Skip to content

Stage 5 — Training M2-S (Alternative-Site)

Pipeline position: features → this stage → evaluation → reporting

M2-S is the alternative-site refiner. Where M1-S learns the curated canonical (MANE) sites, M2-S is trained on the broader Ensembl annotation so it can recognize the alternative sites MANE leaves out — competing donors/acceptors, minor isoforms, tissue-specific sites.

This doc is deliberately short: M2-S uses the same trainer as M1-S. Read Stage 4 first; below is only what changes.

What differs from M1-S

Labels: Ensembl instead of MANE (auto-selected by --mode m2). Base scores: the Ensembl-windowed precomputed dir (--base-scores-dir). Everything else — script, architecture, channels, hyperparameters — is identical.

The command

python examples/meta_layer/07_train_sequence_model.py \
    --mode m2 \
    --use-shards \
    --bigwig-cache <bigwig-cache-dir> \
    --base-scores-dir data/ensembl/GRCh38/openspliceai_eval/precomputed \
    --output-dir output/meta_layer/m2s_v4_cleanannot

--mode m2 does two things automatically: it stamps the variant as M2-S, and it switches the label source to Ensembl (annotation_source = ensembl). You still supply the Ensembl base-scores directory explicitly with --base-scores-dir, because the base model scored the Ensembl gene windows in Stage 2.

A discrepancy to be aware of

The pod runner ops_train_m2s_pod.sh invokes the trainer as --mode m1 --annotation-source ensembl. That trains on equivalent Ensembl data but stamps the config variant = M1-S. The promoted m2s_v4_cleanannot was produced with --mode m2 (stamped M2-S). Prefer --mode m2 — it sets the Ensembl label source for you and records the correct variant.

Why M2-S is the harder model

M1-S is trained and tested on the same curated distribution, so it scores very high. M2-S targets the out-of-distribution sites — the ones a MANE-only model systematically misses — which is a genuinely harder problem. That shows up in the numbers: the promoted M2-S reaches a validation macro PR-AUC of 0.953 (epoch 6), lower than M1-S's 0.998 but on a much broader, noisier label set. The right way to judge M2-S is not the aggregate but its lift on the alternative sites specifically — which is what the dedicated alt-site evaluation in Stage 6 measures.

The motivation and the architecture-search lessons behind M2-S are worth reading:

Artifacts

Identical layout to M1-S, under output/meta_layer/m2s_v4_cleanannot/. The companion alternative-site evaluation writes to a sibling m2s_v4_cleanannot_alt_eval/ directory (produced by 09 in the next stage).

→ Next: Stage 6 — Evaluation