The question
How can we connect transcription factors, chromatin accessibility, and gene expression to study nonlinear regulatory patterns in Alzheimer’s disease?
For my undergraduate thesis, I developed a two-stage graph attention model using paired hippocampal single-nucleus RNA-seq and ATAC-seq data from the ROSMAP cohort. The source dataset contains 50,381 nuclei from 111 donors.
A regulatory path from TFs to genes
I constructed a directed TF → Peak → Gene graph with 115 transcription factors, 46,626 chromatin-accessibility peaks, and 2,921 genes. TF–Peak edges use JASPAR/FIMO motif evidence; Peak–Gene candidates lie within ±500 kb of a gene’s transcription start site. Aligning the two edge sets removes peaks without a complete two-hop path.
The model uses two eight-head graph attention layers and a differentiable gate based on predicted ATAC values. This allows the gene-expression loss to propagate back through the gate to the first attention layer. Cell-subtype and disease-stage embeddings provide conditional information. The architecture contains 88,660 trainable parameters.
My contributions
- Integrated the paired modalities, scanned motifs, and constructed and aligned the regulatory graph.
- Designed and implemented the two-stage attention model, differentiable gate, and multitask training objective.
- Compared prediction against non-graph baselines, tested architectural ablations, and analysed attention patterns and conditional perturbations.
Prediction: what was evaluated?
The main prediction experiment used a 5,000-cell subset, randomly split into 4,000 training and 1,000 validation cells. On the validation cells, mean per-gene R² was 0.9904 across 2,921 genes, with expression measured on the log1p scale. Mean Pearson correlation was 0.9952, while mean Spearman correlation was 0.6711.
The model had lower validation error than the compared per-gene mean, Ridge, and random-forest baselines. Its reported MSE was 0.0029, compared with 0.1178 for random forest. These are within-cohort prediction results under the thesis evaluation protocol.
Ablations: why use a learned gate?
Replacing the learned prediction gate with hard binary ATAC gating increased validation MSE to 0.2177. Setting edge weights to one while retaining the topology had a smaller effect, increasing MSE to 0.0037. In these experiments, the learned gate had the larger effect on prediction quality.
From prediction to biological hypotheses
I examined attention-derived TF scores across cell types and disease stages. The thesis’s separate stage analysis used 20,000 cells with a model trained on the full dataset; the heatmap highlights stage-associated patterns for candidates including MEF2C, RORA, and REL.
For a virtual perturbation, I changed the model’s disease-stage embedding to lateAD while holding the molecular inputs fixed. GO enrichment of the 200 genes with the largest predicted changes highlighted neuron-projection-development processes. This probes the model’s conditional predictions; it is not a biological intervention.
Earlier work in computational biology
Before this thesis, I contributed to knowledge-graph research for circRNA–disease association prediction (KGRACDA), and to data-processing experiments and simulations for polyploid genome-size estimation (findGSEP). These projects led to two co-authored publications in 2024.