I work at the intersection of epigenomics and machine learning, building models that investigate and interpret histone modifications, gene regulation, and genomic variation. My background spans wet-lab biotechnology and computational research, which means I think about biology when I write code and think about code when I design experiments.

My MSc thesis focused on H3K27me3 in breast cancer. Since then, I've moved into deep learning for regulatory genomics, with a specific interest in how chromatin state predicts gene expression. I'm currently building toward a PhD in AI/bioinformatics.

This is a PyTorch replication of Singh et al. (2016). It uses a CNN trained on five histone marks (H3K4me1/3, H3K36me3, H3K9me3, H3K27me3). The training data comes from cell line E047, which consists of primary T CD8+ naive cells from REMC. The model predicts binary gene expression based on RPKM median splits across 19,300 genes. I replicated the original Lua/Torch7 architecture. Then, I added early stopping and ran 30 Optuna trials. The AUROC improved from 0.906 with the paper's hyperparameters to 0.920. This shows that the original hyperparameters were already close to optimal. The biggest improvement came from early stopping.

Python, PyTorch, CNN, Optuna, Scikit-learn

I built the AttentiveChrome architecture from scratch in PyTorch, based on Singh et al. (2017). Five BinEncoder BiLSTMs process each histone mark's bin sequence independently. A bin-level attention (α) reduces each mark to a single vector. Then, a mark-level attention (β) weights all five marks before the final classifier. The AUROC is 0.920 on E047, matching DeepChrome with Optuna tuning, but with added interpretability. The β weights indicate that H3K4me3 has a value of 0.345, while H3K27me3 is second with 0.221. This is consistent with bivalent chromatin at genes ready for activation in naive CD8+ T cells. The bin-level α for H3K36me3 peaked away from the TSS, reflecting its known gene-body biology. The model learned this spatial pattern without explicit positional guidance.

Python, PyTorch, BiLSTM, Attention Mechanism, Attention Visualization

Transformer-based model for histone mark → gene expression prediction

A follow-up to deepchrome that replaces the recurrent layers with a Transformer encoder. Explores whether self-attention alone can capture the long-range chromatin dependencies that BiLSTM handles sequentially.

• Direct comparison with deepchrome on the same benchmark
• Investigates positional encoding strategies for fixed-length genomic windows
• Part of an ongoing line of work on attention mechanisms for regulatory genomics

Python · PyTorch · Transformer · Self-Attention · Genomics

MSA and conservation analysis of SARS-CoV-2 genes from Egyptian patient samples

Multiple sequence alignment and conservation scoring across key SARS-CoV-2 genes (Spike, ORF1ab, N protein) using sequences collected from Egyptian COVID-19 patients. Identifies mutation hotspots relative to the Wuhan reference genome and tracks variant-specific substitutions.

• Egyptian patient cohort — locally relevant epidemiological context
• Conservation scoring with Shannon entropy across aligned positions
• Phylogenetic analysis to place Egyptian sequences in the global variant landscape

Python · Biopython · MSA · Phylogenetics · SARS-CoV-2 · Genomic Epidemiology

Languages Python R SQL Bash

ML / Deep Learning PyTorch TensorFlow Scikit-learn · Random Forest · XGBoost · Logistic Regression · CNNs · LSTMs · Transformers

Bioinformatics NGS analysis GATK BWA STAR DESeq2 Biopython · Variant calling · ChIP-seq · RNA-seq · Multiple sequence alignment

Tools & platforms Git Linux Jupyter Google Colab · ENCODE · UCSC Genome Browser

Publication: Badr M. et al. (2016). Procaine-induced epigenetic changes in HCT116 colon cancer cells. BioMed Research International.