Add COMPASS transfer cohort and standalone preparation script

LICENSE-CODE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 The Authors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

LICENSE.md

@@ -0,0 +1,11 @@
+# Sources and licenses
+
+Third-party data retain their original terms. The code license does not replace the data licenses.
+
+| Material | Source and terms |
+| --- | --- |
+| COMPASS sequences and measurements in `compass-transfer/compass.tsv.gz` | Koplik et al., [COMPASS](https://doi.org/10.1101/2025.10.12.681955), [GEO GSE307247](https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE307247). CC BY-NC 4.0; [license](https://creativecommons.org/licenses/by-nc/4.0/). |
+| OpenSplice source selection | Quarantani et al., [OpenSplice experimental master table, version 5](https://doi.org/10.6084/m9.figshare.32337414.v5). CC BY 4.0; [license](https://creativecommons.org/licenses/by/4.0/). |
+| BEND 2.0 preparation code, documentation and exclusion metadata | The Authors, 2026. [MIT license](LICENSE-CODE). |
+
+Changes to the source data comprise the BEND 2.0 cohort selection, matching reference and alternate constructs, averaging available cell-line effects, and constructing the paired-repeat reference. The preparation script documents these changes. Please retain source attribution and applicable licenses when redistributing derived data.

README.md

@@ -0,0 +1,33 @@
+---
+pretty_name: BEND 2.0 extras
+language:
+- en
+license: other
+license_name: mixed-source-licenses
+license_link: LICENSE.md
+tags:
+- biology
+- genomics
+- splicing
+- benchmark
+- arxiv:2611.03806
+size_categories:
+- 100K<n<1M
+configs:
+- config_name: compass-transfer
+  sep: "\t"
+  data_files:
+  - split: test
+    path: compass-transfer/compass.tsv.gz
+  default: true
+---
+
+# BEND 2.0 extras
+
+Additional data and preparation scripts for the [BEND 2.0 paper](https://arxiv.org/abs/2611.03806).
+
+[Code](https://github.com/MachineLearningLifeScience/BEND2) · [Leaderboard](https://machinelearninglifescience.github.io/BEND2/) · [Main benchmark data](https://huggingface.co/datasets/yhsure/BEND2)
+
+**[OpenSplice → COMPASS transfer](compass-transfer/README.md)** — the Fig. 2B evaluation cohort, OpenSplice source-row list, exclusions and a standalone preparation script. Includes 4,618 COMPASS variants across 88 genes and the 282,610 OpenSplice substitutions used for fitting.
+
+See [source credits and licenses](LICENSE.md). File checksums are in [SHA256SUMS](SHA256SUMS).

SHA256SUMS

@@ -0,0 +1,11 @@
+9f1abe28a25b5f6f95c58a81a647de3fa72ff6388d228c59bddd079255709989  LICENSE-CODE
+664009adaa97f76ae2380e9a17ec19d5b8de3a312241fc145811aad90646101b  LICENSE.md
+63e4aa7347b274ed23ec1b545ff119d91e90a4d291773af7bf4737d8dba71d11  README.md
+39b84a8afd148d3242aaef7ba3ac9fa7c5a92983d3557f14e91d51122ec0d2d8  compass-transfer/README.md
+027f35fec031e194da26b36f2dfa429817c8916ef5394be2af6ebc156acb012f  compass-transfer/compass.tsv.gz
+94ead6d28f3431471e94de863730e8083e00f8b71cbc942afb42421f9415e501  compass-transfer/exclusions.json
+d09a3b736803e6569696c5a77de338829b0189245eee0c3dcb0b280f95edeb37  compass-transfer/opensplice-source.tsv.gz
+ab9a7d93cd4cdce29f733b489d93f55f292825a2c570a31cc337446ce617eb4a  compass-transfer/prepare.py
+d412c34408124d9680009f7e40684bfdd78535df45cec5ba8644fd88bca845da  compass-transfer/sequence-overlaps.csv
+20fafe717d7efd88b6b75e5c9859a5194bb7e91f620a038344649408b828963d  compass-transfer/sources.json
+686af41f9440708e67395568ff4eaef5708641f6e130f0da45e6ed188b02d94d  compass-transfer/target-genes.tsv

compass-transfer/README.md

@@ -0,0 +1,63 @@
+# OpenSplice → COMPASS transfer
+
+Data preparation for BEND 2.0 Fig. 2B and Appendix C. The evaluation contains **4,618 single variants in 88 genes** (97 exon constructs). Fitted readouts use **282,610 OpenSplice substitutions in 443 genes**.
+
+## Files
+
+| File | Contents |
+| --- | --- |
+| [compass.tsv.gz](compass.tsv.gz) | The exact evaluation cohort: variant IDs, genes, reference/alternate DNA, pooled effects and two repeat effects. |
+| [target-genes.tsv](target-genes.tsv) | All 88 target genes, with variant and exon counts. |
+| [opensplice-source.tsv.gz](opensplice-source.tsv.gz) | The 282,610 source record IDs and genes, in fitting order. |
+| [exclusions.json](exclusions.json) | The 23 shared genes and 12 excluded COMPASS sequence-overlap events. |
+| [sequence-overlaps.csv](sequence-overlaps.csv) | The frozen sequence-overlap audit behind those event exclusions. |
+| [prepare.py](prepare.py) | Reconstructs the cohorts from the original assay releases and checks them against these files. |
+| [sources.json](sources.json) | Pinned input URLs and checksums. |
+
+Read the cohort directly:
+
+```python
+import pandas as pd
+
+target = pd.read_csv("compass.tsv.gz", sep="\t")
+source_ids = pd.read_csv("opensplice-source.tsv.gz", sep="\t")
+```
+
+`record_id` is `event_id|mutation`. The mutation position is zero-based within the 161-nt assayed insert. `sequence_ref` and `sequence_alt` are the complete assayed inserts. `target_logit_delta` is the evaluation label; `target_contexts` counts the cell lines contributing to it.
+
+## Rebuild from the original releases
+
+Download this folder, then install the small CPU dependencies:
+
+```bash
+python -m pip install numpy pandas pyarrow
+```
+
+Download the [COMPASS processed table](https://ftp.ncbi.nlm.nih.gov/geo/series/GSE307nnn/GSE307247/suppl/GSE307247_Processed_PSIs_All_Cells.csv.gz) and [OpenSplice v5 table](https://ndownloader.figshare.com/files/64886694). OpenSplice is also available as `opensplice.tsv.gz` in the [main BEND 2.0 data release](https://huggingface.co/datasets/yhsure/BEND2). Verify the input hashes against `sources.json`.
+
+```bash
+python prepare.py \
+  --compass /path/to/GSE307247_Processed_PSIs_All_Cells.csv.gz \
+  --opensplice /path/to/open_splice_experimental_master_unique_sequence_all_predictor_cols.tsv.gz \
+  --out rebuilt
+```
+
+This writes `rebuilt/compass.parquet` and `rebuilt/opensplice-source.parquet`. The source table includes sequences and transformed fitting labels, not just the IDs. Omit `--opensplice` to rebuild only the evaluation cohort. Use a new output directory for each run.
+
+## Cohort construction
+
+The historical COMPASS preparation starts from complete WT/A/B/AB quartets: a wild-type construct, its two constituent single mutations and their double mutation must have nonmissing pooled measurements in at least one cell line. The transfer cohort takes the distinct single variants appearing in those quartets. Wild types are matched by `event_id`; single variants by event and mutation. Sequence identities and raw mutation labels are checked when joining the measurements.
+
+The 12 events in the frozen sequence-overlap audit are removed first. The remaining COMPASS genes are intersected with the OpenSplice substitution genes, and all 23 shared genes are removed from both source and target. OpenSplice fitting retains `mut_type == "sub"`; deletions from the broader Fig. 2A task are excluded. The source IDs preserve the original OpenSplice row indexing.
+
+## Measurements and readouts
+
+COMPASS labels use the release's `*_delta_logit_pooled` columns. Replicates are already pooled by the release; we average these effects across the available HeLa, K562, MCF7, HMC3 and HEK measurements. All five cell lines are not required. This is not an average of the two repeat columns supplied here.
+
+For the experimental reference, `rep1_logit_delta` and `rep2_logit_delta` are computed separately as alternate minus matched wild-type logit, then averaged across the same available cell lines. A cell line contributes only when both alternate and wild-type measurements exist in both repeats; `replicate_contexts` records that count.
+
+OpenSplice fitting labels are `logit(clip(PSI_alt / 100, .01, .99)) - logit(clip(PSI_ref / 100, .01, .99))`. Each embedding-based transfer readout is a separate ridge regression (α = 10, fitted intercept, no feature standardization) on alternate-minus-reference embeddings, fitted once on all 282,610 eligible source rows. The same frozen representation and projection are used for source and target. The transfer fit does **not** average the five Fig. 2A cross-validation heads or ensemble independently fitted checkpoints. Released native-output methods are evaluated separately. COMPASS labels are not used for fitting, calibration or model selection.
+
+The reported metric is the mean of within-gene Spearman correlations. Genes with constant outcomes have undefined correlations; 86 of the 88 target genes have varying outcomes. This release supplies the published cohort and its preparation, without model prediction arrays.
+
+Please cite [BEND 2.0](https://arxiv.org/abs/2611.03806), [OpenSplice](https://doi.org/10.6084/m9.figshare.32337414.v5) and [COMPASS](https://doi.org/10.1101/2025.10.12.681955). See [licenses](../LICENSE.md).

compass-transfer/compass.tsv.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:027f35fec031e194da26b36f2dfa429817c8916ef5394be2af6ebc156acb012f
+size 207036

compass-transfer/exclusions.json

@@ -0,0 +1,41 @@
+{
+  "shared_genes_removed": [
+    "AP4M1",
+    "APC",
+    "ASL",
+    "BIN1",
+    "BRCA1",
+    "CACNA1C",
+    "CFTR",
+    "COL11A2",
+    "COLQ",
+    "ELN",
+    "GNAS",
+    "HMBS",
+    "L1CAM",
+    "MEF2C",
+    "MYH11",
+    "NEK1",
+    "NF1",
+    "NRXN1",
+    "PDE6B",
+    "SLC26A4",
+    "TNNT2",
+    "TSC2",
+    "TSEN2"
+  ],
+  "excluded_overlap_events": [
+    "chr11:119088168-119088328:+",
+    "chr11:119092023-119092183:+",
+    "chr11:31801851-31802011:-",
+    "chr19:38570553-38570713:+",
+    "chr1:45332898-45333058:-",
+    "chr20:63431320-63431480:-",
+    "chr2:232527305-232527465:+",
+    "chr2:71549251-71549411:+",
+    "chr3:128806309-128806469:+",
+    "chr3:38557211-38557371:-",
+    "chr4:113348168-113348328:+",
+    "chrX:31177912-31178072:-"
+  ]
+}

compass-transfer/opensplice-source.tsv.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d09a3b736803e6569696c5a77de338829b0189245eee0c3dcb0b280f95edeb37
+size 4434521

compass-transfer/prepare.py

@@ -0,0 +1,202 @@
+"""Rebuild the BEND 2.0 Fig. 2B cohorts from the released assay tables."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+HERE = Path(__file__).resolve().parent
+CELLS = ("HeLa", "K562", "MCF7", "HMC3", "HEK")
+
+
+def require(condition, message):
+    if not condition:
+        raise ValueError(message)
+
+
+def dna(value):
+    value = str(value).upper().replace("U", "T")
+    require(not (set(value) - set("ACGTN")), "Invalid DNA sequence")
+    return value
+
+
+def prepare_compass(path, exclusions):
+    raw = pd.read_csv(path, low_memory=False)
+    raw = pd.concat([raw, raw.snp.map(lambda x: "".join(str(x).split())).rename("snp_norm")], axis=1)
+    wt = raw.loc[raw.seq_type.eq("none")].copy()
+    singles = raw.loc[raw.seq_type.eq("single")].copy()
+    require(not wt.event_id.duplicated().any(), "Duplicate wild-type event")
+    require(not singles[["event_id", "snp_norm"]].duplicated().any(), "Duplicate single variant")
+    wt = wt.set_index("event_id")
+    single_index = singles.set_index(["event_id", "snp_norm"])
+
+    # The historical canonical table contains complete WT/A/B/AB quartets.
+    # Recover its single-variant union, rather than taking every raw single.
+    identities = []
+    for double in raw.loc[raw.seq_type.eq("double")].itertuples(index=False):
+        mutations = double.snp_norm.split(";")
+        require(len(mutations) == 2, "Expected two mutations in a double")
+        event = double.event_id
+        keys = [(event, m) for m in mutations]
+        if event not in wt.index or any(k not in single_index.index for k in keys):
+            continue
+        reference = wt.loc[event]
+        a, b = (single_index.loc[k] for k in keys)
+        covered = any(all(pd.notna(v) for v in (
+            reference[f"{c}_pooled_logit"], a[f"{c}_pooled_logit"],
+            b[f"{c}_pooled_logit"], getattr(double, f"{c}_pooled_logit"),
+        )) for c in CELLS)
+        if not covered or event in exclusions["excluded_overlap_events"]:
+            continue
+        require(int(mutations[1].split(":")[0]) > int(mutations[0].split(":")[0]),
+                "Double components must be ordered by position")
+        require(len(dna(double.full_seq)) == 161, "Expected a 161-nt double construct")
+        for mutation, alternate in zip(mutations, (a, b)):
+            identities.append((str(event), str(double.gene_exon).split(" exon ", 1)[0],
+                               dna(reference.full_seq), dna(alternate.full_seq), mutation))
+    target = pd.DataFrame(identities, columns=[
+        "event_id", "gene_id", "sequence_ref", "sequence_alt", "mutation",
+    ]).drop_duplicates()
+    require(not target[["event_id", "mutation"]].duplicated().any(),
+            "Inconsistent gene or sequence for a single variant")
+    require(target.sequence_ref.str.len().eq(161).all() and target.sequence_alt.str.len().eq(161).all(),
+            "Expected 161-nt reference and alternate constructs")
+
+    values = singles.loc[singles.event_id.isin(target.event_id)].copy()
+    pooled = [f"{c}_delta_logit_pooled" for c in CELLS]
+    values["target_logit_delta"] = values[pooled].mean(axis=1, skipna=True)
+    values["target_contexts"] = values[pooled].notna().sum(axis=1)
+    values["intron1_length"] = values.intron1.str.len()
+    values["exon_length"] = values.exon.str.len()
+    values = values.rename(columns={"full_seq": "sequence_alt", "snp": "raw_mutation"})
+    target = target.merge(values[[
+        "event_id", "sequence_alt", "raw_mutation", "intron1_length", "exon_length",
+        "target_logit_delta", "target_contexts",
+    ]], on=["event_id", "sequence_alt"], how="left", validate="one_to_one")
+    require(target.target_logit_delta.notna().all(), "Missing target measurement")
+    require(target.mutation.eq(target.raw_mutation).all(), "Raw mutation label mismatch")
+    target = target.drop(columns="raw_mutation").sort_values(["event_id", "mutation"])
+    target["record_id"] = target.event_id + "|" + target.mutation
+    all_target_genes = set(target.gene_id)
+    target = target.loc[~target.gene_id.isin(exclusions["shared_genes_removed"])].copy()
+    for row in target.itertuples(index=False):
+        position, change = row.mutation.split(":")
+        reference_base, alternate_base = change.split(">")
+        position = int(position)
+        require(row.sequence_ref[position] == reference_base and
+                row.sequence_alt == row.sequence_ref[:position] + alternate_base + row.sequence_ref[position + 1:],
+                f"Mutation does not match reference/alternate DNA: {row.record_id}")
+
+    # Repeat effects are for the experimental reference, not the pooled label.
+    replicate_rows = []
+    for row in target.itertuples(index=False):
+        alternate = single_index.loc[(row.event_id, row.mutation)]
+        reference = wt.loc[row.event_id]
+        deltas = [[], []]
+        for cell in CELLS:
+            measurements = [(alternate[f"{cell}_rep{r}_logit"], reference[f"{cell}_rep{r}_logit"])
+                            for r in (1, 2)]
+            if all(np.isfinite(a) and np.isfinite(w) for a, w in measurements):
+                for k, (a, w) in enumerate(measurements):
+                    deltas[k].append(float(a - w))
+        replicate_rows.append((
+            float(np.mean(deltas[0])) if deltas[0] else np.nan,
+            float(np.mean(deltas[1])) if deltas[1] else np.nan, len(deltas[0]),
+        ))
+    target[["rep1_logit_delta", "rep2_logit_delta", "replicate_contexts"]] = replicate_rows
+    target["replicate_contexts"] = target.replicate_contexts.astype(int)
+    target = target.reset_index(drop=True)
+    require((len(target), target.gene_id.nunique(), target.event_id.nunique()) == (4618, 88, 97),
+            "COMPASS cohort differs from the published 4,618 variants / 88 genes / 97 exons")
+    require(target.record_id.is_unique, "Duplicate target ID")
+    return target, all_target_genes
+
+
+def prepare_opensplice(path, exclusions, target_genes):
+    # A historical canonical Parquet may also be used to avoid rereading the raw TSV.
+    if path.suffix == ".parquet":
+        source = pd.read_parquet(path)
+    else:
+        aliases = {
+            "sequence_alt": ("nt_seq", "mut_seq", "sequence", "seq"),
+            "sequence_ref": ("wt_seq", "reference_sequence", "ref_seq"),
+            "unit_id": ("ensembl_exon_id", "exon_id", "exon"),
+            "gene_id": ("gene", "gene_name", "gene_id"),
+            "delta_y": ("delta_psi", "dpsi", "deltaPSI"),
+            "y_alt": ("psi", "PSI"), "y_ref": ("wt_psi", "psi_wt", "WT_PSI"),
+        }
+        columns = set(pd.read_csv(path, sep="\t", nrows=0).columns)
+        resolved = {key: next((c for c in choices if c in columns), None)
+                    for key, choices in aliases.items()}
+        require(all(resolved[k] for k in aliases if k not in {"sequence_ref", "gene_id"}),
+                "Unrecognized OpenSplice columns")
+        extras = columns & {"variant_id", "start", "end", "wt", "mut_type"}
+        source = pd.read_csv(path, sep="\t", low_memory=False,
+                             usecols=sorted(extras | {c for c in resolved.values() if c}))
+        source = source.dropna(subset=[resolved["sequence_alt"], resolved["unit_id"]]).copy()
+        source["sequence_alt"] = source[resolved["sequence_alt"]].map(dna)
+        if resolved["sequence_ref"]:
+            source["sequence_ref"] = source[resolved["sequence_ref"]].map(dna)
+        else:
+            source["sequence_ref"] = [
+                seq[:int(start)-1] + dna(wt) + seq[int(end) if kind == "sub" else int(start)-1:]
+                for seq, start, end, wt, kind in zip(source.sequence_alt, source.start,
+                                                     source.end, source.wt, source.mut_type)
+            ]
+        source["unit_id"] = source[resolved["unit_id"]].astype(str)
+        source["gene_id"] = source[resolved["gene_id"]].astype(str) if resolved["gene_id"] else source.unit_id
+        source["record_id"] = [hashlib.sha256(f"opensplice|{unit}|{index}".encode()).hexdigest()[:24]
+                               for index, unit in zip(source.index, source.unit_id)]
+        for name in ("y_ref", "y_alt", "delta_y"):
+            source[name] = pd.to_numeric(source[resolved[name]], errors="coerce")
+        source["intervention_type"] = source.mut_type.astype(str)
+        source = source.dropna(subset=["delta_y"])
+    source = source.loc[source.intervention_type.eq("sub")].copy()
+    require((len(source), source.gene_id.nunique()) == (306343, 466), "Unexpected pre-exclusion source cohort")
+    shared = sorted(set(source.gene_id) & target_genes)
+    require(shared == exclusions["shared_genes_removed"], "Shared-gene exclusion differs")
+    source = source.loc[~source.gene_id.isin(shared)].reset_index(drop=True)
+    reference = np.clip(source.y_ref.to_numpy(float) / 100, .01, .99)
+    alternate = np.clip(source.y_alt.to_numpy(float) / 100, .01, .99)
+    source["target_logit_delta"] = np.log(alternate / (1 - alternate)) - np.log(reference / (1 - reference))
+    require((len(source), source.gene_id.nunique()) == (282610, 443), "Unexpected strict source cohort")
+    require(source.sequence_ref.str.len().eq(source.sequence_alt.str.len()).all(), "Source contains a length-changing edit")
+    require(source.record_id.is_unique, "Duplicate source ID")
+    columns = ["record_id", "gene_id", "unit_id", "sequence_ref", "sequence_alt", "target_logit_delta"]
+    if "variant_id" in source:
+        columns.insert(1, "variant_id")
+    return source[columns]
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--compass", type=Path, required=True, help="GSE307247 processed CSV.gz")
+    parser.add_argument("--opensplice", type=Path, help="OpenSplice v5 TSV.gz or historical canonical Parquet")
+    parser.add_argument("--out", type=Path, required=True, help="New output directory")
+    args = parser.parse_args()
+    exclusions = json.loads((HERE / "exclusions.json").read_text())
+    target, target_genes = prepare_compass(args.compass, exclusions)
+    # Compare to the published data before writing anything.
+    expected = pd.read_csv(HERE / "compass.tsv.gz", sep="\t")
+    pd.testing.assert_frame_equal(target[expected.columns], expected, check_dtype=False,
+                                  check_exact=False, rtol=1e-12, atol=1e-12)
+    source = prepare_opensplice(args.opensplice, exclusions, target_genes) if args.opensplice else None
+    if source is not None:
+        manifest = pd.read_csv(HERE / "opensplice-source.tsv.gz", sep="\t")
+        pd.testing.assert_frame_equal(source[["record_id", "gene_id"]], manifest, check_dtype=False)
+    args.out.mkdir(parents=True, exist_ok=False)
+    target.to_parquet(args.out / "compass.parquet", index=False)
+    if source is not None:
+        source.to_parquet(args.out / "opensplice-source.parquet", index=False)
+    print(f"Verified COMPASS: {len(target):,} variants, {target.gene_id.nunique()} genes")
+    if source is not None:
+        print(f"Verified OpenSplice: {len(source):,} substitutions, {source.gene_id.nunique()} genes")
+
+
+if __name__ == "__main__":
+    main()

compass-transfer/sequence-overlaps.csv

@@ -0,0 +1,13 @@
+dataset_a,dataset_b,sequence_id_a,sequence_id_b,role_a,role_b,gene_id_a,gene_id_b,length_a,length_b,shared_seeds,edit_distance,identity,orientation,overlap_class,target_start,target_end,query_is_a
+compass,opensplice,chr2:232527305-232527465:+,ENSE00003555045,reference_exon,reference_exon,CHRND,CHRND,161,140,125,5,0.9642857142857143,forward,near_complete_shorter,26,160,False
+compass,opensplice,chr3:38557211-38557371:-,ENSE00004014126,reference_exon,reference_exon,SCN5A,SCN5A,161,149,134,5,0.9664429530201343,forward,near_complete_shorter,17,160,False
+compass,opensplice,chr11:119092023-119092183:+,ENSE00003528094,reference_exon,reference_exon,HMBS,HMBS,161,134,119,5,0.9626865671641791,forward,near_complete_shorter,32,160,False
+compass,opensplice,chrX:31177912-31178072:-,ENSE00001158912,reference_exon,reference_exon,DMD,DMD,161,134,119,5,0.9626865671641791,forward,near_complete_shorter,32,160,False
+compass,opensplice,chr2:71549251-71549411:+,ENSE00001072853,reference_exon,reference_exon,DYSF,DYSF,161,137,122,5,0.9635036496350365,forward,near_complete_shorter,29,160,False
+compass,opensplice,chr20:63431320-63431480:-,ENSE00003997833,reference_exon,reference_exon,KCNQ2,KCNQ2,161,125,110,5,0.96,forward,near_complete_shorter,41,160,False
+compass,opensplice,chr1:45332898-45333058:-,ENSE00003522055,reference_exon,reference_exon,MUTYH,MUTYH,161,137,122,5,0.9635036496350365,forward,near_complete_shorter,29,160,False
+compass,opensplice,chr19:38570553-38570713:+,ENSE00000826243,reference_exon,reference_exon,RYR1,RYR1,161,182,151,0,1.0,forward,exact_containment,16,176,True
+compass,opensplice,chr11:119088168-119088328:+,ENSE00003527962,reference_exon,reference_exon,HMBS,HMBS,161,149,134,5,0.9664429530201343,forward,near_complete_shorter,17,160,False
+compass,opensplice,chr11:31801851-31802011:-,ENSE00003499781,reference_exon,reference_exon,PAX6,PAX6,161,137,122,5,0.9635036496350365,forward,near_complete_shorter,29,160,False
+compass,opensplice,chr4:113348168-113348328:+,ENSE00001644280,reference_exon,reference_exon,ANK2,ANK2,161,128,113,5,0.9609375,forward,near_complete_shorter,38,160,False
+compass,routing,chr3:128806309-128806469:+,chr3:128805722-128806302,reference_exon,enhancer_window,RAB7A,,161,1000,151,0,1.0,forward,exact_containment,796,956,True

compass-transfer/sources.json

@@ -0,0 +1,20 @@
+{
+  "sources": [
+    {
+      "dataset": "compass",
+      "url": "https://ftp.ncbi.nlm.nih.gov/geo/series/GSE307nnn/GSE307247/suppl/GSE307247_Processed_PSIs_All_Cells.csv.gz",
+      "doi": "10.1101/2025.10.12.681955",
+      "bytes": 36222316,
+      "sha256": "c5ec382ce6a54097fe15eda70a5738f891dbe831fab5242cf5ff3671f3c2cd07",
+      "md5": "fa6a4973260cb9534a72b096dd470f10"
+    },
+    {
+      "dataset": "opensplice",
+      "url": "https://ndownloader.figshare.com/files/64886694",
+      "doi": "10.6084/m9.figshare.32337414.v5",
+      "bytes": 425453242,
+      "sha256": "1ec5aa239793bf6e84d5d771990d04486f03017b8d8cfb67d323076575d05bf4",
+      "md5": "374443aa36bc1070dfd36318d8e77c59"
+    }
+  ]
+}

compass-transfer/target-genes.tsv

@@ -0,0 +1,89 @@
+gene_id	variants	exons
+ABCB7	65	2
+ACAD9	29	1
+ACADVL	39	1
+ADA	50	1
+ADAM15	78	1
+AFF2	61	1
+AKAP9	18	1
+ALDH6A1	59	1
+ALDOA	60	1
+AMPD1	22	1
+ANK1	73	1
+ANK3	38	1
+AP4S1	56	1
+APPL1	40	1
+ART4	46	1
+ASAH1	22	1
+ASS1	53	1
+AUH	51	1
+BCKDHA	49	1
+BEST1	52	1
+BLOC1S6	48	1
+CACNA2D4	61	1
+CASK	38	1
+CLN3	68	1
+CRYM	23	1
+DEPDC5	116	2
+DIAPH1	17	1
+DNM1L	126	3
+DRD2	51	1
+ECEL1	53	1
+EPB41L1	110	2
+ERCC2	40	1
+EYA1	19	1
+FANCA	39	1
+FGF8	77	1
+GALC	45	1
+GCDH	49	1
+GDAP1	47	1
+GFM2	12	1
+GH1	44	1
+GNS	46	1
+GPI	56	1
+GUSB	53	1
+HARS2	67	1
+IL2RA	31	1
+IMPDH1	107	2
+IVD	59	1
+LARS1	48	1
+LMNA	61	1
+MAPT	61	1
+MEFV	60	1
+MFAP5	53	1
+MLC1	122	2
+MYBPC1	75	1
+NDUFAF5	59	1
+NEXN	26	1
+NLRC3	35	1
+NPC2	46	1
+OCA2	56	1
+PAFAH1B1	15	1
+PAX5	79	1
+PCCA	45	1
+PCCB	44	2
+PIGH	63	1
+PPT1	63	1
+SERPINB7	47	1
+SHOX2	75	1
+SLC13A5	29	1
+SLC26A5	63	1
+SNCA	50	1
+SOS1	49	1
+SPINT2	50	1
+STRC	15	1
+STXBP2	65	1
+SUMF1	102	2
+SYNGAP1	60	1
+TAC3	57	1
+TIMM44	52	1
+TNFRSF1A	50	1
+TTC8	36	1
+TTN	26	1
+TTN-AS1	54	1
+UROS	51	1
+USB1	54	1
+USP28	47	1
+VPS33B	56	1
+WDR35	49	1
+WT1	7	1