This repository contains homework assignments I designed as a Teaching Assistant for the HSE course Deep Learning in Bioinformatics. I am responsible for the bioinformatics track of the course, covering the application of modern deep learning methods to biological and medical data.
| # | Topic | Key Methods |
|---|---|---|
| HW2 | Medical Image Segmentation | U-Net, Dice Loss, XAI (Captum) |
| HW3 | DNA Promoter Prediction & Sequence Generation | DNABERT-2, REINFORCE, LSTM Policy |
Goal: Build an end-to-end semantic segmentation pipeline for medical images using U-Net.
Datasets (student's choice):
- BraTS2020 — Brain tumor segmentation on multimodal MRI (T1, T1ce, T2, FLAIR)
- ISIC2018 — Skin lesion segmentation on dermoscopy images
Topics covered:
- Segmentation metrics: Dice Score, IoU (Jaccard), Pixel Accuracy, Precision, Recall
- Data preprocessing and augmentation for medical images
- U-Net architecture with encoder-decoder and skip connections
- Dice Loss for handling class imbalance
- Experiment tracking with MLflow or Weights & Biases
- Model interpretability with Captum (attribution maps)
- Bonus: transfer learning and fine-tuning pretrained encoders
Goal: Fine-tune a genomic language model for promoter classification, then use reinforcement learning to generate novel promoter-like DNA sequences.
Datasets:
nucleotide_transformer_downstream(HuggingFace) — 300 bp promoter / non-promoter sequences
Topics covered:
- Biology background: promoters, TSS, TATA-box
- Fine-tuning DNABERT-2 with a classification head for promoter prediction
- Metrics: AUC, accuracy, confusion matrix, ROC curve
- RL environment design:
DNAEnvironmentwith step-by-step sequence construction - REINFORCE policy gradient with exponential moving average baseline
- LSTM-based policy network for nucleotide selection
- Baseline comparison: random generation vs. position-wise frequency matrix
- Bonus: PPO (Proximal Policy Optimization) for stable training
| Area | Libraries |
|---|---|
| Deep Learning | PyTorch |
| Genomic Models | HuggingFace Transformers, DNABERT-2 |
| Medical Imaging | albumentations, torchvision |
| Explainability | Captum |
| Experiment Tracking | MLflow / Weights & Biases |
| Evaluation | scikit-learn, matplotlib |
Institution: Higher School of Economics (HSE), Moscow Course: Deep Learning in Bioinformatics Role: Teaching Assistant