PHACT-miRBind

This repo contains three runnable miRBind-style model families:

• PairwiseSeqCNN: the seq-only pairwise CNN baseline with the original 28x50 miRNA/target pair grid and 147,249 default parameters.
• PairwiseConservationCNN: the same pair-grid CNN with target-position phyloP, phastCons, or both appended as extra channels before the first convolution.
• PairwisePhactCNN: the pair-grid CNN augmented with position-specific PHACT channels from the miRNA, target, or both axes.

The original PHACTn workflows are kept under PHACTn/. The PyTorch cache, training, and model implementations are under src/phact_mirbind/.

Layout

src/phact_mirbind/
  cli/          command-line entry points
  data/         TSV row parsing, sequence normalization, pair encoding
  cache/        .pt cache writers, manifests, iterable datasets
  models/       seq-only, conservation, and PHACT CNNs
  training/     shared train/eval loop, metrics, logging

Setup

cd /home/dtzim01/PHACT-miRBind
uv sync

Cache

Seq-only training uses a neutral pair cache:

uv run build-pair-cache \
  --input-file data/presplit_original_rows/manakov_original_rows_train.tsv \
  --output-dir data/pair_cache_original_rows/train \
  --output-prefix train

Conservation training uses a pair + conservation cache:

uv run build-conservation-cache \
  --input-file data/presplit_conservation_original_rows/manakov_original_rows_train.tsv \
  --output-dir data/conservation_cache_original_rows/train \
  --output-prefix train \
  --conservation-features phylop,phastcons

phyloP is normalized as clamp(score / 10, -1, 1). phastCons is used as provided, with missing values filled as 0.5.

PHACT cache construction keeps the neutral score fill used by earlier runs and also stores one binary missingness channel per PHACT score group. Newly built caches append those masks to the model input, allowing a model to distinguish a real neutral score from an unavailable score. Caches produced before this addition remain readable and simply omit the mask channels.

Train

Run the seq-only baseline:

scripts/run_seq_original_split_params.sh

Run the conservation-channel model:

scripts/run_conservation_original_split_params.sh

Use only one conservation source by setting CONSERVATION_FEATURES:

CONSERVATION_FEATURES=phylop scripts/run_conservation_original_split_params.sh
CONSERVATION_FEATURES=phastcons scripts/run_conservation_original_split_params.sh

Build and train the PHACT-channel model:

scripts/build_phact_full_cache.sh
scripts/train_phact_full_cache.sh

train-phact-mirbind --phact-channel-mode selects mirna, target, or both. --phact-reduction on cache construction selects the full nucleotide scores, actual_margin, or alt_mean representation.

The scripts create/reuse:

• data/presplit_original_rows and data/pair_cache_original_rows
• data/presplit_conservation_original_rows and data/conservation_cache_original_rows

Test

uv run pytest