The protozoan parasite Giardia lamblia infects hosts via the ingestion of cyst-contaminated water. We recently identified EncystR, a 7-trans membrane protein at the cell surface that acts as a negative regulator of encystation and responds to encystation cues by internalizing. By following the localization of EncystR through an encystation timecourse we discovered a novel compartment of unknown function. The Giardia endocytic pathway was only believed to include hybrid endosome/lysosome compartments statically positioned beneath the plasma membrane. We suspect this newly identified compartment may be a lysosome-like compartment. Using EncystR as a marker for this compartment we found that the compartment is highly acidified based on the florescent reporter pHluoren2. Canonically endomembrane compartments are marked by specific phosphatidylinositol phosphates (PIP). Relevant to endomembrane compartments, PI(3)P marks endosomes and PI(3,5)P2 marks multi-vesicular bodies and lysosomes. To test if this newly identified compartment is evolutionarily related to lysosomes, we will generate reporters for these phosphoinositides. Namely, we utilized the FYVE protein’s PIP binding domain as a sensor for PI(3)P which is located on endocytic membranes, the pH domain of PLC delta as a sensor for PI(4,5)P2 or PIP2 which is necessary for endocytosis and membrane-based cytoskeletal protein regulation, and ML1N for PI(3,5)P2 which is localized to lysosomes. Additionally, a mutant variant of the ML1N protein which is incapable of binding to PIPs is utilized as a negative control. These protein sensors were fused to the fluorescent protein mNeonGreen and imaged via fluorescent microscopy. While the experiment is currently in progress, we hypothesize that the localization of PIPs to the novel compartment will likely feature PI(3,5)P2 due to the previously found acidic nature of the compartment, implying a lysosome-like functionality. This possibly novel or conserved compartment could give insight into the evolution of eukaryotic cellular organisms and could potentially offer treatment routes for Giardia.

Since FPGAs (Field-programmable gate arrays) can achieve specific tasks faster and consume less power than general CPUs or GPUs, converting machine learning models to FPGAs has become more popular nowadays. However, people cannot directly deploy a floating-point model onto an FPGA due to resource limits (DSP, BRAM, etc.) on FPGAs. To solve this problem, some processes need to be done to shrink the models' size. Quantization is one method to reduce the models' size by rounding the floating-point calculations into a lower-bit representation. How to quantize a model without losing precision has become an interesting area of study. Our presentation focuses on configuring Qkeras (a tool designed to quantize TensorFlow Keras models) and hls4ml (a package designed to generate equivalent Verilog code based on the original model) to quantize CNN and RNN models before deploying them onto FPGAs. Our results will show that in the hls4ml process, if we quantize the Keras model into a Qkeras model before converting it into an hls model, we will need fewer bits compared to directly converting a Keras model into an hls model. In general, by studying the quantization of machine learning models, we can deploy more powerful AI into hardware such as FPGAs. The need for AI in everyday life has increased significantly, and this is a possible way to deploy AI in small items such as cameras, televisions, or even furniture.

The ForwArd Search ExpeRiment (FASER), located 480 m downstream from the ATLAS beam interaction point, is designed to study light weakly-interacting long-lived particles (LLPs) and neutrino interactions. One of the main components of FASER, used to detect the decay products of LLPs, is the four tracking stations, each with three layers and 24 semiconductor tracker (SCT) modules. Particles leave electronic hit signals on SCT modules, and particle tracks reconstructed from hits are used for analysis. To exploit the excellent intrinsic resolution of the silicon microstrip detectors, high-accuracy alignment is required. A local χ2 alignment algorithm is designed and tested on both Monte-Carlo and collision data to perform software alignment on the tracking stations. By aligning the FASER tracking stations, the performance of track reconstruction from detector hits is substantially improved, indicating reconstructed tracks are more likely created from their corresponding hits. Furthermore, the alignment allows researchers to match particle track information with other detectors. For the showcase for my research, the software alignment shows promising results when compared to survey data.