Advancing Computational Models and Algorithms for the Analysis of Single-Cell and Neural Data

Abstract

Single-cell technologies and neural activity analyses are at the forefront of modern biological research. Advances in sequencing have provided unprecedented resolution into cellular activities and gene regulation, while understanding brain activity is critical for uncovering how we sense the world, providing a theoretical foundation for developing modern biomedical applications, such as brain-chip interfaces. Simultaneously, breakthroughs in artificial intelligence (AI) have further revolutionized our ability to analyze these complex biological systems. Leveraging these advancements, this thesis presents a series of innovative computational models and algorithms that address challenges in single-cell sequencing data and explore the encoding and decoding of neuronal signals. These contributions establish new frameworks for extracting meaningful insights from these systems. In the first section of works, which focus on single-cell sequencing, I apply machine learning and deep learning algorithms to improve doublet detection and infer gene regulatory networks (GRNs) from single-cell RNA sequencing (scRNA-seq) data. I developed ImageDoubler, an image-based algorithm designed to enhance the detection of doublets in single-cell sequencing data, achieving significant improvement in F1 scores over traditional methods when benchmarking on the datasets from the Fludigm C1 platform. I also introduce PSGRN, our winning solution of the 2023 CausalBench Challenge, which infers GRNs using single-cell perturbational data that employs a self-training mechanism with synthetic gold standards to integrate observational and interventional data, addressing the scalability limitations of previous approaches. The second section of work shifts to the field of neuroscience, where advanced AI models are applied to encode and decode the neuronal signals from the primary visual cortex (V1), creating "digital twins" for understanding neural activity in silico. In this section, I first present a deep learning model that encodes the visual stimuli inputs to the single-neuron responses of mouse V1. This model ranked first in the SENSORIUM 2022 Challenge by incorporating object positions and ensemble learning techniques. Next, I explore the reverse direction, decoding neuronal responses to reconstruct visual stimuli. By combining grid-based embeddings with a conditioned Diffusion Transformer (DiT), I demonstrate superior reconstruction quality compared to state-of-the-art methods and efficient cross-subject inferences. Together, these contributions underscore the transformative potential of computational modeling and AI in single-cell data analysis and neural signal processing, paving the way for deeper biological insights and innovative biomedical applications.