The developing human brain is shaped by environmental stimuli, but how genetic diversity modulates cellular responses remains unclear. While prior studies have identified loci at which genetic variation is associated with gene expression differences (expression quantitative trait loci, eQTLs), interpretation of their effects is limited by cell-type specificity or usage of population-level signals that mask exposure-specific regulation. Few studies have explored how genetic variation influences stimulus-dependent responses, captured by response-expression QTLs (reQTLs), which reveal context-specific mechanisms missed in baseline studies. Here, we present a robust, generalizable, and novel reQTL pipeline combining statistical modeling, permutation testing, and functional annotation. To identify environmental stress response–associated variants, we created a “cell village” of fetal neural progenitor cells from 49 genetically diverse donors, pooled in vitro before cortisol exposure. We used single cell RNA-sequencing and demultiplexing algorithms to characterize the transcriptional responses of each donor. We then applied our reQTL pipeline and detected 3059 reQTLs associated with 148 genes. These include rs696, located in a microRNA recognition site on NFKBIA, a key regulator of glucocorticoid signaling. This scalable approach demonstrates strong power to uncover context-dependent regulatory variation even with modest sample sizes. Future work will focus on in vitro validation and continuous and combinatorial perturbations.

Magnetic resonance imaging (MRI) can be lengthy and patients often need to remain still for extended periods. Accelerating MRI scans by undersampling k-space data is a common strategy, but aggressive undersampling degrades image quality without advanced reconstruction methods. We investigated whether incorporating MRI physics priors into diffusion models can improve reconstruction from highly undersampled k-space data. We propose a k-space diffusion framework that integrates physics-informed inductive biases. Our approach combines complex-valued score networks to preserve magnitude and phase relationships, virtual conjugate coil augmentation to exploit Hermitian symmetry, and structured low-rank and SPIRiT-based auxiliary losses to enforce inter-coil consistency. The model is evaluated on multi-coil cardiac MRI data and compared against real-valued diffusion and conventional reconstruction baselines. We show that complex-valued diffusion improves reconstruction quality, achieving up to a 3 dB increase in PSNR and a 0.09 improvement in SSIM over real-valued diffusion models. Additional gains are observed when enforcing conjugate symmetry and coil consistency constraints, indicating that these priors better capture the structure of k-space. These findings demonstrate that aligning generative models with MRI physics priors contributes to more accurate reconstructions from limited data, supporting the development of faster and more clinically practical MRI acquisition protocols.

Photocatalysis—using light to drive chemical reactions—has emerged as a powerful strategy for controlling chemical reactivity. Despite this promising field of study, the molecular-level understanding required for designing better photocatalysts remains limited. This is in large part due because photochemical processes are difficult to simulate. This project aims to develop computational tools to address that challenge. To accurately describe photochemistry in complex environments, most existing methods use quantum mechanics/molecular mechanics (QM/MM) methods. QM/MM methods perform quantum chemistry calculations, most often Density Functional Theory (DFT) calculations, on the regions containing the chemically active part of the system while treating the surrounding environment with a classical force field. However, in order to get an accurate description of the potential energy surface (PES), many DFT calculations must be performed which is a fundamental setback of QM/MM. Due to this computational barrier, machine learning (ML) methods have become an appealing alternative to purely quantum mechanical approaches. These ML/MM—as opposed to QM/MM—models work by performing DFT on a subset of the trajectory geometries, then use these DFT calculations to train an ML model which learns the PES of the system. This project utilizes these machine learning approaches to bridge the computational barrier between quantum mechanical accuracy and the timescales needed to simulate photochemistry.

Machine-learning methods in biochemistry commonly represent molecules as graphs of pairwise intermolecular interactions for property and structure predictions. Most methods operate on a single graph, typically the minimal free energy (MFE) structure, for low-energy ensembles (conformations) representative of structures at thermodynamic equilibrium. We introduce a thermodynamically parameterized exponential‑family random graph (ERGM) embedding that models molecules as Boltzmann‑weighted ensembles of interaction graphs. We evaluate this embedding on SELEX datasets, where experimental biases (e.g., PCR amplification or sequencing noise) can obscure true aptamer–ligand affinity, producing anomalous candidates whose observed abundance diverges from their actual binding strength. We show that the proposed embedding enables robust community detection and subgraph‑level explanations for aptamer-ligand affinity, even in the presence of biased observations. This approach may be used to identify low-abundance aptamer candidates for further experimental evaluation.