// selected work

Projects

A collection of computational biology, deep learning, and population genetics work — each built with reproducibility and software engineering discipline in mind.

3Projects
2Languages
500+HPC Jobs
← QTL locus deme 1 deme 10 CNN kernel 0 -1 0 1 2 1 0 -1 0 P(adaptation) 0.87 100% val. accuracy · 3,000 simulations · ~27 epochs

// 01

✓ Model trained — 100% validation accuracy

DeepLocalAdaptation

Multi-branch CNN with attention mechanisms trained on 3,000 forward-time SLiM simulations to distinguish local adaptation from genetic drift. Achieves 100% validation accuracy in ~27 epochs — directly targeting the scenarios where classical QST–FST comparisons structurally fail.

PyTorch SLiM 5.0 Population Genetics SLURM Python
Case Study
input Conv Encoder μ, σ² z latent Conv Decoder recon recon t-SNE latent space Type A Type B Type C VAE architecture · t-SNE cluster projection

// 02

Complete

BioVAE-Phenotyper

Production-grade Convolutional Variational Autoencoder for unsupervised cell phenotype classification. Learns a compressed latent space from cell images without any labels, then projects it with t-SNE to reveal biological clusters. Built with CI/CD, Poetry, and Pytest from day one — software engineering discipline beyond academic standards.

PyTorch VAE GitHub Actions Pytest Poetry t-SNE
Case Study
30% 20% 10% 0% α=5% 1 5 10 50 100 1000 N QTLs Island model Stepping-stone False Positive Rate FPR vs. N QTLs — additive architecture FPR simulation results · 500 replicates per condition

// 03

✓ Thesis validated — UNIL 2025

QST–FST Simulation Framework

Large-scale Python/R simulation framework using SLiM 5.0 to stress-test the QST–FST evolutionary test across 1–1,000 QTLs, four trait architectures, and two population structures. 500 replicates per condition on SLURM HPC. Key finding: the test breaks down under stepping-stone structure regardless of polygenicity.

SLiM 5.0 Python R SLURM / HPC Population Genetics
Case Study