Repurposing Alzheimer’s and ovarian cancer drugs as sonosensitizers for glioblastoma via a positive-unlabeled learning and 3D bioprinting-based new approach methodology (NAM)

Rudrajit Majumder; Priyankan Datta; Sreejesh Moolayadukkam; Ishwar K. Puri

doi:10.1371/journal.pone.0354981

Abstract

Glioblastoma (GBM) remains a lethal primary brain tumor, in part because therapeutic efficacy is limited by the blood–brain barrier (BBB) and the complex tumor microenvironment (TME). Sonodynamic therapy (SDT), i.e., use of ultrasound to activate chemical sensitizers and generate cytotoxic stress, offers a non-invasive strategy for treating deep-seated intracranial disease, but progress is constrained by the scarcity of validated sonosensitizers and the inefficiency of conventional in vitro screening methods. Here, we introduce a New Approach Methodology (NAM) that couples a neural network-based positive-unlabeled (PU) learning framework with a high-throughput, magnetic field–guided 3D bioprinting platform to accelerate identification and experimental validation of SDT-sensitizing agents. Using curated drug and small-molecule data and RDKit-derived molecular descriptors, the PU classifier identifies candidate ultrasound-responsive compounds without requiring reliable negative labels. We then validate the AI-based predictions in physiologically relevant U-87 MG glioblastoma spheroids that reproduce key TME features, including spatial heterogeneity and a hypoxic core. The NAM identifies two FDA-approved drugs, carboplatin (advanced ovarian cancer) and memantine hydrochloride (Alzheimer’s disease), as effective ultrasound-responsive agents. In 3D spheroids, combining low-intensity pulsed ultrasound with either drug significantly reduces viability compared with drug-only controls, and both combinations outperform temozolomide (TMZ), the current standard chemotherapeutic. Time-resolved responses reveal distinct kinetics: memantine produces strong early cytotoxicity (24 h) enhanced by ultrasound, whereas carboplatin shows delayed but pronounced cytotoxicity (72 h), also improved by ultrasound. Together, these results establish an integrated computational–experimental NAM that enables rapid repurposing of approved drugs as SDT sensitizers and provides a scalable framework for advancing GBM therapeutic discovery while reducing reliance on animal studies.

Citation: Majumder R, Datta P, Moolayadukkam S, Puri IK (2026) Repurposing Alzheimer’s and ovarian cancer drugs as sonosensitizers for glioblastoma via a positive-unlabeled learning and 3D bioprinting-based new approach methodology (NAM). PLoS One 21(8): e0354981. https://doi.org/10.1371/journal.pone.0354981

Editor: Michael C. Burger, Goethe University Hospital Frankfurt, GERMANY

Received: February 25, 2026; Accepted: July 15, 2026; Published: August 12, 2026

Copyright: © 2026 Majumder et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: The complete curated database of small molecules and their RDKit-computed descriptors can be found in the Supplementary Information (S2 Table), and the Python codes for the machine learning workflow are included in S1 Text.

Funding: The author(s) received no specific funding for this work.

Competing interests: The authors have declared that no competing interests exist.

1. Introduction

Glioblastoma multiforme (GBM) is the most common and lethal primary brain tumor, with a five-year survival rate of merely 5.5% [1]. Current first-line treatment for newly diagnosed GBM involves maximal safe surgical resection followed by radiotherapy and adjuvant chemotherapy [2]. Yet despite these aggressive interventions, GBM invariably recurs due to its highly invasive nature, resulting in substantial treatment-related morbidity. Conventional chemotherapeutic agents encounter formidable physicochemical barriers within the tumor microenvironment (TME) that severely compromise their efficacy. Additionally, the blood-brain barrier (BBB), composed of tightly sealed endothelial cells, presents a critical obstacle by preventing effective transport of therapeutic molecules from the systemic circulation to the tumor site [3]. This dual resistance mechanism highlights the critical need for novel therapeutic strategies capable of circumventing these barriers and improving clinical outcomes.

These challenges have generated interest in sensitizer-based therapies, particularly photodynamic therapy (PDT). PDT utilizes light-activated small molecules to generate cytotoxic reactive oxygen species (ROS), facilitating targeted tumor destruction [4,5]. While PDT is effective for superficial lesions, adequate light penetration to treat deep-seated tumors is a significant challenge. To address this limitation, therapeutic ultrasound-mediated sonodynamic therapy (SDT) is a promising alternative for deep-seated tumors, particularly those within the central nervous system. SDT employs low-intensity pulsed ultrasound (LIPU) to convert acoustic energy into light through sonoluminescence, effectively activating responsive sensitizers and enabling non-invasive access to intracranial targets. LIPU operates at substantially lower energy levels than conventional ablative techniques, thereby minimizing patient risk. Furthermore, the pulsatile mechanical forces generated by ultrasound can induce transient cellular membrane disruptions, including reversible blood-brain barrier opening [6], which enhances therapeutic agent delivery directly to the tumor site and amplifies treatment efficacy.

However, the therapeutic efficacy and clinical translation of SDT depend primarily on the molecular and electronic properties of the sensitizers [7]. Despite growing understanding of how these features govern photo- and sonosensitivity, predictive design rules are yet poorly defined. Additionally, sensitizer screening using traditional 2D in vitro cell culture models frequently yields poor clinical translation, as these systems fail to recapitulate the complex in vivo tumor microenvironment. Meanwhile, animal models are costly, time-intensive, and raise ethical concerns. As a result, 3D spheroid models are emerging as a gold standard in vitro platform, providing a more physiologically relevant bridge between 2D culture systems and animal models [8].

For scientific, financial, and ethical reasons, there has been a paradigm shift toward developing and validating human-relevant “new approach methodologies” (NAMs) that can reduce and replace animal use in drug research and development [9]. A NAM typically integrates two core components [10]: human-relevant in vitro systems and in silico modeling techniques, including artificial intelligence and machine learning. While NAMs have been widely employed for toxicity screening [11], their application in cancer drug development remains relatively limited.

Advances in artificial intelligence (AI) and cheminformatics enable the classification of molecules based on their physicochemical properties [12]. AI-based sensitizer screening offers a faster and less expensive approach to expedite preclinical-to-clinical translation for SDT. A critical stage during sensitizer development is the accurate classification and prediction of its activity. Machine learning (ML) approaches can categorize sensitizers, for instance, into Type I (free radical generators) and Type II (singlet oxygen generators) based on their molecular properties and ROS generation mechanisms [13]. Predictive models can forecast the singlet oxygen quantum yield of Type II photosensitizers, a key determinant of their therapeutic efficacy [14–16]. While many models rely on Density Functional Theory (DFT) for descriptors [17], DFT’s computational costs and limited biological realism constrain its utility.

Sensitizer discovery suffers from a lack of reliable negative examples, as limited experimental testing means that evidence confirming whether molecules function as sonosensitizers is often unavailable. Standard supervised learning methods, which classify all unlabeled compounds as negatives, are susceptible to false-negative propagation and reduced predictive accuracy.

Positive-unlabeled (PU) learning is a powerful alternative to overcome these constraints [18,19]. Unlike traditional binary classification, PU learning does not assume that unlabeled instances are negative. Instead, it models the unlabeled set as a mixture of hidden positives and true negatives, enabling more accurate and realistic inference. PU learning algorithms help identify candidate molecules without using confirmed negative examples, thereby broadening the scope of discovery and mitigating biases due to mislabeling or incomplete datasets. [20].

Here, we present a NAM that integrates PU learning with physiologically relevant 3D bioprinted tumor models to accelerate the discovery of sensitizers, as shown in Fig 1. To inform sonosensitizer selection, we compared two neural network architectures—a graph neural network (GNN) and a standard deep neural network (DNN)—and found that the standard neural network yielded superior predictive accuracy, leading to its selection for final predictions. This AI-driven model guides the in vitro evaluation of sensitizers using ultrasound application on tumor spheroids derived from glioblastoma. This scalable platform combines deep learning with high-throughput 3D bioprinting, enabling the rapid identification and validation of ultrasound-responsive sensitizers.

Download:

Fig 1. Schematic overview of the proposed New Approach Methodology (NAM).

A positive–unlabeled deep neural network (PU-DNN) serves as the computational backbone, classifying drugs and small molecules according to their predicted ultrasound-responsive sonosensitizer activity. Shortlisted candidates are subsequently validated through a high-throughput, magnetic field-assisted 3D cancer spheroid bioprinting platform, providing physiologically relevant in vitro confirmation of computational predictions. Applying this integrated in silico-in vitro approach for GBM, the NAM successfully identifies two US FDA-approved drugs that can be repurposed as viable sonosensitizers for sonodynamic therapy, one molecule indicated for Alzheimer’s disease and the other for advanced ovarian cancer.

https://doi.org/10.1371/journal.pone.0354981.g001

Our recent work demonstrated that the FDA-approved chemotherapeutic drug for glioblastoma treatment, Temozolomide (TMZ), can be activated using LIPU [21]. This finding suggests that approved drugs may be repurposed as sonosensitizers, enhancing their therapeutic effectiveness by reducing off-target toxicity, lowering required doses, and enabling localized ROS-mediated tumor killing through ultrasound activation.

We investigate the ultrasound responsiveness of two FDA-approved drugs, carboplatin and memantine hydrochloride, which are currently being used treat advanced ovarian cancer and Alzheimer’s disease, respectively. Their efficacies are compared against TMZ, the current standard of care for GBM. Our AI model predicts that both carboplatin and memantine hydrochloride exhibit ultrasound responsiveness. To validate this AI-based prediction, we generate 3D tumor spheroids using a magnetic field-guided cell printing technology [21,22] with the glioblastoma cell line U-87 MG. The tumor spheroids are treated with carboplatin, memantine hydrochloride, or TMZ in combination with LIPU. Both carboplatin and memantine hydrochloride demonstrate superior efficacy compared to TMZ and, when combined with ultrasound, LIPU reduces both treatment duration and required dosage. Collectively, the findings provide compelling evidence that this NAM serves as a powerful platform for screening high-potential drug candidates and validating their therapeutic efficacy in sonodynamic therapy (SDT) for glioblastoma.

2. Materials and methods

2.1. Reagents and chemicals

Temozolomide (98%, molecular weight 194.15 g/mol, T2577-100MG), GLPBIO Cell Counting Kit-8 (CCK-8), Promega CellTiter-Glo (R) 3D cell viability assay (G9681), Gibco^TM DPBS (14040133), Gibco^TM DMEM (11995040), Trypan Blue (0.4%, Thermofisher Scientific, USA), Invitrogen ready probes cell viability imaging kit, Blue/Green (R37609), JC-1 dye (T3168), CellEvent^TM Caspase 3/7 green ready probes^TM (R37111) are purchased from Fisher Scientific, USA, and Dimethyl Sulfoxide (for molecular biology, D8418), Millipore Sigma, USA. Trypsin-EDTA solution (1X, 30-2101) is purchased from ATCC, USA. Carboplatin and memantine hydrochloride are purchased from Millipore Sigma, USA.

2.2. Cell line and cell culture

U-87 MG cell lines are purchased from ATCC, USA. Tumor cells are cultured in a complete medium of Dulbecco’s Modified Eagle Medium, supplemented with 10% Fetal Bovine Serum (Thermofisher Scientific, USA, Cat. No. A5256801), 1% Penicillin-Streptomycin (Fisher Scientific, USA), and maintained at 5% CO₂ and 37 ⁰C in vitro.

2.3. Glioblastoma tumor spheroid printing

Glioblastoma tumor spheroids are printed using a magnetic-field-guided 3D cell printing technology [22], following the protocol described in our earlier work [21]. Briefly, cells are harvested and counted before printing using an uncoated, flat, square 96-well plate (Ibidi, Cat. No. 89621) placed on a cube magnet array (4.5 mm × 4.5 mm × 4.5 mm). The harvested cells are added to the wells (10,000 cells/well) containing a paramagnetic salt (Gadovist®) solution diluted into the culture medium. The magnetic susceptibility difference between the diamagnetic cells and their surrounding paramagnetic medium exerts a net driving force on each cell, causing them to be displaced towards the region of the lowest magnetic field. This aggregates cells into a layer-by-layer 3D structure within a few hours, depending on cell type.

2.4. Ultrasound stimulation and experimental setup

The experimental setup for ultrasound stimulation and the procedure for SDT experiments are similar to those in earlier work [21]. Briefly, 3D glioblastoma spheroids are incubated for 24 hours after printing. On Day 2, the IC₅₀ and IC₂₅ concentrations of carboplatin, memantine hydrochloride, and TMZ are added and incubated for an additional 4 hours before ultrasound stimulation [21]. The experimental design includes six groups per drug to systematically dissect the sonodynamic contribution: (i) untreated control, (ii) ultrasound alone (US), (iii) drug alone at IC₅₀, (iv) US combined with drug at IC₅₀, (v) drug alone at IC₂₅, and (vi) US combined with drug at IC₅₀. The US-alone group serves as a reference for the intrinsic mechanical cytotoxicity of LIPU, which arises from acoustic cavitation, radiation pressure, and activation of mechanosensitive ion channels even in the absence of an exogenous sensitizer. The drug-attributable enhancement of ultrasound-mediated killing is therefore quantified by comparing each US + drug group against both the corresponding drug-alone and US-alone groups. An ultrasound transducer (Delta Digi Sound) is mounted on a 3D-printed (Prusa MK4) stage, built in-house using polylactic acid. The transducer head (3 cm diameter) is placed below the well plate, and an ultrasonic transmission gel is applied between them to minimize the acoustic impedance mismatch. The ultrasound stimulation parameters for all experiments are 1 MHz, 1 W/cm², 20% Duty factor, 200 milliseconds pulse duration, and 5 minutes of stimulation duration. The spheroid viability for each cell line is evaluated 24 hours after ultrasound stimulation through ATP release using the Promega CellTiter-Glo® 3D cell viability assay, similar to that described in Section 2.6.

2.5. Confocal imaging

Confocal images are taken 24 hours after cell seeding. For this purpose, printed spheroids are stained with NucBlue® Live reagent (Hoechst 33342) and NucGreen® Dead reagent following the manufacturer’s protocol, and images are captured using the Leica microscope. Briefly, two drops of NucBlue® Live and NucGreen® Dead reagents are added to the spheroid culture medium and incubated at 5% CO₂ and 37 ⁰C for 30 minutes. The live reagent stains only the live cell nuclei and is detected with a standard DAPI filter (excitation/emission maxima: 360/460 nm). The dead reagent stains only the nuclei of dead cells with compromised plasma membranes and is detected with a standard FITC (green) filter set (excitation/emission maxima: 504/523 nm). The captured images are further processed using ImarisViewer 10.2.0.

2.6. Spheroid cell viability and IC₅₀ determination for carboplatin, memantine hydrochloride, and TMZ

IC₅₀ concentrations as well as cell viability following SDT for 3D cultures are measured by ATP release using a Promega CellTiter-Glo® 3D cell viability assay, as per the manufacturer’s protocol, which is similar to the procedure we previously employed [21]. Briefly, glioblastoma (U-87 MG) spheroids with 10,000 cells/well are printed and incubated for 24 hours. The drugs are added on Day 2 at different concentrations. Following 72 hours of incubation, the old medium is replaced, and the assay reagent is added at a 1:1 (v/v) ratio. The mixture is then incubated for 25 minutes in the dark. Finally, the supernatant from each treated group is collected into an opaque white tissue culture-treated 96-well plate (353296, Becton Dickinson Labware, USA) to minimize signal loss. Luminescence is measured over an integration time of 0.01 s using a plate reader (Synergy H4), and cell viability is evaluated using GraphPad Prism software. In all experiments, the molecules are dosed only once.

2.7. Oxidative stress determination

Oxidative stress following SDT across different treatment groups is determined using the Invitrogen^TM cellROX green reagent (Thermofisher Scientific, USA, Cat. No. C10444) according to the manufacturer’s protocol. Briefly, after spheroid printing and treatment, the old medium is carefully aspirated to minimize loss of structural integrity. The spheroids are washed with pre-warmed PBS after the addition of complete cell culture medium containing the assay reagent (5 μM). Spheroids are incubated for an additional 30 min in the dark at 37 °C and washed with PBS before taking confocal imaging using a FITC filter (495/520 nm) on a Leica microscope. The green fluorescence intensity change of the dye, due to its oxidation by reactive oxygen species, indicates oxidative stress generation across the groups.

2.8. Mitochondrial membrane potential

Mitochondrial membrane potential is evaluated after SDT experiments using the JC-1 dye in a 2D monolayer culture condition. Briefly, U-87 MG cells are seeded (10⁴ cells/well) into the tissue-culture-treated 96-well plate (Genesee Scientific, Cat. 25-109). 24 hours after drug addition at their IC₅₀ concentrations, followed by ultrasound stimulation, the old medium is replaced, and cells are washed with pre-warmed PBS. After adding the JC-1 dye (2 μM) to each well and incubating for 30 minutes in the dark at 37 °C, images are captured with the standard FITC (495/520 nm) and TRITC (560/590 nm) filter settings using a Leica microscope. In the polarized mitochondrial membrane, JC-1 exhibits an orange fluorescence signal due to its aggregation. However, a loss of membrane polarization led to the emission of a green fluorescence signal in its monomer form.

2.9. Cellular apoptosis assay

Cellular apoptosis is determined using the CellEvent^TM Caspase 3/7 green ready probes^TM reagent according to the manufacturer’s protocol. Briefly, 24 hours after SDT, the assay reagent is added to each treatment group and incubated for 45 minutes at 37 °C. The standard FITC filter (495/520 nm) is used in a Leica microscope to capture the green fluorescence signal, indicating caspase-3/7 activation in the apoptotic cells within the tumor spheroids.

2.10. Statistical analysis

Data are expressed as mean ± standard deviation. The one-way ANOVA method determines whether there are significant differences among groups; P < 0.05 is considered statistically significant. All experiments are performed at least in triplicate.

2.11. Ethics statement

This study does not involve any human participants or animal subjects. All experiments are conducted using the U-87 MG human glioblastoma cell line, which is a commercially available, immortalized cell line; therefore, it does not require Institutional Animal Care and Use Committee (IACUC) approval. The database utilized for deep learning frameworks does not include any human- or animal-specific experimental assay and is sourced from open repositories and literature.

3. Results

3.1. Database curation for AI model

We compile a database of small-molecule drugs from multiple public repositories. Antineoplastic and immunomodulating compounds are obtained from ChEMBL [23], while photosensitivity data are curated from PubChem [24] and DrugBank [25]. Molecules with confirmed photosensitive activity are labeled as positives, whereas those lacking such evidence are retained as unlabeled compounds since they cannot be confidently assigned to a negative class without experimental validation. Several organic sensitizers identified in our recent review of sonodynamic therapy are also incorporated into the dataset [7].

To build a predictive model, we characterize each compound through a comprehensive set of molecular descriptors computed using RDKit, a robust and widely adopted open-source cheminformatics toolkit well-suited for enumerating key molecular properties [26]. These descriptors represent critical structural and physicochemical characteristics, including molecular weight, topological polar surface area, logP (a measure of lipophilicity), number of hydrogen bond donors and acceptors, atom counts, and various topological indices, which collectively inform the potential behavior of a molecule in a biological context.

The final training database comprises 364 small-molecule drugs, for which the RDKit is used to generate 204 distinct molecular descriptors for each molecule, resulting in a feature-rich dataset for machine learning. Within this set, 175 molecules are positively labeled as sensitizers. The unlabeled set is considered a mixture of latent positive and negative instances, necessitating a positive-unlabeled (PU) learning framework for model development. Consequently, a trained model will serve as a screening tool to estimate the probability that a candidate molecule is ultrasound responsive when its molecular descriptors are input into the model. The complete curated database of small molecules and their RDKit-computed descriptors can be found in S2 Table.

3.2. Model development

We select neural network-based architecture as our classifier due to its capacity as a universal function approximator [27,28]. This property ensures that a multilayer feedforward network can approximate any function, making it highly suitable for learning complex, nonlinear relationships inherent in molecular data. Within this framework, we explore two distinct neural network architectures that leverage different representations of the input molecules.

The first architecture is a standard Deep Neural Network (DNN), which operates on fixed-length feature vectors. For each molecule, the DNN uses 204-dimensional molecular descriptors computed using RDKit, providing a numerical representation of molecular structure and substructure presence. The DNN learns a mapping from these feature vectors to a prediction of sonosensitizer activity.

The second architecture is a Graph Neural Network (GNN), which operates directly on the molecular graph representation. Unlike the DNN, which receives precomputed descriptors, the GNN consumes 20 node features (atom properties) and 4 edge features (bond properties). MolGraph, a recently released Python library, is used to encode each molecule into graph tensors containing those node and edge features [29]. This architecture is explored under the hypothesis that learning directly from the molecular graph can capture meaningful structural patterns within a more compact feature space, potentially reducing dimensionality while retaining chemically relevant information.

In a standard supervised binary classification task, the goal is to learn a mapping from an input feature matrix to a discrete output , where . The model learns this mapping by adjusting its internal parameters, weights () and biases (), to minimize a loss function that quantifies the discrepancy between the predicted probabilities () and the true labels ().

For a conventional binary classification problem with confirmed positive and negative samples, the appropriate loss function is the Binary Cross-Entropy (BCE). The BCE loss measures the divergence between the true label distribution and the predicted probabilities. For a single data point, it is defined as:

(1)

For a dataset with M samples, this is averaged to form the total loss:

(2)

Optimization of the model parameters () is achieved via gradient descent, an iterative algorithm that minimizes the loss function [30]. The gradients of the loss with respect to each parameter are computed through the backpropagation algorithm, which applies the chain rule to propagate the error backward through the network layers [31]. This enables precise weight and bias updates, guiding the model toward a state that accurately classifies the input data.

Our specific problem lacks reliable negative samples () since the data consist only of confirmed sensitizers (positive samples, ) and non-confirmed molecules (unlabeled samples, ). The unlabeled set contains a mixture of both hidden positive samples and true negative samples. Directly applying BCE loss is inappropriate because it would incorrectly treat all unlabeled samples as negatives.

The connection between BCE and PU loss is derived from a risk minimization perspective. The expected BCE risk for a classifier,

(3)

where is the class prior (the proportion of true positives in the entire dataset), and and denote expectations over the true positive and negative distributions, respectively. Since there are no explicit negative samples (), the risk over the negative distribution is represented as a risk over the unlabeled and positive distributions.

The approximate relationship among these risks that yields the unbiased PU (uPU) risk estimator [32],

(4)

Using the BCE loss components and , the empirical uPU loss for training,

(5)

This loss is unbiased, implying it is equivalent to training on fully labeled data. However, in practice, once the classifier becomes too confident in the unlabeled set, a common occurrence in overfitting, this leads to negative loss values and severe overfitting.

To address the instability of the unbiased PU loss, the non-negative PU (nnPU) loss has been proposed with a simple, yet critical, modification [32]. If the empirical risk term for the unlabeled data becomes smaller than the risk term for the positives, it is clipped to zero. This prevents the loss from becoming negative and stabilizes the training process.

The nnPU loss is defined as:

(6)

The operation ensures the loss component from the unlabeled data is non-negative, making the training process significantly more robust.

The PU learning framework requires an accurate estimate of the class prior before training. The labeled positive set is selected completely at random (SCAR) from the entire set of positive instances [33], The SCAR assumption introduces a key scalar constant , defined as the probability that a positive sample () is labeled (), i.e.,

(7)

Physically, c represents the labeling frequency. It is the fraction of true sensitizers that have been correctly identified and included in the positive set . A low c value indicates that many sensitizers have been missed and remain unrecognized (and they reside in ), while a higher value suggests the positive set is relatively complete.

The core idea is to train a probabilistic classifier, (e.g., Random Forest), to distinguish the labeled positive set from the unlabeled set . This classifier learns to model the probability , i.e., the probability that a given sample x is in the labeled set.

The value of c is determined using the trained classifier g and a validation set of labeled positives . The estimator is the average value of for all :

(8)

where is the number of labeled positive samples in the validation set. The class prior = and the labeling frequency c are closely related to each other [34]. Given a PU dataset, if one is known, the expected value of the other can be determined as follows,

(9)

(10)

In practice is simply the empirical proportion of labeled positives in the entire dataset, . Therefore, the class prior can be estimated as,

(11)

We estimate the class prior to be 0.5454, indicating a 54.54% probability that a randomly selected molecule from the database is a true sensitizer. Fig 2a shows the class distributions. With π determined, we proceeded to train our DNN and GNN. The nnPU loss function is applied identically to both architectures, enabling a fair comparison of their respective capabilities.

Download:

Fig 2. Data distribution, model development, validation, and application for sonosensitizer prediction.

(a) Pie chart illustrating the inferred dataset composition based on the Elkan & Noto class prior ( = 0.5454), decomposing the unlabeled set into hidden positives and probable negatives to reveal the underlying PU learning scenario. (b) Training and validation loss curves of the PU-DNN model across 500 epochs, demonstrating stable convergence without significant overfitting. (c) Training and validation loss curves of the PU-GNN model across 500 epochs, revealing notable divergence and consistently higher loss values relative to the PU-DNN. (d) Sensitivity analysis of the class prior () for the PU-DNN model, reporting recall across a range of α values (mean ± standard deviation, n = 5 runs).

https://doi.org/10.1371/journal.pone.0354981.g002

We refer to the DNN-based model as PU-DNN (Positive-Unlabeled Deep Neural Network) and the GNN-based model as PU-GNN (Positive-Unlabeled Graph Neural Network). The architecture of PU-DNN consists of an input layer followed by three hidden layers with 64, 16, and 8 neurons, respectively. In contrast, the PU-GNN architecture starts with a graph convolutional layer, constructed using MolGraph [29], equipped with 64 channels to process the molecular graphs. This is followed by two standard hidden layers containing 16 and 8 neurons, respectively.

Both models include a single-neuron output layer. For both models, each hidden layer is equipped with a Rectified Linear Unit (ReLU) activation function to introduce non-linearity, while the output layer employs a sigmoid activation to constrain predictions to the probabilistic range [0, 1]. To mitigate overfitting, both architectures incorporate kernel regularization at each hidden layer, along with dropout layers following each hidden layer. The models are compiled using the Adam optimizer with a learning rate of 0.001, and the dataset is partitioned into training (70%), validation (15%), and test (15%) sets, and stratified to preserve class distributions in all splits. Implemented in TensorFlow [35], the models are trained for 500 epochs. The convergence of the training and validation loss for PU-DNN and PU-GNN is depicted in Fig 2b. and Fig 2c, respectively.

Following training, the models generate predicted probabilities for each molecule in the test set. A decision threshold is applied such that molecules with predicted probability are classified as positive (sensitizers), while those with are classified as negative. Model performance is evaluated using standard classification metrics such as precision, recall, and F-score. Precision quantifies the accuracy of the positive predictions, and is defined as the proportion of molecules classified as positive that are truly positive:

(12)

Where represents true positives and represents false positives.

Recall, on the other hand, quantifies the model’s ability to identify all positive samples, and is defined as the proportion of true positives that are correctly identified:

(13)

Where represents false negatives.

The F-score provides a balanced measure by taking the harmonic mean of precision and recall, is defined as:

(14)

In a conventional supervised setting, the decision threshold is typically selected by maximizing a desired metric, such as the F-score. However, in the PU learning setting, precision is not directly computable because the unlabeled set contains both true positives and true negatives, and the true labels for unlabeled samples are unknown [34]. To address this limitation, Lee and Liu [36] proposed a performance metric that approximates the behavior of the F-score while remaining computable within the PU framework. Following their work, this metric, which we denote as , is defined as:

(15)

We set the for each model by maximizing . Using this threshold, model performance is evaluated on the held-out test set, with results summarized in Table 1. The comparison reveals that PU-DNN consistently outperforms PU-GNN across all evaluated metrics, demonstrating superior predictive accuracy for sonosensitizer identification. Consequently, we select the PU-DNN model for the downstream prediction of sonosensitivity for candidate compounds, carboplatin, and memantine hydrochloride.

Download:

Table 1. Performance comparison of PU-DNN and PU-GNN models.

https://doi.org/10.1371/journal.pone.0354981.t001

It is often necessary to assess the model’s sensitivity to variations in the class prior [37]. To evaluate the robustness of the PU-DNN model, we explore its sensitivity by varying from 0.2 to 0.8. We retrain the model five times for each value and record the resulting recall on the test set. The model shows a sensitivity of 7.7% to changes in the class prior, indicating reasonable stability (Fig 2d).

Our PU-DNN model predicts carboplatin and memantine hydrochloride as sonosensitive with prediction confidences of 99.45% and 98.35%, respectively. The Python programs for the deep learning workflow are included in the supporting information (S1 Text).

3.3. Characterization of the glioblastoma spheroid 24 hours after printing

Glioblastoma spheroids are printed using U-87 MG cells with the high-throughput, magnetic field-guided 3D bioprinting method and incubated for 24 hours. Observation shows that (Fig 3a) spheroids have an average diameter of 306.7 ± 6.2 μM.

Download:

Fig 3. U-87 MG spheroid characterization.

(a) brightfield image, and (b) confocal image stained with NucBlue® Live reagent (Hoechst 33342) and NucGreen® Dead. Spheroids have an average diameter of 306.7 ± 6.2 μM after 24 hours of printing. Further observation reveals the formation of a hypoxic tumor core, demonstrating the reliability and physiological relevance of our printed spheroids.

https://doi.org/10.1371/journal.pone.0354981.g003

Confocal images captured after staining the spheroids with NucBlue® Live reagent (Hoechst 33342) and NucGreen® Dead show that a central hypoxic core forms, which is surrounded by a proliferative outer cell layer (Fig 3a and Fig 3b). The spatial heterogeneity is a hallmark of tumor physiology that resists standard chemotherapy, which also establishes the reliability of the 3D bioprinting method.

3.4. IC₅₀ for carboplatin, memantine hydrochloride, and TMZ

The effects of carboplatin, memantine hydrochloride, and TMZ on glioblastoma spheroids derived from the U-87 MG cell line are evaluated by determining the IC₅₀ value for each drug in the absence of ultrasound stimulation. These drugs are added at varying concentrations to separate U-87 MG spheroids 24 hours after they are formed, and their IC₅₀ values are determined after 72 hours of incubation by measuring the total ATP as described in Section 2.6.

Memantine hydrochloride has the lowest IC₅₀ value (131 μM) among the three drugs. A higher IC₅₀ value for TMZ (804.5 μM) compared to carboplatin (252.3 μM) indicates greater chemotherapeutic resistance to it in the U-87 MG spheroids (Fig 4).

Download:

Fig 4. Comparison of IC₅₀ values for glioblastoma (U-87 MG) spheroids treated with carboplatin, memantine hydrochloride, and TMZ.

The IC₅₀ values determined 72 hours after treating with (a) carboplatin (252.3 μM), (b) memantine hydrochloride (131 μM), and (c) TMZ (804.5 μM). Memantine hydrochloride has the lowest IC₅₀ value among carboplatin and TMZ.

https://doi.org/10.1371/journal.pone.0354981.g004

ADMET predictions (S1 Table) for carboplatin, memantine hydrochloride, and TMZ are also obtained from ADMET-AI [38] and ADMETlab 3.0 [39]. The data reveal that memantine hydrochloride has the lowest drug-induced liver injury and hematotoxic value (related to the bone marrow damage) among the three drugs. It also has the highest lipophilicity (3.116) and the lowest topological polar surface area (26.02 ), which increases its likelihood of crossing the blood-brain barrier. Memantine is a CNS-active drug with a documented BBB-permeability mechanism [40] consistent with our ADMET prediction (BBB crossing probability = 0.988, S1 Table). Additionally, studies by Carpentier et al. [41], Idbaih et al. [42], and Sonabend et al. [43] revealed enhanced BBB penetration of carboplatin in conjunction with LIPU. These pharmacokinetic and toxicity data for memantine hydrochloride, along with its lowest IC₅₀ value establish its greater potency over carboplatin and the standard chemotherapeutic drug for GBM, TMZ.

3.5. Sonodynamic therapy (SDT) on glioblastoma spheroid treated with carboplatin, memantine hydrochloride, and TMZ

We assess the sonosensitizer potential of carboplatin (an ovarian cancer drug) and memantine hydrochloride (an Alzheimer’s drug) in comparison with TMZ, using U-87 MG spheroids treated with each drug individually, with or without ultrasound stimulation. For sonodynamic treatment, drugs are introduced 24 hours after spheroid printing at their IC₅₀ and IC₂₅ concentrations and incubated for 4 hours prior to ultrasound exposure (1 MHz, 1 W/cm², 20% duty cycle, 5 minutes). Cell viability is subsequently evaluated at 24 and 72 hours post-treatment using the ATP-based Promega CellTiter-Glo® 3D assay (Section 2.6).

At 24 hours without ultrasound stimulation, memantine demonstrates the highest cytotoxicity on U-87 MG spheroids (Fig 5c, lowest cell viability, 69.5 ± 9.9%), while carboplatin exhibits the lowest cytotoxicity (Fig 5a), with TMZ showing intermediate effects (Fig 5e). The cytotoxic effect of memantine is markedly enhanced with ultrasound stimulation (Fig 5c, cell viability 36.7 ± 13.7%) compared to spheroids treated with carboplatin (Fig 5a, cell viability 66.6 ± 14.6%) or TMZ (Fig 5e, 51 ± 7.1%). This pattern is consistent across all drug dosages.

Download:

Fig 5. Viability of glioblastoma (U-87 MG) spheroids in the presence of carboplatin (CPT), memantine hydrochloride (M), and TMZ (T) after 24 and 72 hours of ultrasound (US) stimulation.

U-87 MG spheroid viability following ultrasound stimulation in the presence of carboplatin (a–b), memantine hydrochloride (c–d), and TMZ (e–f), assessed at 24 hours (a, c, e) and 72 hours (b, d, f) post-stimulation. At 24 hours, memantine produces the greatest cytotoxicity, which is further enhanced by ultrasound stimulation, while carboplatin demonstrates the weakest effect and TMZ shows intermediate efficacy. At 72 hours, however, memantine’s cytotoxic effect plateaus with little improvement over its 24-hour performance. Carboplatin, by contrast, emerges as the most cytotoxic agent at this timepoint regardless of ultrasound stimulation, while TMZ exhibits the lowest overall efficacy among all drugs tested. This pattern is consistent across all drug concentrations. Statistical comparisons are performed using one-way ANOVA (n = 4; *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001).

https://doi.org/10.1371/journal.pone.0354981.g005

At 72 hours, memantine-mediated cytotoxicity shows no significant improvement over its 24-hour performance, regardless of ultrasound stimulation (Fig 5d). Carboplatin, by contrast, emerges as the most cytotoxic agent at this timepoint (Fig 5b), reducing spheroid viability to 44.9 ± 1.6% without ultrasound and 23.2 ± 6.2% with ultrasound stimulation. TMZ remains the least effective treatment at 72 hours (Fig 5f), irrespective of ultrasound stimulation. This trend is consistent at the lower IC₂₅ dosage across all three drugs, mirroring the pattern observed at 24 hours.

4. Discussion

Carboplatin is an FDA-approved drug for advanced ovarian cancer treatment. The pharmacological and molecular properties indicate that it is hydrophilic (log P = −1.012) and has the highest molecular weight (371.249 g/mol) among memantine hydrochloride (log P = 3.116, Mw = 215.76 g/mol) and TMZ (log P = −0.587, Mw = 194.15 g/mol). These characteristics result in delayed carboplatin accumulation within the cellular cytoplasm via active transport and passive diffusion in the absence of ultrasound stimulation (Fig 6a). Following accumulation, carboplatin undergoes hydrolysis, becoming positively charged [44] and thereby forming a covalent bond with the N7 atom of the DNA purine base. This forms DNA adducts, which ultimately arrest the cell cycle in G2/M phases and trigger apoptosis [40]. The delayed carboplatin accumulation and its mechanism of action account for its lowest potency at 24 hours (Fig 5a) on U-87 MG spheroids.

Download:

Fig 6. Schematic representation of the working mechanisms for carboplatin, memantine hydrochloride, and TMZ with and without ultrasound application.

(a) The hydrophilic drug carboplatin enters the cellular cytoplasm via active and passive transport in the absence of ultrasound, subsequently arresting the cell cycle and triggering cellular apoptosis. (b) Ultrasound application enhances sonoporation-driven drug transport and ROS generation, ultimately increasing therapeutic efficacy. (c) Memantine, being lipophilic, rapidly binds to and blocks NMDA receptors in the absence of ultrasound. This reduces Ca²⁺ influx into the cellular cytoplasm, thereby inhibiting cell proliferation driven by excess glutamate secretion from GBM cells. (d) With ultrasound, mechanosensitive ion channels open, causing Ca²⁺ overload and ROS-mediated oxidative stress that leads to cancer cell death. This results in higher therapeutic efficacy compared to conditions without ultrasound. (e) TMZ, a lipophilic chemotherapeutic drug, diffuses rapidly into the cellular cytoplasm via active and passive transport pathways without ultrasound stimulation. At physiological pH (>7.0), TMZ hydrolysis generates the reactive methyldiazonium ion, which methylates the O⁶ position of guanine and induces DNA damage. (f) In response to ultrasound stimulation, sonoporation further enhances TMZ diffusion and triggers ROS-mediated oxidative stress generation, ultimately improving therapeutic efficacy compared to non-stimulated conditions.

https://doi.org/10.1371/journal.pone.0354981.g006

Ultrasound stimulation enhances carboplatin transport (Fig 6b) due to transient opening of the cell membrane [45]. Furthermore, carboplatin activation in response to ultrasound stimulation induces oxidative stress and loss of mitochondrial membrane potential (S1 and S2 Figs) via reactive oxygen species (ROS) generation, triggering cellular apoptosis (S3 Fig), which results in reduced cell viability compared to unstimulated conditions (Fig 5a). However, at 72 hours, enhanced carboplatin accumulation in the cytoplasm produces the highest cytotoxicity (i.e., lowest cell viability) among all three drugs, with or without ultrasound stimulation, and regardless of dosage (Fig 5b).

Memantine, on the other hand, is an FDA-approved drug for Alzheimer’s treatment. It is highly lipophilic and has a lower molecular weight than carboplatin. Therefore, memantine can block NMDA receptors on the U-87 MG cell membrane [46] more rapidly (Fig 6c). Blocking these receptors reduces NMDA receptor-associated signaling, such as glutamate-mediated autocrine/paracrine signaling [47,48], thereby inhibiting GBM cell proliferation. The experiments also reveal that within 24 hours of memantine addition, it exhibits the highest cytotoxic effect on U-87 MG spheroids, even without ultrasound (Fig 5c). Application of ultrasound (Fig 6d) opens mechanically activated piezo-1 channels [7], leading to Ca⁺² overloading followed by loss of the mitochondrial membrane potential and cellular apoptosis (S2 and S3 Figs). This ultimately reduces spheroid cell viability compared to the unstimulated condition (Fig 5c). However, memantine does not show a significant change in potency after 72 hours, regardless of ultrasound stimulation or dosage (Fig 5d).

TMZ is the current standard of care chemotherapeutic drug for GBM treatment. Its lipophilicity is intermediate between memantine and carboplatin. Under physiological conditions (pH > 7.0), TMZ undergoes hydrolysis to produce an active metabolite, MTIC, which further dissociates into AIC and the reactive methyldiazonium ion (Fig 6e). The methyldiazonium ion enters the GBM cell nucleus and methylates the O⁶ position of DNA guanine bases. This alteration can lead to error accumulation by incorporating thymine instead of cytosine in the complementary strand opposite O⁶-MeG, resulting in double-strand breaks that trigger cellular apoptosis [49]. TMZ and its first derivative MTIC have a half-life of about 2 hours [50]. Our experiments also demonstrate that TMZ exhibits minimal change in potency at 72 hours compared to 24 hours, regardless of ultrasound stimulation (Fig 5e and Fig 5f). We find that at 24 hours, the TMZ-mediated cytotoxic effect is more pronounced with ultrasound stimulation than without it (Fig 5e), which is attributed to enhanced TMZ transport into the cellular cytoplasm via sonoporation (Fig 6f), followed by its singlet-to-triplet state transition (= 2.1 eV) and cytotoxic singlet oxygen-mediated oxidative stress [7,21]. This is further corroborated by our experimental observations of oxidative stress generation, mitochondrial membrane potential disruption, and apoptotic signaling induced by TMZ, both with and without ultrasound stimulation (S1-S3 Figs). Additionally, we evaluated the singlet-to-triplet transition of TMZ (= 2.1 eV) and the maximum spin-orbit coupling (39.94 cm^-1) using time-dependent density functional theory (TD-DFT) with the B97X-D3 functional implemented in ORCA.

5. Conclusions

This work establishes a scalable New Approach Methodology (NAM) that integrates positive–unlabeled deep neural network (PU-DNN)-based screening with high-throughput, magnetic field–guided 3D bioprinted glioblastoma spheroids to accelerate discovery and validation of sonodynamic therapy (SDT) sensitizers. By explicitly addressing two persistent bottlenecks, i.e., (i) lack of trustworthy negative labels in sensitizer datasets and (ii) limited physiological relevance of standard 2D screening platforms, this NAM provides a practical computational–experimental route for identifying clinically actionable candidates for GBM treatment.

Using PU learning on curated drug collections and molecular descriptors, the model prioritizes FDA-approved compounds with predicted ultrasound responsiveness. Experimental testing in U-87 MG spheroids, chosen to better capture TME-driven resistance mechanisms, confirms that carboplatin and memantine hydrochloride function as ultrasound-enhanced cytotoxic agents. Across doses and timepoints, both drug–ultrasound combinations reduce spheroid viability relative to drug-only treatment and compare favorably with TMZ. Importantly, the spheroid studies reveal distinct efficacy kinetics consistent with each drug’s physicochemical and mechanistic profile: memantine produces strong early effects that are further amplified by ultrasound, whereas carboplatin exhibits delayed but pronounced cytotoxicity that is also improved by ultrasound exposure. These differences underscore a key practical insight for SDT development, i.e., optimal sensitizer choice and dosing schedule may depend not only on activation by ultrasound but also on transport, accumulation, and downstream cell-death pathways within the 3D tumor architecture.

Beyond identifying two repurposable candidates, the broader contribution is methodological. The combined PU learning/3D-bioprinting workflow offers a generalizable template for sensitizer discovery in settings where labels are incomplete and biological context critically shapes therapeutic response. Future work should extend this framework to additional GBM models (including patient-derived spheroids), incorporate multi-cellular TME components, and systematically map ultrasound parameter spaces to maximize efficacy while preserving safety. Additionally, the AI model could be extended to jointly predict BBB permeability, ADMET liabilities, or GBM-specific IC₅₀ values. Collectively, this NAM advances SDT sensitizer discovery toward faster, more human-relevant preclinical translation and supports repurposing strategies that may shorten the path to improved GBM therapies.

Supporting information

S1 Table. ADMET properties for carboplatin, memantine hydrochloride, and TMZ.

https://doi.org/10.1371/journal.pone.0354981.s001

(PDF)

S2 Table. Small molecule database containing RDKit-computed descriptors.

https://doi.org/10.1371/journal.pone.0354981.s002

(XLSX)

S1 Fig. Oxidative stress determination 1 hour after SDT via staining the spheroids with cellROX green reagent.

https://doi.org/10.1371/journal.pone.0354981.s003

(PDF)

S2 Fig. Mitochondrial membrane potential (MMP) determination using JC1 dye 24 hours after SDT.

https://doi.org/10.1371/journal.pone.0354981.s004

(PDF)

S3 Fig. Cellular apoptosis determination via caspase 3/7 green ready probe staining 24 hours after SDT.

https://doi.org/10.1371/journal.pone.0354981.s005

(PDF)

S1 Text. Python programs for the deep learning workflow.

https://doi.org/10.1371/journal.pone.0354981.s006

(PDF)

References

1. Zhang H, Li Y. Temozolomide-derived therapeutic strategies to overcome resistance in glioblastoma. J Med Chem. 2025;68(18):18810–28. pmid:40913573
- View Article
- PubMed/NCBI
- Google Scholar
2. Fernandes C, Costa A, Osório L, Lago RC, Linhares P, Carvalho B, et al. Current standards of care in glioblastoma therapy. Exon Publications; 2017.
3. Löscher W, Potschka H. Drug resistance in brain diseases and the role of drug efflux transporters. Nat Rev Neurosci. 2005;6(8):591–602. pmid:16025095
- View Article
- PubMed/NCBI
- Google Scholar
4. Yuan H, Chong H, Wang B, Zhu C, Liu L, Yang Q, et al. Chemical molecule-induced light-activated system for anticancer and antifungal activities. J Am Chem Soc. 2012;134(32):13184–7. pmid:22830933
- View Article
- PubMed/NCBI
- Google Scholar
5. Li X, Kwon N, Guo T, Liu Z, Yoon J. Innovative strategies for hypoxic-tumor photodynamic therapy. Angew Chem Int Ed Engl. 2018;57(36):11522–31. pmid:29808948
- View Article
- PubMed/NCBI
- Google Scholar
6. McHale AP, Callan JF, Nomikou N, Fowley C, Callan B. Sonodynamic therapy: concept, mechanism and application to cancer treatment. Therapeutic Ultrasound. 2016:429–50.
- View Article
- Google Scholar
7. Datta P, Moolayadukkam S, Chowdhury D, Rayes A, Lee NS, Sahu RP, et al. Recent advances and future directions in sonodynamic therapy for cancer treatment. BME Front. 2024;2024:0080. pmid:39735354
- View Article
- PubMed/NCBI
- Google Scholar
8. Chatzinikolaidou M. Cell spheroids: the new frontiers in in vitro models for cancer drug validation. Drug Discov Today. 2016;21(9):1553–60. pmid:27371506
- View Article
- PubMed/NCBI
- Google Scholar
9. Mehta K, Maass C, Cucurull-Sanchez L, Pichardo-Almarza C, Subramanian K, Androulakis IP, et al. Modernizing preclinical drug development: the role of new approach methodologies. ACS Pharmacol Transl Sci. 2025;8(6):1513–25. pmid:40567279
- View Article
- PubMed/NCBI
- Google Scholar
10. Levine SL, Riter LS, Lagadic L, Bejarano AC, Burden N, Burgoon LD, et al. Challenges and opportunities in the development and adoption of new approach methods (NAMs). J Agric Food Chem. 2025;73(13):7519–21. pmid:40106668
- View Article
- PubMed/NCBI
- Google Scholar
11. Mirlohi MS, Yousefi T, Aref AR, Seyfoori A. Integrating new approach methodologies (NAMs) into preclinical regulatory evaluation of oncology drugs. Biomimetics. 2025;10(12):796.
- View Article
- Google Scholar
12. Begam BF, Kumar JS. A study on cheminformatics and its applications on modern drug discovery. Procedia Eng. 2012;38:1264–75.
- View Article
- Google Scholar
13. Chen W, Mao X-Q, Wang X-Z, Liao Y-C, Yin X-Y, Wu H-L, et al. Data-driven discovery of near-infrared type I photosensitizers for RNA-targeted tumor photodynamic therapy. Chem Sci. 2025;16(32):14455–67. pmid:40698170
- View Article
- PubMed/NCBI
- Google Scholar
14. Buglak AA, Telegina TA, Kritsky MS. A quantitative structure-property relationship (QSPR) study of singlet oxygen generation by pteridines. Photochem Photobiol Sci. 2016;15(6):801–11. pmid:27216311
- View Article
- PubMed/NCBI
- Google Scholar
15. He L, Dong J, Yang Y, Huang Z, Ye S, Ke X, et al. Accelerating the discovery of type Ⅱ photosensitizer: Experimentally validated machine learning models for predicting the singlet oxygen quantum yield of photosensitive molecule. J Mol Struct. 2025;1321:139850.
- View Article
- Google Scholar
16. Aldhafferi N. Alternative cancer therapy through modeling pteridines photosensitizer quantum yield singlet oxygen production using swarm-based support vector regression and extreme learning machine. Cogent Eng. 2024;11(1).
- View Article
- Google Scholar
17. Xu S, Li J, Cai P, Liu X, Liu B, Wang X. Self-improving photosensitizer discovery system via Bayesian search with first-principle simulations. J Am Chem Soc. 2021;143(47):19769–77. pmid:34788033
- View Article
- PubMed/NCBI
- Google Scholar
18. Niu G, Du Plessis MC, Sakai T, Ma Y, Sugiyama M. Theoretical comparisons of positive-unlabeled learning against positive-negative learning. Adv Neural Inform Process Syst. 2016;29.
- View Article
- Google Scholar
19. Yang P, Li X-L, Mei J-P, Kwoh C-K, Ng S-K. Positive-unlabeled learning for disease gene identification. Bioinformatics. 2012;28(20):2640–7. pmid:22923290
- View Article
- PubMed/NCBI
- Google Scholar
20. Jaskie K, Spanias A. Positive and unlabeled learning algorithms and applications: A survey. In: 2019 10th International conference on information, intelligence, systems and applications (IISA). IEEE; 2019. pp. 1–8.
21. Datta P, Lee NS, Moolayadukkam S, Sahu RP, Yu X, Guo T, et al. In vitro sonodynamic therapy using a high throughput 3D glioblastoma spheroid model with 5-ALA and TMZ Sonosensitizers. Adv Healthc Mater. 2024;13(32):e2402877. pmid:39434433
- View Article
- PubMed/NCBI
- Google Scholar
22. Fattah AR, Sahu RP, Geng F, Mishriki S, Meleca E, Puri IK, Ghosh S. Inventors; McMaster University, assignee. 3D label-free contactless formation of cellular structures and co-cultures through diamagnetophoresis. United States. 2023.
23. Gaulton A, Bellis LJ, Bento AP, Chambers J, Davies M, Hersey A, et al. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res. 2012;40(D1):D1100–7. pmid:21948594
- View Article
- PubMed/NCBI
- Google Scholar
24. Wang Y, Xiao J, Suzek TO, Zhang J, Wang J, Bryant SH. PubChem: a public information system for analyzing bioactivities of small molecules. Nucleic Acids Res. 2009;37(suppl_2):W623–33. pmid:19498078
- View Article
- PubMed/NCBI
- Google Scholar
25. Wishart DS, Knox C, Guo AC, Cheng D, Shrivastava S, Tzur D, et al. DrugBank: a knowledgebase for drugs, drug actions and drug targets. Nucleic Acids Res. 2008;36(suppl_1):D901–6. pmid:18048412
- View Article
- PubMed/NCBI
- Google Scholar
26. Landrum G. RDKit: A software suite for cheminformatics, computational chemistry, and predictive modeling. J Cheminform. 2013;8(31.10):5281.
- View Article
- Google Scholar
27. Hornik K, Stinchcombe M, White H. Multilayer feedforward networks are universal approximators. Neural Netw. 1989;2(5):359–66.
- View Article
- Google Scholar
28. Cybenko G. Approximation by superpositions of a sigmoidal function. Math Control Signal Syst. 1992;5(4):455–455.
- View Article
- Google Scholar
29. Kensert A, Desmet G, Cabooter D. MolGraph: a Python package for the implementation of molecular graphs and graph neural networks with TensorFlow and Keras. J Comput Aided Mol Des. 2024;39(1):3. pmid:39636382
- View Article
- PubMed/NCBI
- Google Scholar
30. Ruder S. An overview of gradient descent optimization algorithms. arXiv preprint arXiv:1609.04747. 2016.
- View Article
- Google Scholar
31. Rojas R. Neural networks: a systematic introduction. Springer Science & Business Media; 2013.
32. Kiryo R, Niu G, Du Plessis MC, Sugiyama M. Positive-unlabeled learning with non-negative risk estimator. Adv Neural Inform Process Syst. 2017;30.
- View Article
- Google Scholar
33. Elkan C, Noto K. Learning classifiers from only positive and unlabeled data. In: Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. 2008. pp. 213–20.
34. Bekker J, Davis J. Learning from positive and unlabeled data: a survey. Mach Learn. 2020;109(4):719–60.
- View Article
- Google Scholar
35. Abadi M, Agarwal A, Barham P, Brevdo E, Chen Z, Citro C, et al. Tensorflow: large-scale machine learning on heterogeneous distributed systems. arXiv preprint. 2016.
- View Article
- Google Scholar
36. Lee WS, Liu B. Learning with positive and unlabeled examples using weighted logistic regression. ICML. 2003;3(2003)448–55.
- View Article
- Google Scholar
37. Bepler T, Morin A, Rapp M, Brasch J, Shapiro L, Noble AJ, et al. Positive-unlabeled convolutional neural networks for particle picking in cryo-electron micrographs. Nat Methods. 2019;16(11):1153–60. pmid:31591578
- View Article
- PubMed/NCBI
- Google Scholar
38. Swanson K, Walther P, Leitz J, Mukherjee S, Wu JC, Shivnaraine RV, et al. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics. 2024;40(7):btae416. pmid:38913862
- View Article
- PubMed/NCBI
- Google Scholar
39. Fu L, Shi S, Yi J, Wang N, He Y, Wu Z, Peng J, Deng Y, Wang W, Wu C, Lyu A. ADMETlab 3.0: an updated comprehensive online ADMET prediction platform enhanced with broader coverage, improved performance, API functionality and decision support. Nucleic Acids Res. 2024;52(W1):W422–31.
- View Article
- Google Scholar
40. Mehta DC, Short JL, Nicolazzo JA. Memantine transport across the mouse blood-brain barrier is mediated by a cationic influx H+ antiporter. Mol Pharm. 2013;10(12):4491–8. pmid:24098999
- View Article
- PubMed/NCBI
- Google Scholar
41. Carpentier A, Canney M, Vignot A, Reina V, Beccaria K, Horodyckid C, et al. Clinical trial of blood-brain barrier disruption by pulsed ultrasound. Sci Transl Med. 2016;8(343):343re2. pmid:27306666
- View Article
- PubMed/NCBI
- Google Scholar
42. Idbaih A, Canney M, Belin L, Desseaux C, Vignot A, Bouchoux G, et al. Safety and feasibility of repeated and transient blood–brain barrier disruption by pulsed ultrasound in patients with recurrent glioblastoma. Clin Cancer Res. 2019;25(13):3793–801.
- View Article
- Google Scholar
43. Sonabend AM, Gould A, Amidei C, Ward R, Schmidt KA, Zhang DY, et al. Repeated blood–brain barrier opening with an implantable ultrasound device for delivery of albumin-bound paclitaxel in patients with recurrent glioblastoma: a phase 1 trial. The Lancet Oncology. 2023;24(5):509–22.
- View Article
- Google Scholar
44. Mikuła-Pietrasik J, Witucka A, Pakuła M, Uruski P, Begier-Krasińska B, Niklas A, et al. Comprehensive review on how platinum- and taxane-based chemotherapy of ovarian cancer affects biology of normal cells. Cell Mol Life Sci. 2019;76(4):681–97. pmid:30382284
- View Article
- PubMed/NCBI
- Google Scholar
45. Paškevičiūtė M, Januškevičienė I, Sakalauskienė K, Raišutis R, Petrikaitė V. Evaluation of low-intensity pulsed ultrasound on doxorubicin delivery in 2D and 3D cancer cell cultures. Sci Rep. 2020;10(1):16161. pmid:32999381
- View Article
- PubMed/NCBI
- Google Scholar
46. Müller-Längle A, Lutz H, Hehlgans S, Rödel F, Rau K, Laube B. NMDA receptor-mediated signaling pathways enhance radiation resistance, survival and migration in glioblastoma cells-a potential target for adjuvant radiotherapy. Cancers (Basel). 2019;11(4):503. pmid:30970642
- View Article
- PubMed/NCBI
- Google Scholar
47. Takano T, Lin JH, Arcuino G, Gao Q, Yang J, Nedergaard M. Glutamate release promotes growth of malignant gliomas. Nat Med. 2001;7(9):1010–5. pmid:11533703
- View Article
- PubMed/NCBI
- Google Scholar
48. Lyons SA, Chung WJ, Weaver AK, Ogunrinu T, Sontheimer H. Autocrine glutamate signaling promotes glioma cell invasion. Cancer Res. 2007;67(19):9463–71. pmid:17909056
- View Article
- PubMed/NCBI
- Google Scholar
49. Ortiz R, Perazzoli G, Cabeza L, Jiménez-Luna C, Luque R, Prados J, et al. Temozolomide: an updated overview of resistance mechanisms, nanotechnology advances and clinical applications. Curr Neuropharmacol. 2021;19(4):513–37. pmid:32589560
- View Article
- PubMed/NCBI
- Google Scholar
50. Günther W, Pawlak E, Damasceno R, Arnold H, Terzis AJ. Temozolomide induces apoptosis and senescence in glioma cells cultured as multicellular spheroids. Br J Cancer. 2003;88(3):463–9. pmid:12569392
- View Article
- PubMed/NCBI
- Google Scholar

[ref1] 1. Zhang H, Li Y. Temozolomide-derived therapeutic strategies to overcome resistance in glioblastoma. J Med Chem. 2025;68(18):18810–28. pmid:40913573
View Article
PubMed/NCBI
Google Scholar

[2] View Article

[3] PubMed/NCBI

[4] Google Scholar

[ref2] 2. Fernandes C, Costa A, Osório L, Lago RC, Linhares P, Carvalho B, et al. Current standards of care in glioblastoma therapy. Exon Publications; 2017.

[ref3] 3. Löscher W, Potschka H. Drug resistance in brain diseases and the role of drug efflux transporters. Nat Rev Neurosci. 2005;6(8):591–602. pmid:16025095
View Article
PubMed/NCBI
Google Scholar

[7] View Article

[8] PubMed/NCBI

[9] Google Scholar

[ref4] 4. Yuan H, Chong H, Wang B, Zhu C, Liu L, Yang Q, et al. Chemical molecule-induced light-activated system for anticancer and antifungal activities. J Am Chem Soc. 2012;134(32):13184–7. pmid:22830933
View Article
PubMed/NCBI
Google Scholar

[11] View Article

[12] PubMed/NCBI

[13] Google Scholar

[ref5] 5. Li X, Kwon N, Guo T, Liu Z, Yoon J. Innovative strategies for hypoxic-tumor photodynamic therapy. Angew Chem Int Ed Engl. 2018;57(36):11522–31. pmid:29808948
View Article
PubMed/NCBI
Google Scholar

[15] View Article

[16] PubMed/NCBI

[17] Google Scholar

[ref6] 6. McHale AP, Callan JF, Nomikou N, Fowley C, Callan B. Sonodynamic therapy: concept, mechanism and application to cancer treatment. Therapeutic Ultrasound. 2016:429–50.
View Article
Google Scholar

[19] View Article

[20] Google Scholar

[ref7] 7. Datta P, Moolayadukkam S, Chowdhury D, Rayes A, Lee NS, Sahu RP, et al. Recent advances and future directions in sonodynamic therapy for cancer treatment. BME Front. 2024;2024:0080. pmid:39735354
View Article
PubMed/NCBI
Google Scholar

[22] View Article

[23] PubMed/NCBI

[24] Google Scholar

[ref8] 8. Chatzinikolaidou M. Cell spheroids: the new frontiers in in vitro models for cancer drug validation. Drug Discov Today. 2016;21(9):1553–60. pmid:27371506
View Article
PubMed/NCBI
Google Scholar

[26] View Article

[27] PubMed/NCBI

[28] Google Scholar

[ref9] 9. Mehta K, Maass C, Cucurull-Sanchez L, Pichardo-Almarza C, Subramanian K, Androulakis IP, et al. Modernizing preclinical drug development: the role of new approach methodologies. ACS Pharmacol Transl Sci. 2025;8(6):1513–25. pmid:40567279
View Article
PubMed/NCBI
Google Scholar

[30] View Article

[31] PubMed/NCBI

[32] Google Scholar

[ref10] 10. Levine SL, Riter LS, Lagadic L, Bejarano AC, Burden N, Burgoon LD, et al. Challenges and opportunities in the development and adoption of new approach methods (NAMs). J Agric Food Chem. 2025;73(13):7519–21. pmid:40106668
View Article
PubMed/NCBI
Google Scholar

[34] View Article

[35] PubMed/NCBI

[36] Google Scholar

[ref11] 11. Mirlohi MS, Yousefi T, Aref AR, Seyfoori A. Integrating new approach methodologies (NAMs) into preclinical regulatory evaluation of oncology drugs. Biomimetics. 2025;10(12):796.
View Article
Google Scholar

[38] View Article

[39] Google Scholar

[ref12] 12. Begam BF, Kumar JS. A study on cheminformatics and its applications on modern drug discovery. Procedia Eng. 2012;38:1264–75.
View Article
Google Scholar

[41] View Article

[42] Google Scholar

[ref13] 13. Chen W, Mao X-Q, Wang X-Z, Liao Y-C, Yin X-Y, Wu H-L, et al. Data-driven discovery of near-infrared type I photosensitizers for RNA-targeted tumor photodynamic therapy. Chem Sci. 2025;16(32):14455–67. pmid:40698170
View Article
PubMed/NCBI
Google Scholar

[44] View Article

[45] PubMed/NCBI

[46] Google Scholar

[ref14] 14. Buglak AA, Telegina TA, Kritsky MS. A quantitative structure-property relationship (QSPR) study of singlet oxygen generation by pteridines. Photochem Photobiol Sci. 2016;15(6):801–11. pmid:27216311
View Article
PubMed/NCBI
Google Scholar

[48] View Article

[49] PubMed/NCBI

[50] Google Scholar

[ref15] 15. He L, Dong J, Yang Y, Huang Z, Ye S, Ke X, et al. Accelerating the discovery of type Ⅱ photosensitizer: Experimentally validated machine learning models for predicting the singlet oxygen quantum yield of photosensitive molecule. J Mol Struct. 2025;1321:139850.
View Article
Google Scholar

[52] View Article

[53] Google Scholar

[ref16] 16. Aldhafferi N. Alternative cancer therapy through modeling pteridines photosensitizer quantum yield singlet oxygen production using swarm-based support vector regression and extreme learning machine. Cogent Eng. 2024;11(1).
View Article
Google Scholar

[55] View Article

[56] Google Scholar

[ref17] 17. Xu S, Li J, Cai P, Liu X, Liu B, Wang X. Self-improving photosensitizer discovery system via Bayesian search with first-principle simulations. J Am Chem Soc. 2021;143(47):19769–77. pmid:34788033
View Article
PubMed/NCBI
Google Scholar

[58] View Article

[59] PubMed/NCBI

[60] Google Scholar

[ref18] 18. Niu G, Du Plessis MC, Sakai T, Ma Y, Sugiyama M. Theoretical comparisons of positive-unlabeled learning against positive-negative learning. Adv Neural Inform Process Syst. 2016;29.
View Article
Google Scholar

[62] View Article

[63] Google Scholar

[ref19] 19. Yang P, Li X-L, Mei J-P, Kwoh C-K, Ng S-K. Positive-unlabeled learning for disease gene identification. Bioinformatics. 2012;28(20):2640–7. pmid:22923290
View Article
PubMed/NCBI
Google Scholar

[65] View Article

[66] PubMed/NCBI

[67] Google Scholar

[ref20] 20. Jaskie K, Spanias A. Positive and unlabeled learning algorithms and applications: A survey. In: 2019 10th International conference on information, intelligence, systems and applications (IISA). IEEE; 2019. pp. 1–8.

[ref21] 21. Datta P, Lee NS, Moolayadukkam S, Sahu RP, Yu X, Guo T, et al. In vitro sonodynamic therapy using a high throughput 3D glioblastoma spheroid model with 5-ALA and TMZ Sonosensitizers. Adv Healthc Mater. 2024;13(32):e2402877. pmid:39434433
View Article
PubMed/NCBI
Google Scholar

[70] View Article

[71] PubMed/NCBI

[72] Google Scholar

[ref22] 22. Fattah AR, Sahu RP, Geng F, Mishriki S, Meleca E, Puri IK, Ghosh S. Inventors; McMaster University, assignee. 3D label-free contactless formation of cellular structures and co-cultures through diamagnetophoresis. United States. 2023.

[ref23] 23. Gaulton A, Bellis LJ, Bento AP, Chambers J, Davies M, Hersey A, et al. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res. 2012;40(D1):D1100–7. pmid:21948594
View Article
PubMed/NCBI
Google Scholar

[75] View Article

[76] PubMed/NCBI

[77] Google Scholar

[ref24] 24. Wang Y, Xiao J, Suzek TO, Zhang J, Wang J, Bryant SH. PubChem: a public information system for analyzing bioactivities of small molecules. Nucleic Acids Res. 2009;37(suppl_2):W623–33. pmid:19498078
View Article
PubMed/NCBI
Google Scholar

[79] View Article

[80] PubMed/NCBI

[81] Google Scholar

[ref25] 25. Wishart DS, Knox C, Guo AC, Cheng D, Shrivastava S, Tzur D, et al. DrugBank: a knowledgebase for drugs, drug actions and drug targets. Nucleic Acids Res. 2008;36(suppl_1):D901–6. pmid:18048412
View Article
PubMed/NCBI
Google Scholar

[83] View Article

[84] PubMed/NCBI

[85] Google Scholar

[ref26] 26. Landrum G. RDKit: A software suite for cheminformatics, computational chemistry, and predictive modeling. J Cheminform. 2013;8(31.10):5281.
View Article
Google Scholar

[87] View Article

[88] Google Scholar

[ref27] 27. Hornik K, Stinchcombe M, White H. Multilayer feedforward networks are universal approximators. Neural Netw. 1989;2(5):359–66.
View Article
Google Scholar

[90] View Article

[91] Google Scholar

[ref28] 28. Cybenko G. Approximation by superpositions of a sigmoidal function. Math Control Signal Syst. 1992;5(4):455–455.
View Article
Google Scholar

[93] View Article

[94] Google Scholar

[ref29] 29. Kensert A, Desmet G, Cabooter D. MolGraph: a Python package for the implementation of molecular graphs and graph neural networks with TensorFlow and Keras. J Comput Aided Mol Des. 2024;39(1):3. pmid:39636382
View Article
PubMed/NCBI
Google Scholar

[96] View Article

[97] PubMed/NCBI

[98] Google Scholar

[ref30] 30. Ruder S. An overview of gradient descent optimization algorithms. arXiv preprint arXiv:1609.04747. 2016.
View Article
Google Scholar

[100] View Article

[101] Google Scholar

[ref31] 31. Rojas R. Neural networks: a systematic introduction. Springer Science & Business Media; 2013.

[ref32] 32. Kiryo R, Niu G, Du Plessis MC, Sugiyama M. Positive-unlabeled learning with non-negative risk estimator. Adv Neural Inform Process Syst. 2017;30.
View Article
Google Scholar

[104] View Article

[105] Google Scholar

[ref33] 33. Elkan C, Noto K. Learning classifiers from only positive and unlabeled data. In: Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. 2008. pp. 213–20.

[ref34] 34. Bekker J, Davis J. Learning from positive and unlabeled data: a survey. Mach Learn. 2020;109(4):719–60.
View Article
Google Scholar

[108] View Article

[109] Google Scholar

[ref35] 35. Abadi M, Agarwal A, Barham P, Brevdo E, Chen Z, Citro C, et al. Tensorflow: large-scale machine learning on heterogeneous distributed systems. arXiv preprint. 2016.
View Article
Google Scholar

[111] View Article

[112] Google Scholar

[ref36] 36. Lee WS, Liu B. Learning with positive and unlabeled examples using weighted logistic regression. ICML. 2003;3(2003)448–55.
View Article
Google Scholar

[114] View Article

[115] Google Scholar

[ref37] 37. Bepler T, Morin A, Rapp M, Brasch J, Shapiro L, Noble AJ, et al. Positive-unlabeled convolutional neural networks for particle picking in cryo-electron micrographs. Nat Methods. 2019;16(11):1153–60. pmid:31591578
View Article
PubMed/NCBI
Google Scholar

[117] View Article

[118] PubMed/NCBI

[119] Google Scholar

[ref38] 38. Swanson K, Walther P, Leitz J, Mukherjee S, Wu JC, Shivnaraine RV, et al. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics. 2024;40(7):btae416. pmid:38913862
View Article
PubMed/NCBI
Google Scholar

[121] View Article

[122] PubMed/NCBI

[123] Google Scholar

[ref39] 39. Fu L, Shi S, Yi J, Wang N, He Y, Wu Z, Peng J, Deng Y, Wang W, Wu C, Lyu A. ADMETlab 3.0: an updated comprehensive online ADMET prediction platform enhanced with broader coverage, improved performance, API functionality and decision support. Nucleic Acids Res. 2024;52(W1):W422–31.
View Article
Google Scholar

[125] View Article

[126] Google Scholar

[ref40] 40. Mehta DC, Short JL, Nicolazzo JA. Memantine transport across the mouse blood-brain barrier is mediated by a cationic influx H+ antiporter. Mol Pharm. 2013;10(12):4491–8. pmid:24098999
View Article
PubMed/NCBI
Google Scholar

[128] View Article

[129] PubMed/NCBI

[130] Google Scholar

[ref41] 41. Carpentier A, Canney M, Vignot A, Reina V, Beccaria K, Horodyckid C, et al. Clinical trial of blood-brain barrier disruption by pulsed ultrasound. Sci Transl Med. 2016;8(343):343re2. pmid:27306666
View Article
PubMed/NCBI
Google Scholar

[132] View Article

[133] PubMed/NCBI

[134] Google Scholar

[ref42] 42. Idbaih A, Canney M, Belin L, Desseaux C, Vignot A, Bouchoux G, et al. Safety and feasibility of repeated and transient blood–brain barrier disruption by pulsed ultrasound in patients with recurrent glioblastoma. Clin Cancer Res. 2019;25(13):3793–801.
View Article
Google Scholar

[136] View Article

[137] Google Scholar

[ref43] 43. Sonabend AM, Gould A, Amidei C, Ward R, Schmidt KA, Zhang DY, et al. Repeated blood–brain barrier opening with an implantable ultrasound device for delivery of albumin-bound paclitaxel in patients with recurrent glioblastoma: a phase 1 trial. The Lancet Oncology. 2023;24(5):509–22.
View Article
Google Scholar

[139] View Article

[140] Google Scholar

[ref44] 44. Mikuła-Pietrasik J, Witucka A, Pakuła M, Uruski P, Begier-Krasińska B, Niklas A, et al. Comprehensive review on how platinum- and taxane-based chemotherapy of ovarian cancer affects biology of normal cells. Cell Mol Life Sci. 2019;76(4):681–97. pmid:30382284
View Article
PubMed/NCBI
Google Scholar

[142] View Article

[143] PubMed/NCBI

[144] Google Scholar

[ref45] 45. Paškevičiūtė M, Januškevičienė I, Sakalauskienė K, Raišutis R, Petrikaitė V. Evaluation of low-intensity pulsed ultrasound on doxorubicin delivery in 2D and 3D cancer cell cultures. Sci Rep. 2020;10(1):16161. pmid:32999381
View Article
PubMed/NCBI
Google Scholar

[146] View Article

[147] PubMed/NCBI

[148] Google Scholar

[ref46] 46. Müller-Längle A, Lutz H, Hehlgans S, Rödel F, Rau K, Laube B. NMDA receptor-mediated signaling pathways enhance radiation resistance, survival and migration in glioblastoma cells-a potential target for adjuvant radiotherapy. Cancers (Basel). 2019;11(4):503. pmid:30970642
View Article
PubMed/NCBI
Google Scholar

[150] View Article

[151] PubMed/NCBI

[152] Google Scholar

[ref47] 47. Takano T, Lin JH, Arcuino G, Gao Q, Yang J, Nedergaard M. Glutamate release promotes growth of malignant gliomas. Nat Med. 2001;7(9):1010–5. pmid:11533703
View Article
PubMed/NCBI
Google Scholar

[154] View Article

[155] PubMed/NCBI

[156] Google Scholar

[ref48] 48. Lyons SA, Chung WJ, Weaver AK, Ogunrinu T, Sontheimer H. Autocrine glutamate signaling promotes glioma cell invasion. Cancer Res. 2007;67(19):9463–71. pmid:17909056
View Article
PubMed/NCBI
Google Scholar

[158] View Article

[159] PubMed/NCBI

[160] Google Scholar

[ref49] 49. Ortiz R, Perazzoli G, Cabeza L, Jiménez-Luna C, Luque R, Prados J, et al. Temozolomide: an updated overview of resistance mechanisms, nanotechnology advances and clinical applications. Curr Neuropharmacol. 2021;19(4):513–37. pmid:32589560
View Article
PubMed/NCBI
Google Scholar

[162] View Article

[163] PubMed/NCBI

[164] Google Scholar

[ref50] 50. Günther W, Pawlak E, Damasceno R, Arnold H, Terzis AJ. Temozolomide induces apoptosis and senescence in glioma cells cultured as multicellular spheroids. Br J Cancer. 2003;88(3):463–9. pmid:12569392
View Article
PubMed/NCBI
Google Scholar

[166] View Article

[167] PubMed/NCBI

[168] Google Scholar

Figures

Abstract

1. Introduction

2. Materials and methods

2.1. Reagents and chemicals

2.2. Cell line and cell culture

2.3. Glioblastoma tumor spheroid printing

2.4. Ultrasound stimulation and experimental setup

2.5. Confocal imaging

2.6. Spheroid cell viability and IC50 determination for carboplatin, memantine hydrochloride, and TMZ

2.7. Oxidative stress determination

2.8. Mitochondrial membrane potential

2.9. Cellular apoptosis assay

2.10. Statistical analysis

2.11. Ethics statement

3. Results

3.1. Database curation for AI model

3.2. Model development

3.3. Characterization of the glioblastoma spheroid 24 hours after printing

3.4. IC50 for carboplatin, memantine hydrochloride, and TMZ

3.5. Sonodynamic therapy (SDT) on glioblastoma spheroid treated with carboplatin, memantine hydrochloride, and TMZ

4. Discussion

5. Conclusions

Supporting information

S1 Table. ADMET properties for carboplatin, memantine hydrochloride, and TMZ.

S2 Table. Small molecule database containing RDKit-computed descriptors.

S1 Fig. Oxidative stress determination 1 hour after SDT via staining the spheroids with cellROX green reagent.

S2 Fig. Mitochondrial membrane potential (MMP) determination using JC1 dye 24 hours after SDT.

S3 Fig. Cellular apoptosis determination via caspase 3/7 green ready probe staining 24 hours after SDT.

S1 Text. Python programs for the deep learning workflow.

References

2.6. Spheroid cell viability and IC₅₀ determination for carboplatin, memantine hydrochloride, and TMZ

3.4. IC₅₀ for carboplatin, memantine hydrochloride, and TMZ