shown, with the line of identity (1 : 1) and coefficient of determination (R2) reported for each model.

Fig. 3

Open in a new tab

Because the overlapping confidence intervals make this quantitative advantage largely statistically inconclusive, a qualitative examination of detection behavior was used to determine which model was better suited for this use-case. Both models perform well in sparsely populated images (Fig. 4A) precisely detecting nearly every object. Additionally, both maintain accurate coverage of sub-15 µm vesicles (Fig. 4B) and of highly crowded fields (Fig. 4C). However, RT-DETRv3-R50 fails to capture several large GUVs that YOLOv11 detects (Fig. 4D) and similarly drops multiple objects around the densest vesicle cores (Fig. 4B and C), consistent with its lower average recall across five folds. Although these missed detections may seem infrequent, they are disproportionately impactful in the quantification of membrane disruption; since detected pixel area is converted to lipid amount (in moles), a single missed large GUV biases the estimated disruption data far more than its impact on the raw object counts. YOLOv11's superiority in precisely detecting these large and densely packed vesicles, along with its faster and more stable training convergence, compelled us to select it to quantify all GUV membrane disruption data in the experiments reported hereafter.

Fig. 4. The qualitative detection results for RT-DETRv3-R50 (top row) and YOLOv11 (middle row) across four image conditions, with raw fluorescence images shown for reference (bottom row). Green circles indicate accepted GUVs; red circles indicate geometrically rejected detections; blue circles indicate brightness-rejected detections. All images are shown at identical scales, reflected by the 10 µm scale bar in the top-left image. The four image conditions included in the comparison are (A) samples with low GUV density, (B) samples with small GUVs, (C) samples with high GUV density, and (D) samples with large GUVs.

Fig. 4

Open in a new tab

Characterization of analogs using GUV membrane disruption assays

Membrane disruption assays on GUVs were performed to assess whether carmofur and its urethane analogs disrupt lipid bilayers. GUVs serve as cell-sized model membranes whose responses to membrane-active small molecules can be visualized through fluorescence microscopy. To quantify membrane disruption, the moles of vesicles before and after the addition of compounds 1–9, 5-fluorouracil (5FU), sodium dodecyl sulfate (SDS), 1-octanesulfonic acid, and dimethyl sulfoxide (DMSO) (negative control) were calculated by quantifying intact GUVs with the computer vision model YOLOv11 and comparing the results before and 15 minutes after the assay. Using previously reported methods for formation and sedimentation of GUVs, we systematically evaluated membrane disruption for each carmofur analog over concentrations ranging from 1.15 mM to 1.15 pM to determine whether treatment with these compounds produced concentration and tail-dependent disruption. Fig. 5A–D shows selected fluorescence microscopy images of typical GUV populations before and after the administration of various concentrations of DMSO, carmofur (1), decyl urethane (4), and hexadecyl urethane (7).

Fig. 5. Evaluation of GUV membrane disruption induced by carmofur and its analogs. The top row displays vesicles before treatment, and the following three rows display vesicles 15 minutes after treatment of compounds at 1.15 mM, 1.15 µM, and 1.15 pM. Scale bar is 100 µm. Column (A) GUVs treated with 0.0225% v/v DMSO, acting as a negative control. Columns (B–D) GUVs treated with compounds 1, 4, and 7. (E) Quantitative analysis for GUV disruption of compounds 1–9, ranging from 1.15 mM to 1.15 pM. Data are presented as mean ± SD.

Fig. 5

Open in a new tab

Before the addition of compounds 1–9, all wells exhibited fairly uniform populations of intact fluorescent GUVs with commensurate density, size distributions, and membrane morphology (Fig. 5A–D). Intact vesicles appeared as demarcated circular structures with even orange fluorescence coloring within the membrane boundary. The post-treatment fluorescence images indicated visible differences in GUV membrane disruption across the analogs. Notably, at high concentrations, decyl urethane (4) produced more pronounced GUV patches and brighter fluorescent debris relative to the other analogs, such as carmofur (1) and hexadecyl urethane (7). Similar trends were also observed in the expanded pre- and post-treatment image sets presented in Fig. S1, which compares compounds 1–9 evaluated at 1.15 mM and 115 nM. In Fig. S1, dose-dependency was visibly observed across all compounds except decyl urethane (4), where GUV disruption was consistently high across all concentrations displayed.

Our quantitative results, obtained using YOLOv11, confirmed dose-dependency across compounds 1–9 (Fig. 5E). Overall, compounds 1–9 resulted in distinct disruption profiles, indicating that membrane activity was influenced by both concentration and lipid tail length. At 1.15 mM, compounds 1–9 produced GUV disruption values of 67.20 ± 3.68%, 77.32 ± 4.61%, 79.40 ± 2.74%, 88.36 ± 3.51%, 82.88 ± 3.52%, 73.35 ± 3.70%, 50.66 ± 4.50%, 52.87 ± 4.33%, and 67.42 ± 3.93%, respectively. Decyl urethane (4) yielded the strongest disruption at this concentration and maintained high activity across the entire dilution series.

1-Octanesulfonic acid was included to evaluate whether removing the 5-FU moiety impacted membrane disruption. As shown in Fig. 5E, 1-octanesulfonic acid exhibited low disruption across all concentrations, indicating that the 5-FU headgroup may potentiate the observed membrane activity. Due to the low activity of 1-octanesulfonic acid, the remaining sulfonic acid analogs were not prioritized for further testing. The minor apparent disruption and associated variability measured in the 1-octanesulfonic acid and DMSO trials were attributed to vesicle displacement outside the image frame following compound addition rather than compound-induced disruption.

After probing the significance of the 5-FU head group, we sought to evaluate whether chain saturation influences membrane-disrupting ability by comparing the unsaturated oleyl urethane (9) to its saturated counterpart, octadecyl urethane (8). Oleyl urethane (9) produced greater GUV disruption than octadecyl urethane (8) at the highest concentration by approximately 1.28-fold (Fig. 5E). This suggests that the unsaturated oleyl tail may enhance membrane–disruptive activity relative to the saturated octadecyl analog. This increased activity could be ascribed to the presence of a cis-double bond, introducing a kink that may alter the compound's interaction with the lipid membrane. However, both compounds showed a similar decrease in disruption upon dilution, indicating that chain saturation does not substantially alter dose-dependence.

Since GUVs are reductionist models of the cell plasma membrane, we elected to investigate the potential role of various, more complex lipid compositions of GUVs in membrane disruption. In previous studies,21,25 we observed minimal difference in GUV percent disruption based on GUV osmolarity and negative charge. Here we further considered potential effects due to various cholesterol and lipid-conjugated dye compositions, and we also find quantitatively that there were minimal membrane disruption differences for these modified vesicles against SDS and DMSO as shown in Fig. S5 and S6.

To succinctly summarize the GUV disruption results, we fit Hill slope curves and extracted half-maximal effective concentration (EC50) values to all of the GUV intact data (GUV intact = 100 − GUV disrupt) across the entire analog library. The positive control, SDS, performed exactly as expected for an anionic detergent, yielding a biophysically valid EC50 of 0.115 mM, while the parent drug 5-fluorouracil along with 1-octanesulfonic acid exhibited no true membrane–disruption activity and thus no extractable EC50 values. Among the active urethane analogs, membrane disruption increased with increasing carbon chain length, moving from the moderate activity of the shorter heptyl (C7) and octyl (C8) chains to peak potency in the decyl (C10) and dodecyl (C12) analogs, which resulted in a precipitous lowering of EC50 into the nanomolar ranges. However, this exponential increase in potency abruptly halted at the tetradecyl (C14) chain length, demonstrating a cut-off effect where the C14, C16, and C18 analogs completely lost their ability to disrupt the vesicles, potentially due to their tendency to self-aggregate before reaching the lipid bilayer.

Chain length dependent self-assembly and hydrodynamic size distributions of carmofur alkyl urethane analogs

To determine whether attaching various medium- and long-chain length hydrophobic tails to a 5-FU head group causes complex self-assembly behavior that could alter drug uptake, we conducted DLS measurements to characterize the analogs at 0.5 mM. Fig. 6A shows that the size distributions of each analog reached a peak at around 100 nm regardless of tail length, suggesting that at these high concentrations, the lipid-tails are forming much larger, kinetically trapped structures rather than simple micelles. Additionally, the higher mean values and slightly higher median values of carmofur (1), the heptyl (2), and octyl (3) analogs (Fig. 6B) quantitatively confirm the skewed, right-tailed appearance of their size distributions. These right-tails suggest the presence of transient large flocculates, which are likely unstable once the lipid tail becomes long enough to dictate the self-assembly behavior of the molecule.

Fig. 6. Chain length-dependent self-assembly and critical micelle concentration analysis of carmofur alkyl urethane analogs. (A) Relative intensity size distributions are shown as a function of radius for each of the analogs. (B) Mean and median radius are shown as a function of alkyl tail length. (C) Individual contact angle versus concentration plots of dodecyl urethane (5), tetradecyl urethane (6), hexadecyl urethane (7), and octadecyl urethane (8). (D) Inflection point as a function of compound. The inflection point was defined as the concentration preceding the largest drop in contact angle and was assigned to a change in contact angle. Note: higher concentrations were unable to be tested due to the low solubility without heating of the compounds in 2% DMSO in water.

Fig. 6

Open in a new tab

To further evaluate this chain length-dependent assembly behavior and investigate whether these differences in self-assembly could contribute to the distinct GUV popping behaviors observed across compounds 1–9, we measured contact angle as a function of concentration and used the inflection point as an estimate of micellization. As shown in Fig. 6C, the log3 dilutions initially showed a relatively unchanging contact angle before the inflection point. Once the inflection point was reached, the contact angle decreased with increasing concentration.

Across the analogs with measurable inflection points, the inflection point concentration decreased almost linearly with increasing alkyl tail length, as shown in Fig. 6D. Carmofur (1) and heptyl urethane (2) were not included in this plot because no clear inflection point was observed within the tested concentration range (Fig. S2). This indicates that longer chain analogs undergo apparent self-assembly at significantly lower concentrations. This trend is consistent with stronger hydrophobic driving forces in the longer chain compounds. However, because GUV popping did not exhibit a similarly abrupt change upon crossing the inflection point, this suggests that there is no clear relationship between self-assembly and GUV disruption.

In vitro & live cell imaging

In vitro cell viability and cytotoxicity assays (MTT/LDH)

To assess the in vitro antineoplastic efficacy of compounds 1–9, we performed 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT)38 and lactate dehydrogenase (LDH)39 cytotoxicity assays against a panel of mammalian cancer cell lines, including A549 (human lung adenocarcinoma), HCT-116 (human colorectal carcinoma), B16-F10 (murine skin cancer), and HT-29 (human colorectal adenocarcinoma). We evaluated our compounds alongside a known surfactant, sodium dodecyl sulfate (SDS) as a direct comparison with our dodecyl urethane (5) analog to investigate whether the 5-FU headgroup has a significant effect on cytotoxicity. Cell viability assessment at the 96 h timepoint demonstrated dose-dependent activity across all four cell lines (Fig. 7A). While limited antiproliferative activity was observed in the B16-F10 and HT-29 cells, prominent antiproliferative activity was observed in the HCT-116 cell line, with compounds 1–9 exhibiting IC50 values of 2.743 µM, 2.091 µM, 1.985 µM, 4.039 µM, 3.075 µM, 3.642 µM, 3.233 µM, 3.481 µM, and 3.714 µM, respectively. Intriguingly, analogs exhibiting the most potent membrane-disrupting activity in GUV assays did not demonstrate superior activity in in vitro assays. Irrespective of the length of the alkyl chain, each compound exhibited virtually indistinguishable dose-dependent antiproliferative activity, suggesting that the primary mechanism of anticancer activity of aliphatic carmofur analogs may not be directly influenced by the GUV membrane–disruption activity observed on model systems.

Fig. 7. Evaluation of compounds 1–9 in cancer cell viability, cytotoxicity, and membrane permeabilization assays. (A) Cell viability (MTT) in A549, HCT-116, B16-F10, and HT-29 cells treated with compounds 1–9. (B) Representative fluorescence microscopy images of HCT-116 cells following 24 h treatment, stained with propidium iodide (PI) and fluorescein to assess membrane permeabilization. (C) ImageJ quantification of the normalized fluorescence intensities from panel B. Data are presented as means ± SD relative to the negative control (0.5% v/v DMSO).

Fig. 7

Open in a new tab

Additional cytotoxic evaluation through lactate dehydrogenase assays (LDH) revealed that all compounds exhibited minimal perturbation of the cell membrane, indicating a limited induction of necrotic cell death through cell membrane lysis (Fig. S5). Consistent with similar antiproliferative activity observed across all compounds, these observations suggest that membrane disruption is not an operative mechanism towards the underlying anticancer activity of these compounds.

In vitro imaging

Visualization of cellular membrane integrity was quantified using fluorescence microscopy to evaluate the effects of our lipidated compounds on cell membrane morphology and lipid packing. Fluorescein, a fluorescent small molecule probe typically impermeable to intact cellular membranes, together with a counter propidium iodide (PI) stain, was used as a proxy for determining intracellular leakage and membrane permeabilization. HCT-116 cells seeded in an imaging plate were treated with compounds 1–9 at 2.5 µM, and imaging was performed following compound exposure for 24 h. After an incubation period of 30 minutes with the addition of 1 mg mL−1 solution of fluorescein and PI (Fig. 7B), individual wells were imaged using standard fluorescence microscopy at the 24 h timepoint under 10× magnification to visualize alterations in membrane permeability. Using the ImageJ software, we extracted the average pixel intensity for the fluorescein images to quantify membrane permeabilization, allowing for comparison between compounds. At the 24 h timepoint, membrane permeabilization progressively increased with extension of the lipid tail length from a hexyl to decyl alkyl chain, where the decyl urethane analog (4) exhibited the maximal average fluorescence intensity, followed by a decline in membrane permeabilization from dodecyl urethane (5) to octadecyl urethane (8) (Fig. 7C). This corresponds with results from the GUV disruption assay, in which decyl urethane (4) also exhibited the highest membrane disruption activity. Intriguingly, the unsaturated oleyl analog (9) exhibited increased fluorescence intensity, similar to that of decyl urethane (4). Crucially, it was generally noted that our lipidated compounds did not induce significant perturbation in cell membrane morphology.

Discussion & conclusions

Lipid conjugation has been demonstrated to be an effective means to alter the biological function of pharmaceutical compounds. Carmofur and related lipidated urethane derivatives of 5-fluorouracil, which are reported to possess greater bioavailability than 5-fluorouracil and potentially act through the inhibition of acid ceramidase, also possess a unique amphiphilic character. In connection to this, we had previously observed that a dodecyl urethane analog (5) exhibited potent, rapid, and concentration-dependent activity in the disruption of GUVs, suggesting a potential membrane-directed mechanism of action. In order to further interrogate this phenomenon, we systematically assessed the membrane disruption and self-assembly of a library of carmofur analogs with increasing alkyl chain lengths in a panel of cell-free and in vitro experiments.

While we have previously reported computer vision-based algorithms capable of GUV detection on large-scale imaging, the fine-grained structure–activity relationships central to this study necessitated the optimization of more precise computer vision algorithms to capture nuanced differences in membrane–disruption activity across the library of analogs. Altogether, the results from our benchmarking and performance comparison of the optimized YOLOv11n and RT-DETRv3-R50 indicate that the two architectures are statistically indistinguishable in the detection accuracy on this task. What distinguishes the models, however, is their practicality: YOLOv11n trains faster and runs with a fraction of RT-DETRv3's parameters and runtime. It also more completely detects the large, densely packed vesicles that dominate the lipid-amount readout. In light of these findings, the convolutional model was chosen for membrane disruption quantification.

Enabled by this fine-tuned optimization of a machine learning model for the rapid quantification of GUV disruption, we identified a pronounced dependence of membrane–disruption activity on lipid length, where 5-fluorouracil derivatives possessing a decyl or dodecyl alkyl lipid exhibited the most significant GUV disruption activity. This was further reflected in in vitro membrane permeability experiments, in which imaging of fluorescein-stained cells showed an increase in fluorescence intensity with chain lengths leading up to the decyl tail.

Further, we observed a linear decrease in CMC with increasing alkyl length, demonstrating that self-assembly of lipidated carmofur analogs is principally driven by hydrophobic effects. In DLS measurements of micelle size, lipidated analogs with the shortest alkyl chains demonstrated increased micelle populations of larger radii. This revealed distinct aggregation behavior among shorter-chain analogs, consistent with altered micellar organization and packing. This behavior may arise from reduced hydrophobic packing efficiency in shorter alkyl derivatives, resulting in more weakly organized assemblies with broader size distributions in solution.

Collectively, these findings establish that alkyl chain length is the determinant in the unique biophysical properties of lipidated 5-fluorouracil derivatives analogous to carmofur. These observations further indicate that lipid tail length modulates both supramolecular organization in solution, and in the capacity of these compounds to perturb phospholipid bilayers. More broadly, this work highlights the potential to repurpose such compounds as membrane-active agents, where systematic variation of lipid length leads to subtle yet significant changes to the delivery and self-assembly of these and similar compounds.

Methods

Cell imaging using fluorescent dyes

HCT-116 cells were seeded at 70% confluency in a 96-well tissue culture-treated flat-bottom plate and incubated in a humidified incubator at 37 °C in 5% CO2. A drug medium solution was prepared by adding 200-fold DMSO solutions of compounds 1–9, in addition to a sodium dodecyl sulfate (SDS) control, to cell culture media to reach a desired concentration of 2.5 µM. Six wells of negative controls (0.5% v/v DMSO) were also included. The plates were then incubated for 24 hours at 37 °C in a humidified incubator (5% CO2). After incubation, 10 µL of a 1 mg mL−1 solution of fluorescein in PBS was added to each well. Upon incubation at 37 °C in 5% CO2 for 1 h, 1 µL of a 100 µM solution of propidium iodide (PI) in PBS was added to each well. Upon incubation at 37 °C in 5% CO2 for 5 minutes, each well was washed with PBS twice. Cells were subsequently imaged using a Zeiss Axiovert 200 M fluorescence microscope with a Zeiss FLUAR 10×/0.50 M27 lens. HCT-116 cells were purchased from European Collection of Authenticated Cell Cultures (ECACC), Millipore Sigma Cat. # 91091005-1. HT-29 cells were purchased from European Collection of Authenticated Cell Cultures (ECACC), Millipore Sigma Cat. # 91072201. A549 cells and B16F10 cells were a generous gift from Dr Feng Wang-Johanning and Dr Gary Johanning (SunnyBay Biotech, Fremont, CA) and were originally purchased from ATCC.

GUV rupture fluorescence imaging assay

A GUV solution was prepared through literature-reported protocols using a 1 mg mL−1 mixture of 1,2-dioleoyl-sn-glycero-3-phosphocholine (DOPC) : 1,2-dioleoyl-sn-glycero-3-phosphoethanolamine-N-(lissamine rhodamine B sulfonyl) (RhodamineDOPE) (99.5 : 0.5 mol%). To prepare for imaging, the wells of a Greiner 96-Well Glass Bottom Plate were passivated with 0.75 mg mL−1 casein solution in 1× Dulbecco's phosphate-buffered saline (DPBS). After incubation for 1 h, 85 µL of a 95 mM glucose solution was added to each of the wells, followed by 2 µL of the GUV solution in 100 mM sucrose. GUVs were then sedimented for 1 h. We collected images of the GUVs at 555 nm using a Zeiss Axiovert 200 M fluorescence microscope with a Zeiss FLUAR 5×/0.25 M27 objective. Compounds 1–9 and 1-octanesulfonic acid were initially heated and sonicated until no visible precipitate remained, then serially diluted ten-fold from 50 mM to 5 pM in a Falcon 96-well clear round-bottom (U-bottom) polystyrene microplate. 2 µL of the appropriate concentrations were subsequently added to the imaging plate in triplicate. A negative control was established with approximately 0.0225% v/v DMSO. Following a 15 minute incubation, representative images of the GUVs in each well were captured to characterize membrane disruption. The percentages of GUVs disrupted were determined using YOLOv11.

Critical micelle concentration evaluation assay

Critical Micelle Concentration measurements were performed by evaporating 30 µL of 500 µM compound solution in DMSO to dryness at high temperature in a SpeedVac (Savant SC100) for approximately 2 h, followed by reconstruction with 30 µL of 2% DMSO in water. Samples were then bath-sonicated (Thermo Fisher) for 15 min and briefly vortexed before preparing log 3 serial dilutions from 500 µM to approximately 1.29 pM in a Falcon 96-well clear round-bottom (U-bottom) polystyrene microplate by adding 7 µL of compound solution to 14 µL of 2% DMSO across four wells per concentration. For each concentration, 2 µL was taken from each of the four wells and put onto a Teflon surface placed in front of a blue background. A 2% DMSO control was also included. The droplets were then photographed37 using an iPhone 13 camera fitted with an external macro lens at a 5× zoom, and contact angles were measured from photos using ImageJ (180 − θE). Changes in contact angle across the dilutions were used to determine the inflection point.

Dynamic light scattering

Dynamic Light Scattering (DLS) measurements were performed on 500 µM solutions of compounds 1–9 dissolved in DMSO and water. The pure compound in powder form was first dissolved in DMSO to reach a final concentration of 50 mM. Then, the DMSO solution was diluted 100× to 500 µM in DI water, with a final solvent composition of 1% DMSO. After this, the 500 µM solution was vortexed before being tip-sonicated for 4 minutes to ensure a homogenous size distribution. After sonication, the solution was then vortexed again and transferred to a quartz cuvette. The cuvette was then placed in a Protein Solutions DynaPro Dynamic Light Scattering machine, and triplicates of 100 acquisitions were run for each analog. The average radius measurements of the triplicates were taken to determine the radius of carmofur and its analogs.

Author contributions

C. C., L. C., K. H., S. V., F. L., E. Y., I. C., M. G., K. H., O. K., J. Z., A. T., V. T. G. D., E. D., and E. N. synthesized the compounds used in this study. U. S. and A. K. prepared GUVs and performed GUV experiments. L. Q., A. K., A. Y., and J. P. performed DLS experiments. A. K., N. T., and J. P. performed CMC experiments. A. A. and J. L. optimized computer vision training and produced the data pipeline for GUV images. C. C., S. V., L. C., K. H. performed cell viability assays. F. L. performed mass spectrometry experiments for compound characterization. C. C. and L. C. performed in vitro imaging experiments. T. J. S., A. Y., J. L., J. P., and E. N. supervised the study and conceived of the project.

Conflicts of interest

There are no conflicts of interest to declare.

Abbreviations

5-FU

5-Fluorouracil

GUV

Giant unilamellar vesicle

YOLOv11

You only look once, version 11

RT-DETRv3

Real-time detection transformer, version 3

CMC

Critical micelle concentration

DLS

Dynamic light scattering

NMR

Nuclear magnetic resonance

FT-IR

Fourier-transform infrared

SAHI

Slice-aided hyper inference

DFL

Distribution focal loss

GIoU

Generalized intersection over union

MLV

Multi-lamellar vesicle

MVV

Multi-vesicular vesicle

DMSO

Dimethyl sulfoxide

SD

Standard deviation

MTT

3-(4,5-Dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide

SDS

Sodium dodecyl sulfate

LDH

Lactate dehydrogenase

PI

Propidium iodide

DMEM

Dulbecco's modified eagle medium

HEPES

4-(2-Hydroxyethyl)-1-piperazineethanesulfonic acid

PBS

Phosphate-buffered saline

β-NAD+

Beta-nicotinamide adenine dinucleotide

Methoxy PMS

1-Methoxy-5-methylphenazinium methyl sulfate

INT

2-p-iodophenyl-3-p-nitrophenyl tetrazolium chloride

DOPC

1,2-Dioleoyl-sn-glycero-3-phosphocholine

Rhodamine DOPE

1,2-Dioleoyl-sn-glycero-3-phosphoethanolamine-N-(lissamine rhodamine B sulfonyl)

DPBS

Dulbecco's phosphate-buffered saline

FBS

Fetal bovine serum

CDCl3

Deuterochloroform

TLC

Thin layer chromatography

LC-MS

Liquid chromatography-mass spectrometry

NMS

Non-maximal suppression

OAT

One-at-a-time

mAP50

Mean average precision at 50% intersection over union

Supplementary Material

RA-OLF-D6RA05448J-s001

Acknowledgments

The authors gratefully acknowledge Dr Stephen Lynch and the Stanford University NMR Facility for providing access to high-field NMR spectra. Additionally, the authors gratefully acknowledge Dr Feng Wang-Johanning and Dr Gary Johanning of SunnyBay Biotech for generous contribution of several of the cell lines used in this study.

Data availability

Characterization for compounds: all 1H, 19F, and 13C{1H} NMR spectra are available in the supporting information document and have been deposited in the Harvard Dataverse Database (DOI: https://doi.org/10.7910/DVN/77ZUIW). Biophysical evaluation of compounds: the datasets generated and analyzed during the current study, including GUV microscopy images, contact angle droplet photographs, and raw dynamic light scattering (DLS) values, are available upon request. YOLOv11 and RTDetR code: all code involved in model training and GUV quantification inferences, as well as external links to model parameters, are actively maintained in the following GitHub repository: https://github.com/josephlaurienzo-sudo/guv-detection.git.

Supplementary information (SI): key experimental, spectroscopic and characterization data for novel compounds, and supplementary figures. All code use for the processing and preparation of data is available on GitHub. See DOI: https://doi.org/10.1039/d6ra05448j.

References

  1. Vadlapudi A. D. Vadlapatla R. K. Kwatra D. Earla R. Samanta S. K. Pal D. Mitra A. K. Targeted Lipid Based Drug Conjugates: A Novel Strategy for Drug Delivery. Int. J. Pharm. 2012;434(1–2):315–324. doi: 10.1016/J.IJPHARM.2012.05.033. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Dahan A. Markovic M. Aponick A. Zimmermann E. M. Ben-Shabat S. The Prospects of Lipidic Prodrugs: An Old Approach with an Emerging Future. Future Med. Chem. 2019;11(19):2563–2571. doi: 10.4155/fmc-2019-0155. [DOI] [PubMed] [Google Scholar]
  3. Zaro J. L. Lipid-Based Drug Carriers for Prodrugs to Enhance Drug Delivery. AAPS J. 2015;17(1):83–92. doi: 10.1208/s12248-014-9670-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Čerpnjak K. Zvonar A. Gašperlin M. Vrečer F. Lipid-Based Systems as a Promising Approach for Enhancing the Bioavailability of Poorly Water-Soluble Drugs. Acta Pharm. 2013;63(4):427–445. doi: 10.2478/acph-2013-0040. [DOI] [PubMed] [Google Scholar]
  5. Seley-Radtke K. L. Yates M. K. The Evolution of Nucleoside Analogue Antivirals: A Review for Chemists and Non-Chemists. Part 1: Early Structural Modifications to the Nucleoside Scaffold. Antiviral Res. 2018;154:66–86. doi: 10.1016/J.ANTIVIRAL.2018.04.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Terabe K. Yamauchi M. Watanabe T. Ichihashi H. Takagi H. Imaizumi M. Kondo T. Prospective Adjuvant Chemotherapy, with Mitomycin C plus Long-Term Carmofur in Colorectal Cancer. Gan No Rinsho. 1986;32(3):261–267. [PubMed] [Google Scholar]
  7. Sakamoto J. Hamada C. Rahman M. Kodaira S. Ito K. Nakazato H. Ohashi Y. Yasutomi M. An Individual Patient Data Meta-Analysis of Adjuvant Therapy with Carmofur in Patients with Curatively Resected Colon Cancer. Jpn. J. Clin. Oncol. 2005;35(9):536–544. doi: 10.1093/JJCO/HYI147. [DOI] [PubMed] [Google Scholar]
  8. Wang X. Vu J. Luk C. Njoo E. Benchtop 19F Nuclear Magnetic Resonance Spectroscopy Enabled Kinetic Studies and Optimization of the Synthesis of Carmofur. Can. J. Chem. 2023;101(8):518–524. doi: 10.1139/cjc-2022-0266. [DOI] [Google Scholar]
  9. Ozaki S. Ike Y. Mizuno H. Ishikawa K. Mori H. 5-Fluorouracil Derivatives. I. The Synthesis of 1-Carbamoyl-5-Fluorouracils. Bull. Chem. Soc. Jpn. 1977;50(9):2406–2412. doi: 10.1246/bcsj.50.2406. [DOI] [Google Scholar]
  10. Dementiev A. Joachimiak A. Nguyen H. Gorelik A. Illes K. Shabani S. Gelsomino M. Ahn E.-Y. E. Nagar B. Doan N. Molecular Mechanism of Inhibition of Acid Ceramidase by Carmofur. J. Med. Chem. 2019;62(2):987–992. doi: 10.1021/acs.jmedchem.8b01723. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Realini N. Solorzano C. Pagliuca C. Pizzirani D. Armirotti A. Luciani R. Costi M. P. Bandiera T. Piomelli D. Discovery of Highly Potent Acid Ceramidase Inhibitors with in Vitro Tumor Chemosensitizing Activity. Sci. Rep. 2013;3(1):1035. doi: 10.1038/srep01035. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Frohbergh M. He X. Schuchman E. H. The Molecular Medicine of Acid Ceramidase. Biol. Chem. 2015;396(6–7):759–765. doi: 10.1515/HSZ-2014-0290/XML. [DOI] [PubMed] [Google Scholar]
  13. Ryland L. K. Fox T. E. Liu X. Loughran T. P. Kester M. Dysregulation of Sphingolipid Metabolism in Cancer. Cancer Biol. Ther. 2011;11(2):138–149. doi: 10.4161/CBT.11.2.14624. [DOI] [PubMed] [Google Scholar]
  14. Cho S. M. Kwon H. J. Acid Ceramidase, an Emerging Target for Anti-Cancer and Anti-Angiogenesis. Arch. Pharmacal Res. 2019;42(3):232–243. doi: 10.1007/S12272-019-01114-3. [DOI] [PubMed] [Google Scholar]
  15. Gao J. Zhang L. Liu X. Li F. Ma R. Zhu Z. Zhang J. Wu J. Shi Y. Pan Y. Ge Y. Ruan K. Repurposing Low-Molecular-Weight Drugs against the Main Protease of Severe Acute Respiratory Syndrome Coronavirus 2. J. Phys. Chem. Lett. 2020;11(17):7267–7272. doi: 10.1021/ACS.JPCLETT.0C01894. [DOI] [PubMed] [Google Scholar]
  16. Grigorenko B. L. Polyakov I. V. Khrenova M. G. Giudetti G. Faraji S. Krylov A. I. Nemukhin A. V. Multiscale Simulations of the Covalent Inhibition of the SARS-CoV-2 Main Protease: Four Compounds and Three Reaction Mechanisms. J. Am. Chem. Soc. 2023;145(24):13204–13214. doi: 10.1021/JACS.3C02229. [DOI] [PubMed] [Google Scholar]
  17. Jin Z. Du X. Xu Y. Deng Y. Liu M. Zhao Y. Zhang B. Li X. Zhang L. Peng C. Duan Y. Yu J. Wang L. Yang K. Liu F. Jiang R. Yang X. You T. Liu X. Yang X. Bai F. Liu H. Liu X. Guddat L. W. Xu W. Xiao G. Qin C. Shi Z. Jiang H. Rao Z. Yang H. Structure of Mpro from SARS-CoV-2 and Discovery of Its Inhibitors. Nature. 2020;582(7811):289–293. doi: 10.1038/s41586-020-2223-y. [DOI] [PubMed] [Google Scholar]
  18. Hu Q. Xiong Y. Zhu G. H. Zhang Y. N. Zhang Y. W. Huang P. Ge G. B. The SARS-CoV-2 Main Protease (Mpro): Structure, Function, and Emerging Therapies for COVID-19. MedComm. 2022;3(3):e151. doi: 10.1002/MCO2.151;WGROUP:STRING. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Ma C. Hu Y. Townsend J. A. Lagarias P. I. Marty M. T. Kolocouris A. Wang J. E. Disulfiram, Carmofur, PX-12, Tideglusib, and Shikonin Are Nonspecific Promiscuous SARS-CoV-2 Main Protease Inhibitors. ACS Pharmacol. Transl. Sci. 2020;3(6):1265–1277. doi: 10.1021/ACSPTSCI.0C00130. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Jin Z. Zhao Y. Sun Y. Zhang B. Wang H. Wu Y. Zhu Y. Zhu C. Hu T. Du X. Duan Y. Yu J. Yang X. Yang X. Yang K. Liu X. Guddat L. W. Xiao G. Zhang L. Yang H. Rao Z. Structural Basis for the Inhibition of SARS-CoV-2 Main Protease by Antineoplastic Drug Carmofur. Nat. Struct. Mol. Biol. 2020;27(6):529–532. doi: 10.1038/s41594-020-0440-6. [DOI] [PubMed] [Google Scholar]
  21. Gu T. Lu A. Wang X. Brahan N. Xu L. Zhang L. Su K. Seow K. Vu J. Luk C. Lee Y. Raman A. Pazzi J. Njoo E. Synthesis and Evaluation of Carmofur Analogs as Antiproliferative Agents, Inhibitors to the Main Protease (Mpro) of SARS-CoV-2, and Membrane Rupture-Inducing Agents. Discov. Chem. 2025;2(1):73. doi: 10.1007/s44371-025-00144-9. [DOI] [Google Scholar]
  22. Ma W. Fu X. Zhao T. Qi Y. Zhang S. Zhao Y. Development and Applications of Lipid Hydrophilic Headgroups for Nucleic Acid Therapy. Biotechnol. Adv. 2024;74:108395. doi: 10.1016/J.BIOTECHADV.2024.108395. [DOI] [PubMed] [Google Scholar]
  23. Heerklotz H. Interactions of Surfactants with Lipid Membranes. Q. Rev. Biophys. 2008;41(3–4):205–264. doi: 10.1017/S0033583508004721. [DOI] [PubMed] [Google Scholar]
  24. Inácio Â. S. Domingues N. S. Nunes A. Martins P. T. Moreno M. J. Estronca L. M. Fernandes R. Moreno A. J. M. Borrego M. J. Gomes J. P. Vaz W. L. C. Vieira O. V. Quaternary Ammonium Surfactant Structure Determines Selective Toxicity towards Bacteria: Mechanisms of Action and Clinical Implications in Antibacterial Prophylaxis. J. Antimicrob. Chemother. 2016;71(3):641–654. doi: 10.1093/jac/dkv405. [DOI] [PubMed] [Google Scholar]
  25. Chang L. Devendiran D. Gard J. Gu T. Guan A. Yamamoto A. Sarkar T. J. Njoo E. Pazzi J. Large-Scale Fluorescence Microscopy Analysis of Lipid Membrane Conformational Changes Optimized and Enabled by an AI-Guided Image Detection Method. J. Exp. Theor. Anal. 2026;4(1):3. doi: 10.3390/JETA4010003. [DOI] [Google Scholar]
  26. Stano P. Minimal Cells: Relevance and Interplay of Physical and Biochemical Factors. Biotechnol. J. 2011;6(7):850–859. doi: 10.1002/BIOT.201100079. [DOI] [PubMed] [Google Scholar]
  27. Van de Cauter L. van Buren L. Koenderink G. H. Ganzinger K. A. Exploring Giant Unilamellar Vesicle Production for Artificial Cells — Current Challenges and Future Directions. Small Methods. 2023;7(12):2300416. doi: 10.1002/SMTD.202300416. [DOI] [PubMed] [Google Scholar]
  28. Kindt J. T. Szostak J. W. Wang A. Bulk Self-Assembly of Giant, Unilamellar Vesicles. ACS Nano. 2020;14(11):14627–14634. doi: 10.1021/ACSNANO.0C03125/SUPPL_FILE/NN0C03125_SI_009.AVI. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Moen E. Bannon D. Kudo T. Graf W. Covert M. Van Valen D. Deep Learning for Cellular Image Analysis. Nat. Methods. 2019;16(12):1233–1246. doi: 10.1038/s41592-019-0403-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Redmon J.; Divvala S.; Girshick R.; Farhadi A.You Only Look Once: Unified, Real-Time Object Detection. IEEE Conference on Computer Vision and Pattern Recognition. 2016, 779–788 [Google Scholar]
  31. Carion N. Massa F. Synnaeve G. Usunier N. Kirillov A. Zagoruyko S. End-to-End Object Detection with Transformers. Lect. Notes Comput. Sci. 2020:213–229. doi: 10.1007/978-3-030-58452-8_13/SAVE-RESEARCH. [DOI] [Google Scholar]
  32. Zhao Y.; Lv W.; Xu S.; Wei J.; Wang G.; Dang Q.; Liu Y.; Chen J.DETRs Beat YOLOs on Real-Time Object Detection. IEEE Conference on Computer Vision and Pattern Recognition. 2024, 16965–16974 [Google Scholar]
  33. Khanam R. and Hussain M., YOLOV11: an Overview of the Key Architectural Enhancements, 2024 [Google Scholar]
  34. Jocher G. and Qiu J., Ultralytics YOLO11. https://github.com/ultralytics/ultralytics
  35. Vaswani A.; Brain G.; Shazeer N.; Parmar N.; Uszkoreit J.; Jones L.; Gomez A. N.; Kaiser Ł.; Polosukhin I.Attention Is All You Need, 31st Conference on Neural Information Processing Systems, NIPS, 2017, 10.48550/arXiv.1706.03762 [DOI] [Google Scholar]
  36. Wang S. Xia C. Lv F. Shi Y. RT-DETRv3: Real-Time End-to-End Object Detection with Hierarchical Dense Positive Supervision. Proc. - 2025 IEEE Winter Conf. Appl. Comput. Vision, WACV. 2025:1628–1636. doi: 10.1109/WACV61041.2025.00166. [DOI] [Google Scholar]
  37. Khan M. S. Wani A. A. Jan Q. Bhat M. M. Kuchey M. Y. Sheikh U. R. Sofi F. A. Bhat M. A. Smartphone-Assisted Contact Angle Measurements: A Simple Approach to Determine the Critical Micelle Concentration of Surfactants. J. Chem. Educ. 2024;101(3):1198–1203. doi: 10.1021/ACS.JCHEMED.3C00953. [DOI] [Google Scholar]
  38. Mosmann T. Rapid Colorimetric Assay for Cellular Growth and Survival: Application to Proliferation and Cytotoxicity Assays. J. Immunol. Methods. 1983;65(1–2):55–63. doi: 10.1016/0022-1759(83)90303-4. [DOI] [PubMed] [Google Scholar]
  39. Kumar P. Nagarajan A. Uchil P. D. Analysis of Cell Viability by the Lactate Dehydrogenase Assay. Cold Spring Harb. Protoc. 2018;2018(6):465–468. doi: 10.1101/pdb.prot095497. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

RA-OLF-D6RA05448J-s001
RA-OLF-D6RA05448J-s001.pdf (2.6MB, pdf)

Data Availability Statement

Characterization for compounds: all 1H, 19F, and 13C{1H} NMR spectra are available in the supporting information document and have been deposited in the Harvard Dataverse Database (DOI: https://doi.org/10.7910/DVN/77ZUIW). Biophysical evaluation of compounds: the datasets generated and analyzed during the current study, including GUV microscopy images, contact angle droplet photographs, and raw dynamic light scattering (DLS) values, are available upon request. YOLOv11 and RTDetR code: all code involved in model training and GUV quantification inferences, as well as external links to model parameters, are actively maintained in the following GitHub repository: https://github.com/josephlaurienzo-sudo/guv-detection.git.

Supplementary information (SI): key experimental, spectroscopic and characterization data for novel compounds, and supplementary figures. All code use for the processing and preparation of data is available on GitHub. See DOI: https://doi.org/10.1039/d6ra05448j.

Articles from RSC Advances are provided here courtesy of Royal Society of Chemistry

RESOURCES