Skip to content

seed_perturbation_and_generation.md

Latest commit

 

History

History
148 lines (109 loc) · 5.12 KB

seed_perturbation_and_generation.md

File metadata and controls

148 lines (109 loc) · 5.12 KB
Raw

Seed Perturbation and miRNA Generation

Overview

scripts/seed_perturbation_and_generation.py simultaneously mutates every nucleotide inside the seed match region of each mRNA sequence in the 30-nt validation set (one random alternative per base, all at once) and generates a candidate miRNA from each perturbed mRNA using the trained TargetGenerationModel.

What the Script Does

1. Load the Validation Dataset

TargetScan_dataset/positive_samples_30_random_samples_validation.csv

19 951 positive miRNA–mRNA pairs with 30-nt mRNA windows. Each row contains the columns:

Column Description
Transcript ID Ensembl transcript identifier
miRNA ID miRBase identifier
miRNA sequence Full miRNA sequence (RNA, 5'→3')
mRNA sequence 30-nt 3'-UTR window (DNA, T-based)
seed start 0-based start of the seed match in the mRNA window
seed end 0-based end of the seed match (inclusive)
label 1 = positive pair

2. Build the Perturbed mRNA Catalogue

For each sample, every base in positions [seed_start, seed_end] (inclusive) is replaced simultaneously with a randomly-chosen alternative nucleotide (any of A, T, C, G excluding the original):

seed region positions:  seed_start, seed_start+1, …, seed_end
each position:          1 random alternative base
records per sample:     1  (entire seed region mutated in one pass)
total output rows:      19 951  (one-to-one with the input)

The perturbed DataFrame carries all original columns plus:

New column Description
perturbed_mRNA Full 30-nt mRNA with all seed-region bases simultaneously substituted

3. Load the TargetGenerationModel Checkpoint

Model configuration (identical to the training run in transformer_model.py):

Hyper-parameter Value
mrna_max_len 30
mirna_max_len 26 (= 24 bases + BOS + EOS)
embed_dim 256
ff_dim 512
num_layers 2
num_heads 2
vocab_size / n_classes 13
dropout_rate 0.1

Checkpoint path:

checkpoints/TargetScan/TwoTowerTransformer/CNN-tokenized/
    TargetGeneration/30/best_token_accuracy_0.9554_epoch19.pth

This checkpoint achieves 95.54 % token accuracy on the validation set after 19 training epochs.

4. Generate miRNAs (Greedy Decoding)

Each perturbed mRNA is tokenised with CharacterTokenizer (character-level, T-based DNA, right-padded to length 30 with an appended [EOS] token) and passed through the model's encoder–decoder in batches of 64.

Greedy decoding procedure (TargetGenerationModel.greedy_generate):

  1. Encode the perturbed mRNA once with the CNN-augmented TransformerEncoder.
  2. Start the decoder with a single [BOS] token.
  3. At each step take the argmax of the last-position logits and append the predicted token.
  4. Stop when all sequences in the batch have emitted [EOS] or the maximum length (26) is reached.
  5. Strip [BOS] from the output; [EOS] and [PAD] tokens are removed during decoding (skip_special_tokens=True).

The resulting sequences are in DNA notation, 3'→5' direction (matching the reference file convention; RNA U was converted to T and the sequence was reversed 3'→5' during tokenisation).

5. Save Output

TargetScan_dataset/generated_mirna_seed_perturbation_30_random_samples_validation.csv

Output Format

The output CSV has the same columns as the input plus two additional columns:

Transcript ID, miRNA ID, miRNA sequence, mRNA sequence,
seed start, seed end, label,
perturbed_mRNA, generated_mirna

Example row (seed region positions 16–23 fully mutated in one pass):

Transcript ID miRNA ID miRNA sequence mRNA sequence seed start seed end label perturbed_mRNA generated_mirna
ENST00000324344.4 hsa-miR-182-5p UUUGGCAAUGGUAGAACUCACACU GCTCTTTCCCCCATCTTTGCCAAATCTCAA 16 23 1 GCTCTTTCCCCCATCTTTACGTACATCTCAA …
… … … … … … … … …

How to Run

cd /home/mcb/users/jgu13/projects/MiRformer/scripts
python seed_perturbation_and_generation.py

Adjust DEVICE, BATCH_SIZE, and OUTPUT_CSV at the top of the script as needed. The script prints progress to stdout and writes one CSV file when finished.

Design Choices

Choice Rationale
Inclusive seed window [seed_start, seed_end] Matches TargetScan annotation convention used throughout the dataset
Simultaneous random perturbation of all seed bases Simulates a fully disrupted seed match in one step; avoids the combinatorial explosion of exhaustive single-base mutagenesis
One output row per input sample Keeps output size identical to the validation set (19 951 rows), making downstream comparisons straightforward
Greedy decoding Deterministic; consistent with the original generation pipeline in transformer_model.py
Batch size 64 Balances GPU memory and throughput on the 30-nt model
Output appended to existing columns Preserves full traceability — each generated miRNA can be linked back to the exact substitution that produced it