PRESENTERS

Department of Computer Science & Engineering

Anomaly Detection in FDM 3D Printing

The increasing adoption of additive manufacturing (AM) has raised the need for reliable, non-destructive defect detection. Active thermography offers an inspection method, but conventional machine learning requires large datasets containing both defective and non-defective samples, increasing costs. This study presents a simulation-assisted anomaly detection framework using an ensemble of convolutional autoencoders (CAEs) trained on only defect-free thermal sequences. A finite-difference thermal model of the FDM printing process, calibrated on digitized reference data, generates nozzle-heating and cooling profiles for a single filament pass, with sensor noise and blur applied to emulate experimentally captured infrared thermography data from heating tests. The CAEs learn from cooling behavior and engineered features, and then identify defects via reconstruction errors. To improve anomaly discrimination beyond reconstruction error, several complementary detectors are applied in parallel: a Mahalanobis distance test in a reduced latent space, an Isolation Forest (IF) on reconstruction error, and One-Class Support Vector Machine (OCSVM) variants (radial basis function (RBF), polynomial (Poly), and a reconstruction-error-based nu-SVM variant (nu)). Four defect classes, nozzle clogging, cooling drafts, interlayer delamination, and underextrusion, were introduced as localized perturbations to thermal conductivity, convective loss, or heat-capacity terms across a fixed region of the filament, at severities ranging from 2-50% depending on class. Detection is evaluated across defect severity using ROC/AUC analysis. Results demonstrate that the OCSVM-RBF method achieved high and severity-consistent detection sensitivity (~0.85), while other methods such as IF underperform at high severities for gradual defects such as draft. 

The findings show that unsupervised learning can lower inspection costs while maintaining high detection accuracy, supporting scalable quality assurance in additive manufacturing. In addition, the relation between the performances of specific models over different defect types and severities suggests a connection between a defect’s physics and the model’s architecture. Ongoing improvements include resolving sensitivity in the OCSVM-RBF model, extending the IF scoring to extended IF (EIF) evaluation, improving the AUC/ROC metrics of all methods, justifying performance of models against their architecture, and developing a user interface (UI) tool to accompany the simulation.

RESEARCHERS: Shrish Jain, American High School '27

ADVISOR: Laurienzo Lab, Applied Mathematics | Computer Modeling | Modern Physics
KEYWORDS: Thermal Simulation | Unsupervised Machine Learning | FDM Defect Detection 

Department of Chemistry, Biochemistry & Physics

Translating nature’s molecular language, from isolation to preclinical leads: Proscillaridin A, Andrographolide, and their analogues as anticancer small molecule therapeutics.

From the earliest examples of human history, phytochemical sources of therapeutic materials have provided the inspiration for the discovery and subsequent design of small molecule agents. In fact, while four percent of modern clinically-approved therapeutics trace their origin directly from natural sources, these compounds are outnumbered nearly tenfold by synthetic derivatives thereof, highlighting both the foundational role of natural products and the enabling power of design-informed synthesis in the search for new therapeutics. Among small molecules used to treat human cancers, these numbers are even higher, with a tenth of small molecules originating from nature, and over two-thirds of all anticancer drugs in the modern day deriving their design from a natural product. Among the many natural product classes interrogated by our laboratory as preclinical leads for cancer, cardiac glycosides such as Proscillaridin A and its analogs have demonstrated unique potential. While their endogenous function as cardiotonic agents relies on high binding specificity and affinity to sodium / potassium ATPases (NKAs), recent evidence in the literature from our laboratory and from others have demonstrated that the therapeutic activity of such compounds may extend to a diverse array of intracellular targets connected to cancer cell proliferation, both in connection to and orthogonal to native NKA inhibitory activity. Last year, we reported the synthesis of five Proscillaridin A glycan analogs incorporating ketal, acetate ester, and silyl ether motifs, and found that several analogs, especially the glycan acetonide, exhibited potent surface expression of Caspase 3, a key apoptotic marker, and concomitant downregulation of CD117 and CD133, key markers for cancer cell stemness. (Somani, et al. Disc. Pharm. Sci. 2025) 

With these results in hand, we and collaborators in biotech advanced the acetonide analog into preclinical studies, in which Proscillaridin A and its analogs demonstrated potent antitumor activity but with dose-limiting off-target toxicity challenges, with recent evidence pointing to a potential role of immunogenic cell death as a plausible mechanism of action of these and like molecules. In yet another example of nature-inspired synthesis, our laboratory has previously investigated silyl and trityl ether analogs of the labdane diterpenoid andrographolide as anticancer agents in colorectal and breast cancer models. (Gu, et al. Bioorg. Med. Chem. Lett. 2025) Our laboratory and others have reported that functionalization at the C19 hydroxyl position both ablates metabolic degradation and re-orients the molecule to target the Wnt1/β-catenin signaling axis, which is overactivated in many cancer models. (Gutti, et al. Essential Chem. 2026) As part of a broader effort and strategy in our laboratory, these silyl ether analogs present a unique strategy to incorporate bio-orthogonal reactivity and structural diversity into natural products and their analogs. (Raval, et al. Molecules 2026) More recently, we have expanded this SAR to include murine cancer cell models as a proxy for collective pathway-associated potency generic to multiple cancer types. We found potent, dose-responsive anticancer activity of C19 andrographolide analogs in CT26 murine colorectal cancer, 4T1 murine ductal adenocarcinoma, B16F10 melanoma, and Neuro2A neuroblastoma cell lines, concomitant with induction of apoptosis at single- to double-digit micromolar concentrations determined by flow cytometry. Further, while such compounds possessed similar, potent activity in in vitro cancer cell models, we encountered unique dose-limiting solubility challenges in preclinical studies.  With this observed improved anticancer activity of A-ring functionalized andrographolide analogs, we were further inspired to pursue the chemical synthesis of unique cyclic A-ring analogs of andrographolide, including an oxetane congener, which is co-isolated from the same plant. En route to accessing this unique and highly strained four-membered ring, we synthetically prepared several cyclic A-ring analogs, including aryl acetals and two diastereomers of a cyclic sulfite, whose structures are unequivocally established by a combination of X-ray crystallographic and elemental analysis spectra. Further, we identify a unique role that the C-ring siloxybutenolide plays in perturbation of A-ring conformational bias, through a collection of experimental observations, NMR spectroscopic data, and computer modeling, but that this conformational bias is generally insufficient to overcome foundational ring strain and conformational distortion required to effect a 4-exo-tet oxetane ring closure via a C-19 leaving group. 

Further, the ultimate biological implications of this distal conformational control proposed by such experiments are underreported, and preliminary biological data suggests that A-ring andrographolide analogs possess unique potency against Calu1 and A549 human lung cancer cells. Altogether then, this colloquium serves to highlight two overlapping perspectives on the translational role of synthetic chemistry, structural design, and preclinical efficacy of natural products and their analogs as preclinical leads for the treatment of cancers.

RESEARCHERS: Ruirui Liu, Mission San Jose High School '27

ADVISOR: Njoo Lab, Synthesis | Physical Organic Chemistry | Catalysis | Chemical Biology | Spectroscopy | Medicinal Chemistry
KEYWORDS: Natural Products | Preclinical Evaluation | Chemical Biology | Synthetic Derivatives | HERV-K | Silyl Ethers | Medicinal Chemistry 

Department of Computer Science & Engineering

Comparative study on three machine learning models in novel autonomous drone-based analysis of bike lane infrastructure

Ensuring the safety of bikers on bike lanes is essential to many cities, but monitoring the bike lanes manually is hard and time consuming. This research proposes a way to use drones to effectively inspect bike lanes using sensors  and cameras. The drones will use sensors to autonomously capture images to identify cracks, debris, and other obstructions. By using drones, this will result in less casualties on the road and a more efficient and safe experience for the riders. Drones are more cost effective, faster, and overall better than manually inspecting the bike lanes. The expected outcome of this project is that bike lanes will be much safer to ride on and accidents will greatly reduce and it will improve the cyclists safety.

RESEARCHERS: Saanvi Buragapu, Emerald High School '28; Samyuktha Bodduluri, McNeil High School '27 

ADVISOR: McMahan Lab, Quantum Computing & Computer Science
KEYWORDS: Machine Learning | Drone-Based Inspection | Bike Lane Safety | Autonomous Monitoring

Department of Chemistry, Biochemistry & Physics

Scalable formal synthesis of (R)-(+)-etomoxir and its analogs enabled by copper catalyzed aerobic oxidation

Etomoxir is a covalent inhibitor of CPT1, a transmembrane mitochondrial protein that acts as the rate-limiting enzyme for fatty acid oxidation. This enzyme plays a major role in metabolic diseases such as diabetes, where regulation of fatty acid biosynthesis and β-oxidation kinetics through CPT1 are effective treatments for such diseases. The 4-Cl phenolic ether on (R)-(+)-etomoxir is a key SAR hotspot for enabling isoform selective inhibition of CPT1. Previously reported syntheses either require early installation of a 4-Cl phenolic ether which precludes the potential for late stage aryl substitution, or employ large scale pyrophoric reactions in early synthetic operations which are challenging to scale. We demonstrate the scalability of a new synthetic route to intercept a late-stage allylic alcohol in route to (R)-(+)-etomoxir. Notably, our alternate retrosynthetic disconnection, which proceeds through a catalytic aerobic oxidation and a one-flask tandem aldol condensation- reduction sequence, to install a key allylic methylene, avoids pyrophoric materials such as n-butyllithium. With a scalable synthesis of a key diversifiable intermediate in hand, our laboratory is currently preparing a library of diverse (R)-(+)-etomoxir analogs to more fully interrogate the SAR of the aryl ring in CPT1 inhibitory activity.

RESEARCHERS: Sophie Chen, The Harker School ‘27; Alisha Hothi, Cupertino High School ‘28

ADVISOR: Njoo Lab, Synthesis | Physical Organic Chemistry | Catalysis | Chemical Biology | Spectroscopy | Medicinal Chemistry
KEYWORDS: Carnitine palmitoyl transferase 1 (CPT-1) | (+)-etomoxir | Catalytic Oxidation

Department of Computer Science & Engineering

Drone-Based Object Detection to Identify Vegetation Hazards Near Power Lines for Wildfire Mitigation

Vegetation encroachment near powerlines poses a significant wildfire risk in various locations around the world, requiring regular inspection to maintain safe clearance distances. Existing inspection methods are often costly and difficult to scale across large geographic areas. This study presents and validates a scalable drone-based system for detecting vegetation-powerline hazards using deep-learning-based object detection. A multi-task detection and hazard inference framework was trained and evaluated using 3,768 manually annotated aerial drone images containing vegetation and powerline infrastructure. The system integrates UAV-based image acquisition with a convolutional neural network–based detection pipeline capable of automatically identifying vegetation–powerline proximity hazards. Model performance was evaluated using k-fold cross-validation and holdout testing. The system achieved strong classification performance scores of 0.91 accuracy, 0.92 precision, 0.97 recall, 0.95 F1 score, and 0.93 ROC-AUC. These results demonstrate that drone-based deep learning systems can enable scalable and cost-effective vegetation hazard monitoring for powerline infrastructure.

RESEARCHERS: Nirupama Balaji, American High School '28
ADVISOR: McMahan Lab, Quantum Computing & Computer Science
KEYWORDS: Powerline Infrastructure | Drone-Based Systems | Deep Learning | Object Detection | Unmanned Aerial Vehicles | Convolutional Neural Networks | Hazard Classification

Department of Chemistry, Biochemistry & Physics

Mechanistic Investigation Into Phenolic Ester-Mediated Site-Selective Amine Acylation

The formation of amide bonds is frequently employed in the synthesis of biologically active compounds, but due to the similar reactivity of different amines, obtaining high yields in reactions with polyamine-containing substrates is challenging. While the use of protecting groups has frequently been employed to address this challenge, developing reagents with exquisite selectivity for the choice of one amine nucleophile over other nucleophiles has the potential to greatly reduce the step count and provide expedient access to mono-acylated products from polyamine substrates. Here, we prepare a library of 10 para and meta-substituted phenolic esters with a variety of electron-withdrawing and electron-donating groups and demonstrate their unique selectivities for amine nucleophiles that were previously not possible to achieve through conventional amide bond-forming reactions. The mechanistic basis for this selectivity is further justified through the use of real-time reaction kinetics by 19F Benchtop NMR. In addition, we demonstrate the applicability of these esters on a broad electrophile substrate scope.

RESEARCHERS: Ellie Leo, Aragon High School '28
ADVISOR: Njoo Lab, Synthesis | Physical Organic Chemistry | Catalysis | Chemical Biology | Spectroscopy | Medicinal Chemistry
KEYWORDS: Organic Synthesis | Chemoselectivity | 19F Benchtop NMR | Phenolic Esters

Department of Chemistry, Biochemistry & Physics

Reactivity-informed pharmacophore editing and biological evaluation of andrographolide and synthesis of a-ring analogs: Closing the loop 

Andrographolide, a labdane diterpenoid isolated from the plant Andrographis paniculata, has been studied as an anticancer small molecule that putatively functions through modulation of the NF-κB signaling pathway. Functionalization of the C-19 hydroxyl of andrographolide has been shown to alter the primary mode of action to modulation of the Wnt/B-catenin signaling pathway from NF-KB inhibition (Reabroi et al. Biomedicine & Pharmacotherapy 2018). We identified that andrographolide A-ring trityl and silyl ethers exhibit greater potency than the natural product in in vitro human colon and breast cancer models (Gu, et al. Bioorganic and Medicinal Chemistry Letters 2025). With this established insight into A-ring modifications on andrographolide’s biological activity in both Wnt1 and NF-κB signaling, as well as the unique overlap of Human Endogenous Retrovirus-K in the two aforementioned pathways, we sought to study the effects of our analogs on the rapid cell proliferation and metastasis associated with HERV-K in breast cancer. Through stably transfected murine cancer cell lines, which exhibit fundamentally different mechanisms of canonical Wnt1 signaling, we observed that our analogs exhibited similar trends in anticancer potency in murine colon carcinoma, breast cancer, and melanoma cell lines as in their human counterparts. In a subsequent investigation to access a natural co-isolated A-ring oxetane (Jantan et al., Phytochemistry 1994), we developed a five-step synthetic sequence to access a related cyclic sulfite of both diastereomers, of whose structures were unequivocally established by 2D NMR spectroscopy and X-ray diffraction, which were evaluated alongside a broader library of cyclic A-ring analogs. These studies demonstrate that the andrographolide A-ring is a tunable site for modulating biological potency and mechanism of action in the context of cancer drug discovery.

RESEARCHERS: Aileen Pak, Castro Valley High ‘27
ADVISOR: Njoo Lab, Synthesis | Physical Organic Chemistry | Catalysis | Chemical Biology | Spectroscopy | Medicinal Chemistry
KEYWORDS: Natural Products | Organic Chemistry | Cancer Therapeutic | Pre-clinical Studies | Medicinal Chemistry 

Department of Computer Science & Engineering

Benchmarking Deep Learning Architectures for Early-Stage Brain Tumor Detection & Classification 

Early-stage brain tumors affect millions worldwide due to the lack of quality diagnosis to identify small and subclinical lumps of abnormal brain activity. Currently, to check if a patient has a brain tumor, doctors perform a neurological exam to check brain-body interaction, and in more severe cases where symptoms like seizures occur, they perform an MRI or CT scan on the patient. However, at this stage, the tumor has likely grown to the point of affecting essential parts of the brain, making treatment harder. This model uses a non-invasive deep learning image detection model to detect brain tumors accurately and at an early stage so that future patients could act before symptoms worsen. Over 100,000 cleaned and diverse tumorous and nontumorous, data-augmented MRI scans were used.

The Convolutional Neural Network (CNNs) algorithm employs feature extraction by highlighting edges and boundaries of local abnormal brain activity. Recurrent Neural Networks’ (RNNs) hidden-state processing and Long-Short Term Memory’s (LSTM) regulation gates led to Vision Transformers (ViT), specifically the Swin ViT. Its self-attention mechanism allows models to prioritize essential parts of the MRI. While CNNs extract hierarchical features layer by layer, ViTs analyze all regions simultaneously, enabling long-range dependency modeling. Introducing quantum concepts of superposition and entanglement allows compact encoding of MRI feature spaces. Across six architectures, the ImageNet-1k–pretrained Swin-ViT achieved ~99% accuracy (AUC≈1.0), while QCNN-based models remained near chance (~25–30%). Errors concentrated between glioma and meningioma, while no-tumor and pituitary were near-perfect, indicating classical pretraining dominates on clean MRI.

RESEARCHERS: Vivaan Sheoran, Leigh High School ‘28; Pranav Das, New Albany High School '27
ADVISOR: McMahan Lab, Quantum Computing & Computer Science
KEYWORDS: Brain tumor detection | MRI classification | Swin Transformer | Convolutional Neural Networks | Quantum Machine Learning

Department of Chemistry, Biochemistry & Physics

Reactivity-informed pharmacophore editing and biological evaluation of andrographolide and synthesis of a-ring analogs: Closing the loop 

Andrographolide, a labdane diterpenoid isolated from the plant Andrographis paniculata, has been studied as an anticancer small molecule that putatively functions through modulation of the NF-κB signaling pathway. Functionalization of the C-19 hydroxyl of andrographolide has been shown to alter the primary mode of action to modulation of the Wnt/B-catenin signaling pathway from NF-KB inhibition (Reabroi et al. Biomedicine & Pharmacotherapy 2018). We identified that andrographolide A-ring trityl and silyl ethers exhibit greater potency than the natural product in in vitro human colon and breast cancer models (Gu, et al. Bioorganic and Medicinal Chemistry Letters 2025). With this established insight into A-ring modifications on andrographolide’s biological activity in both Wnt1 and NF-κB signaling, as well as the unique overlap of Human Endogenous Retrovirus-K in the two aforementioned pathways, we sought to study the effects of our analogs on the rapid cell proliferation and metastasis associated with HERV-K in breast cancer. Through stably transfected murine cancer cell lines, which exhibit fundamentally different mechanisms of canonical Wnt1 signaling, we observed that our analogs exhibited similar trends in anticancer potency in murine colon carcinoma, breast cancer, and melanoma cell lines as in their human counterparts. In a subsequent investigation to access a natural co-isolated A-ring oxetane (Jantan et al., Phytochemistry 1994), we developed a five-step synthetic sequence to access a related cyclic sulfite of both diastereomers, of whose structures were unequivocally established by 2D NMR spectroscopy and X-ray diffraction, which were evaluated alongside a broader library of cyclic A-ring analogs. These studies demonstrate that the andrographolide A-ring is a tunable site for modulating biological potency and mechanism of action in the context of cancer drug discovery.

RESEARCHERS: Aileen Pak, Castro Valley High ‘27
ADVISOR: Njoo Lab, Synthesis | Physical Organic Chemistry | Catalysis | Chemical Biology | Spectroscopy | Medicinal Chemistry
KEYWORDS: Natural Products | Organic Chemistry | Cancer Therapeutic | Pre-clinical Studies | Medicinal Chemistry 

Department of Chemistry, Biochemistry & Physics

Installation of Electrophilic Alkyl Amides and Sulfonamides Modulate the Anticancer Activity of Palbociclib and Ribociclib Towards Achieving Isoform-Selective CDK inhibition

Given their central role in aberrant cell cycle regulation found in many cancer types, cyclin dependent kinases (CDKs) have become an increasingly important target in the development of anticancer therapeutics. Recent efforts in this space have led to the discovery and development of palbociclib, an inhibitor of the closely related isoforms CDK 4 and 6. While these compounds have demonstrated remarkable clinical efficacy, potency, and safety tolerability profiles, their limited selectivity between CDK 4 and 6 pose unique target tolerability challenges. To address this challenge, and in efforts towards the development of similar compounds with improved CDK 4 and 6 isoform discrimination, several have elaborated a solvent exposed piperazine handle to either install a covalent warhead or other functional motifs that may imbue greater isoform selectivity. Inspired by these and other approaches, our laboratory has prepared a series of alkyl, alkyl halo, alkynyl amide, and sulfonate derivatives of palbociclib. One lead compound, bearing a brominated alkyl sidechain, exhibits a remarkably different in vitro potency profile compared to palbociclib its parent compounds. Among several analogs prepared, palbociclib exhibited an IC50 of 9.17 μM in HEK293 cells, while its 4-azidobutyrate amide was found to be slightly more potent, with an IC50 of 5.35 μM. This initial hit has prompted the synthesis of palbociclib and ribociclib derivatives in MDA-MB-231 human breast cancer cells. Here we disclose the chemical synthesis of over a half dozen palbociclib and ribociclib derivatives with both alkylating and non-alkylating motifs on their shared piperazine handle and current progress towards isoform selective biological applications of these compounds as therapeutic leads for the treatment of cancer

RESEARCHERS: Srinidhi Venkatesh, Irvington High School '27
ADVISOR: Njoo Lab, Synthesis | Physical Organic Chemistry | Catalysis | Chemical Biology | Spectroscopy | Medicinal Chemistry
KEYWORDS: Palbociclib | Ribociclib | anticancer activity | cyclin-dependent kinase inhibitors | isoform-selective CDK inhibition

Department of Chemistry, Biochemistry & Physics

Advancing Environmental Mapping and Crop Health Assessments with Drone Imaging

With the increasing concern for environmental conservation, there is a growing need for efficient methods of environmental mapping and forest health assessments. However, traditional methods employed by the U.S Forest Health Monitoring have faced controversy due to limited spatial resolution and integration of modern technologies. This research paper explores the application of machine learning algorithms in autonomous drones to conduct forest health assessments. Autonomous drones have the ability to collect timely, up-to-date data, which offers enhanced accuracy. This study focuses on training Deep Learning (DL) models to classify different environmental features based on aerial imagery captured by drones. To achieve accurate and efficient data collection, we will utilize Red-Green-Blue imaging and Convolutional Neural Networks (CNN) with the appropriate evaluation metrics, such as the Normalized Difference Vegetation Index (NDVI), Leaf Area Index (LAI), and foliage color, to create tree classes and identify forest health indicators. By integrating machine learning algorithms into forest health assessment, this study provides a more efficient, accurate, and up-to-date approach to monitor and evaluate the well-being of forests—supporting ongoing efforts towards environmental management and conservation.

RESEARCHERS: Mina Iqlas, Foothill High School '28; Arnav Rao, Mission San Jose High School '28; Neel Sreenivasa, Mission San Jose High School '29
ADVISOR: McMahan Lab, Quantum Computing & Computer Science
KEYWORDS: Autonomous Drones | Machine Learning | Crop Health Assessment | Environmental Mapping | Deep Learning

Department of Computer Science & Engineering

Recovering Strategic Phase Structure in Recurrent Multi-Agent Socio-Economic Systems

Recurrent multi-agent learning systems are increasingly applied to sequential decision problems involving asymmetric-role interactions and delayed feedback, yet standard episodic-return evaluation cannot determine whether learned behavior recovers the qualitative strategic structure of the environment. We study this question using a two-country trade conflict as a test bed, with phase-topology recovery as the primary evaluation criterion. We establish a tabular double Q-learning baseline as a reference phase topology for evaluating recurrent multi- agent architectures under identical economic dynamics. Naive recurrent deep Q-networks fail to recover this topology due to three systematic failure modes: reward-scale domination, credit assignment collapse across factorized instruments, and recurrent overwrite of sparse strategic signals. We then identify a set of architectural conditions that restore topology recovery, including factorized action heads, a tau-conditional stateless retaliation pathway, and decomposed Bellman targets, instantiated in the Factorized DRQN (FactorDRQN). The resulting system recovers 59 of 64 cells (92.2%) of the reference phase topology on the pri- mary structural parameter grid, a result that is unchanged under both five-seed and ten-seed aggregation protocols, and achieves 93.8% agreement under a second independent perturbation axis. This work contributes a phase-topology evaluation framework, a taxonomy of three recurrent-learning failure modes in strategic environments, and architecture conditions for consistent topology recovery.

RESEARCHERS: Nikhil Muthukumar, Archbishop Mitty High School '27, Micah Chiang, Valley Christian High School '27, Jeffrey Chen, Santa Clara High School '28
ADVISOR: Mui Lab, Computer Science, Machine Learning
KEYWORDS: Multi-agent reinforcement learning | recurrent deep Q-networks | phase-topology evaluation | socio-economic conflict | factorized architecture | regime structure

Department of Chemistry, Biochemistry & Physics

Development of Niclosamide Prodrugs as Cancer Therapeutics Reveals Rapid Bioactivation and Conserved Anticancer Activity in Human Cancer Cells

Niclosamide is an FDA-approved salicylanilide derivative initially discovered as an anthelmintic agent, possessing numerous biological targets. Due to its tendency to enter lipid bilayer membranes of mitochondria, endosomes, and lysosomes, it acts upstream of many various pathways, which allows it to act as an anticancer agent through STAT3, NF-κB and Wnt signaling pathways. However, one of the limiting factors for niclosamide usage in anti-cancer studies is its poor water solubility (0.23μg/mL), prompting new research to reformulate niclosamide for increased bioavailability. Recently, a niclosamide stearate prodrug therapeutic demonstrated heightened potency in in vivo evaluation against osteosarcoma (Reddy et. al. Cancer Research. 2020). In addition, a series of novel O-alkylamino-tethered derivatives of niclosamide exhibited highly improved aqueous solubility which improved biological performance (Chen et. al. ACS Med. Chem. Lett. 2013). To further test the ability of niclosamide analogs to inhibit STAT3 and Wnt1 pathways, we have synthesized a total of 21 niclosamide analogs with varying ester and carbonate tail lengths, utilizing simple and efficient one-step acylation and esterification reactions. In a series of cell proliferation assays and reporter cell assays, we have seen moderate dose dependent inhibitory activity in HCT-116, HT29, MDA-MB-231, and A549 cell lines. Additionally, through our hydrolysis assays, we have concluded that all compounds seem to demonstrate similar anticancer potency, given their ability to hydrolyze into niclosamide in a short period of time under physiological pH. Thus, we have demonstrated that our library of 21 niclosamide carbonate and ester analogs act as anticancer prodrugs against multiple human cancer cells.

RESEARCHERS: Hyunseo Claire Kim, Los Gatos High School '27; Maria Pavlidou, Carlmont High School '29
ADVISOR: Njoo Lab, Synthesis | Physical Organic Chemistry | Catalysis | Chemical Biology | Spectroscopy | Medicinal Chemistry
KEYWORDS: Niclosamide | Anticancer Agent | Drug Repurposing | Prodrug Development | Medicinal Chemistry | Cancer Research | Cancer Therapeutics | Small Molecule Drugs

Department of Computer Science & Engineering

Mechanical properties and fracture surface line roughness determination for 3D printed PLA in two sizes and orthogonal printing orientations

With the growing popularity of additive manufacturing, an alternative technology to traditional manufacturing, its application has been used in a range of fields from structural and civil engineering to dental and precision manufacturing. The knowledge of reliable mechanical properties is important for prototyping, computational modelling and final product quality assurance. Recent studies have shown that the mechanical properties of 3D printed parts depend on the printing parameters, color or brand of the feedstock, printing parameters and even brand of the printer. Both Prusa and Bambu printers have been used for 3D printing of tensile test specimens in this research. Print orientation as one of the most critical factors to the properties of 3D printed parts is the focus of our study. The mechanical properties determined from tensile tests between two printing orientations (vertical and horizontal) of different thicknesses (7mm and 3.2mm) are compared in terms of orientation, size and printer brand. The fractured surfaces after tensile testing for each orientation and thickness size were evaluated and line roughness determined. Fracture surface morphology and line roughness measurements are consistent with mechanical testing results: vertically printed samples are more brittle then horizontally printed parts. Fracture surface morphology and line roughness measurements for two different sizes, 3.2mm vs 7mm, does not seem to discern a fine difference in mechanical properties resulting from tensile test measurements. Parts printed with Bambu appear to be stronger than printed with Prusa, however the difference is rather nonessential.

RESEARCHERS: Annika Hegde, Leland High School '27

ADVISOR: Starostina Lab, Materials Science 

KEYWORDS: Additive Manufacturing | Print Orientation | Mechanical Properties

Department of Computer Science & Engineering

Machine Learning in 0D Nanostructure SEM Image Analysis: Count, Morphology, Size

Artificial intelligence has quickly risen to the forefront of research these past few years, and continues to accelerate in that growth. That growth is not only in everyday usage, but in also assisting in research as well. Electron microscopy is no exception to this trend. Manning Scanning Electron Microscope (SEM) was, and still is, incredibly time consuming. As a result, creating large datasets that are accurate tends to take a seemingly excessive amount of time. Research in recent years has utilized AI to be quicker and more accurate than manning the SEM manually. Our aim is to use machine learning techniques to make more efficient, and maybe even automize, the process of using and analyzing SEM scans. To do this, we will train different models for different kinds of nanostructures that are scanned by the SEM. Using different datasets of nanoparticles, nanopores, nanowires, ordered and nonordered, spherical, and cuboidal scans, we can use these to train our models. Eventually we aim to combine the models in a mixture of experts to be able to identify all types of nanostructures on SEM scans. Model performance is currently being evaluated using precision, recall, and intersection-over-union metrics to assess detection accuracy. The end goal is to hopefully bring an AI model that can be utilized by all SEM and users of SEM. And for them to use such a model to automize/make more efficient both the process of operating and analyzing SEM and their scans.

RESEARCHERS: Carter Tsao, Harvard-Westlake '27

ADVISOR: Starostina Lab, Materials Science 

KEYWORDS: Scanning Electron Microscope | Microscopy | Machine Learning | Nanostructures | Materials Science

Department of Computer Science & Engineering

Detection of Surface Level Cyanobacterial Algae in Freshwater Lakes using Cost Effective RGB Autonomous Unmanned Aerial Vehicles

Toxic cyanobacteria pose a significant health risk to humans through the consumption of poisoned fish or shellfish and hinder recreational activities. Harmful algae blooms (HABs), categorized by a deviation of regular algal biomass, develop when toxic or nontoxic cyanobacteria are exposed to a combination of nutrient runoffs and increased water temperature. Algae blooms deplete the lake of oxygen, causing hypoxia, killing or harming aquatic animals that require aquatic respiration, and killing underwater plants by obstructing sunlight. The rise of algal blooms in frequency is forcing lake managers and government bodies to constantly monitor algal levels and concentrations in their lakes. We aim to determine whether a sole RGB sensor can accurately quantify algae surface levels over a large period of time through autonomous waypoint flying and camera triggers. Images collected are plugged into a CNN (YOLO based model) for bounding boxes of algae areas. Polygons of algae-detected zones will be displayed on a public algal dashboard page. Ultimately, our research contributes to the growing body of autonomous UAVs for algal detection through the creation of a more practical and simpler solution, saving lake managers and governmental bodies time, costs, and personnel.

RESEARCHERS: Jeremiah Welch, Credo High '28; Matthew Chang, The King's Academy '27

ADVISOR: McMahan Lab, Quantum Computing & Computer Science

KEYWORDS:  Algae, Algae Detection, Mapping, Cyanobacteria, RGB, Unmanned Aerial Vehicles, Autonomous, Cost-Effective, Convolutional Neural Network

Department of Computer Science & Engineering

Assessment of GAN-Based Models for Synthetic Pneumonia Chest X-Ray Generation and Quality Improvement

Medical imaging, such as chest X-rays, is crucial for diagnosing pneumonia. However, image noise and poor resolution often hide vital details, limiting the effectiveness of AI-based diagnostic tools. To address this, we used Generative Adversarial Networks (GANs) to synthesize and enhance chest X-rays. GANs use two competing neural networks to learn how to produce highly realistic medical scans. In this study, we trained several models including a baseline Simple GAN, a Deep Convolutional GAN (DCGAN), and a Wasserstein GAN (WGAN) on a dataset of pneumonia X-rays. We evaluated them using the Frechet Inception Distance (FID) score, where a lower score indicates a more realistic image. Our results show that the DCGAN architecture outperforms the alternatives, achieving the lowest FID scores and generating the highest-quality X-rays. This demonstrates that DCGANs are a powerful tool for improving medical image datasets and future diagnostic accuracy.

RESEARCHERS: Arya Addagarla, Amador Valley High School '27; Pranav Pulavarthi, Foothill High School School '27; Kayla Wijesekera, Holy Names Academy, '27; Krish Puthran, Monta Vista High School, '28

ADVISOR: Viktoriia Liu Lab, Chemistry & Computer Neurobiology & Explainable AI & Augmented Reality

KEYWORDS:  Medical Imaging | Generative Adversarial Networks | Machine Learning | Chest Radiography | Image Enhancement

Department of Computer Science & Engineering

Detection of Surface Level Cyanobacterial Algae in Freshwater Lakes using Cost Effective RGB Autonomous Unmanned Aerial Vehicles

Toxic cyanobacteria pose a significant health risk to humans through the consumption of poisoned fish or shellfish and hinder recreational activities. Harmful algae blooms (HABs), categorized by a deviation of regular algal biomass, develop when toxic or nontoxic cyanobacteria are exposed to a combination of nutrient runoffs and increased water temperature. Algae blooms deplete the lake of oxygen, causing hypoxia, killing or harming aquatic animals that require aquatic respiration, and killing underwater plants by obstructing sunlight. The rise of algal blooms in frequency is forcing lake managers and government bodies to constantly monitor algal levels and concentrations in their lakes. We aim to determine whether a sole RGB sensor can accurately quantify algae surface levels over a large period of time through autonomous waypoint flying and camera triggers. Images collected are plugged into a CNN (YOLO based model) for bounding boxes of algae areas. Polygons of algae-detected zones will be displayed on a public algal dashboard page. Ultimately, our research contributes to the growing body of autonomous UAVs for algal detection through the creation of a more practical and simpler solution, saving lake managers and governmental bodies time, costs, and personnel.

RESEARCHERS: Jeremiah Welch, Credo High '28; Matthew Chang, The King's Academy '27

ADVISOR: McMahan Lab, Quantum Computing & Computer Science

KEYWORDS:  Algae, Algae Detection, Mapping, Cyanobacteria, RGB, Unmanned Aerial Vehicles, Autonomous, Cost-Effective, Convolutional Neural Network

Department of Computer Science & Engineering

Applications of Quantum Annealing in Cybersecurity

Adversarial training stands as the most effective defense against adversarial attacks in machine learning, yet its efficacy is significantly constrained by the diversity and quality of the attacks used during training. Traditional classical methods, such as the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD), are inherently local and restricted to gradient-following paths, which leave large regions of the adversarial perturbation space unexplored. Our project proposes a hybrid quantum-classical framework that leverages quantum annealing to overcome these local geometric limitations. By reformulating the adversarial perturbation search as a Quadratic Unconstrained Binary Optimization (QUBO) problem, we apply the combinatorial search capabilities of D-Wave’s quantum architecture to discover qualitatively distinct, non-local adversarial examples. The models will be benchmarked on standard datasets (MNIST/CIFAR-10) against rigorous robustness metrics, alongside an analysis of runtime and hardware efficiency. We aim to determine whether expanding the adversarial search space through quantum optimization yields classifiers with superior, more generalized adversarial robustness.

RESEARCHERS: Nitya Pisolkar, Archbishop Mitty High School '27

ADVISOR: McMahan Lab, Quantum Computing & Computer Science

KEYWORDS:  Machine Learning | Cybersecurity | Adversarial Attacks | Quantum Annealing

Department of Computer Science & Engineering

Autonomous UAV-Based Photogrammetry and Machine Learning for Predictive Coastal Erosion Analysis

Coastal erosion, the gradual wearing away of coastline land because of wave action, rising sea levels, and storms, poses a significant risk to coastal communities and habitats. As global sea levels continue to rise, the problem will only become more increasingly severe. This proposal aims to not only monitor coastal erosion but also use Artificial Intelligence and Machine Learning to predict its progression over time. By analyzing the data collected, this approach can highlight areas of concern, helping preserve infrastructure and natural habitats. This project will use an autonomous drone-based system built on a HolyBro X500 V2 Arf Kit which will be equipped with a M9N GPS, a CADDXFPV Farsight FPV Camera, and a Lidar Sensor to maintain altitude. Flight plans will be pre-determined using the GSHHG waypoint dataset and will be modified into a grid-shaped pattern. Images taken through this grid pattern will overlap 70-80% and will then be put through photogrammetry via open-source software like Meshroom to generate geotagged, timestamped DSMs. The training dataset will consist of historical coastline DSM data from sources such as USGS, customly labeled by subtracting the DSMs to get elevation change as a prediction metric. The live DSMs will be processed and analyzed using a Machine Learning model trained to detect shoreline changes, vegetation loss, and erosion patterns over time through the labeled dataset, outputting a final erosion score. A heatmap of areas at risk will then be generated, allowing coastline erosion mitigation efforts to be prioritized where they will have the greatest impact.

RESEARCHERS: Melody Dai, Basis Independent Silicon Valley High School '27; Srihari Anoop, American High School '28; Akshaj Seetharaman, Tilden Preparatory School '27; Sai Sanjay Devi, American High School '28; Hasini Enugu, Dougherty Valley High School '28; Akshara Gunturi, Dougherty Valley High School '29

ADVISOR: McMahan Lab, Quantum Computing & Computer Science

KEYWORDS:  Coastal Erosion | Machine Learning | Unmanned Aerial Vehicles | Remote Sensing | Mechanical Engineering