feat: add leech as optional GPU-accelerated CCA classifier

Add leech (rnabioco/leech) as a git submodule and integrate it as an
alternative to remora for charging classification. Controlled by
`classifier: "remora" | "leech"` in config (remora remains default).

- Add classify_charging_leech rule with GPU support and ruleorder routing
- Add GPU resource configs for leech in both LSF and SLURM profiles
- Add leech installation to setup-tools.sh and pixi install-leech task
- Add leech tool check in Snakefile onstart when classifier is leech
- Fix pod5 missing 'deprecated' dependency from --no-deps install

Author: jayhesselberth

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "resources/leech"]
+	path = resources/leech
+	url = https://github.com/rnabioco/leech

cluster/lsf/config.yaml

@@ -18,6 +18,10 @@ set-resources:
   - rebasecall:ngpu=1
   - rebasecall:mem_mb=24
   - classify_charging:mem_mb=24
+  - classify_charging_leech:lsf_queue="gpu"
+  - classify_charging_leech:lsf_extra="-gpu num=1:j_exclusive=yes:mode=exclusive_process"
+  - classify_charging_leech:ngpu=1
+  - classify_charging_leech:mem_mb=24
   - remora_signal_stats:mem_mb=24
   - inject_ubam_tags:mem_mb=16
   - bwa_align:mem_mb=24

cluster/slurm/config.yaml

@@ -40,6 +40,14 @@ set-resources:
     mem_mb: 24000              # 24GB
     cpus_per_task: 4
 
+  # Leech charging classification (GPU-accelerated)
+  classify_charging_leech:
+    slurm_partition: "gpu"     # GPU partition (customize for your cluster)
+    gres: "gpu:1"              # Request 1 GPU
+    runtime: 120               # 2 hours (faster than remora with GPU)
+    mem_mb: 24000              # 24GB
+    cpus_per_task: 4
+
   # Tag injection from unaligned BAM
   inject_ubam_tags:
     runtime: 360               # 6 hours

config/config-base.yml

@@ -60,6 +60,10 @@ remora_kmer_table: null
 # read classification model - remora trained model to classify charged vs uncharged reads
 remora_cca_classifier: "resources/models/cca_classifier.pt"
 
+# CCA charging classifier to use: "remora" (default) or "leech"
+# Leech is a GPU-accelerated alternative (requires resources/leech submodule)
+classifier: "remora"
+
 # software tools (modkit and remora managed by pixi)
 # NOTE: dorado_version and dorado_model are coupled. When upgrading dorado,
 # check the release notes for new models:

pixi.toml

@@ -46,7 +46,7 @@ bottleneck = "*"
 scripts = ["scripts/setup-env.sh"]
 
 [tasks]
-setup = { cmd = "bash scripts/setup-tools.sh", description = "One-time setup: install dorado, models, remora, and WarpDemuX" }
+setup = { cmd = "bash scripts/setup-tools.sh", description = "One-time setup: install dorado, models, remora, leech, and WarpDemuX" }
 dl-test-data = { cmd = "cd .tests && bash dl_test_data.sh", description = "Download test data" }
 # Pipeline execution
 dry-run = { cmd = "snakemake -n --configfile=config/config-test.yml", description = "Dry run with test config" }
@@ -62,6 +62,8 @@ collapse-ref = { cmd = "python workflow/scripts/collapse_gtrndb_fasta.py", descr
 lint-snakefmt = { cmd = "snakefmt --check workflow/", description = "Check Snakemake formatting" }
 lint-python = { cmd = "ruff check workflow/scripts/ && ruff format --check workflow/scripts/", description = "Lint and check Python formatting" }
 lint-yaml = { cmd = "yamllint -d \"{extends: default, rules: {line-length: {max: 120}, document-start: disable}}\" config/ .github/", description = "Lint YAML files" }
+# Leech tasks
+install-leech = { cmd = "git submodule update --init resources/leech && uv pip install -e resources/leech && CONDA_PYARROW=$(pixi list pyarrow 2>/dev/null | awk '/^pyarrow[[:space:]]/ {print $2}') && [ -n \"$CONDA_PYARROW\" ] && uv pip install --force-reinstall pyarrow==$CONDA_PYARROW || true", description = "Install leech CCA classifier from submodule" }
 # WarpDemuX tasks
 install-warpdemux = { cmd = "bash -c '[ -d resources/tools/WarpDemuX ] || git clone --recursive https://github.com/KleistLab/WarpDemuX.git resources/tools/WarpDemuX' && uv pip install -e resources/tools/WarpDemuX", description = "Clone and install WarpDemuX" }
 dry-run-demux = { cmd = "snakemake -n --configfile=config/config-demux-test.yml", description = "Dry run demux test" }

resources/leech

@@ -156,6 +156,8 @@ if python -c "from packaging.version import Version; exit(0 if Version('${curren
 else
     echo "Installing pod5 >= ${POD5_MIN_VERSION}..."
     uv pip install --no-deps "pod5>=${POD5_MIN_VERSION}"
+    # pod5 needs 'deprecated' but --no-deps skips it
+    uv pip install deprecated
     echo "Pod5 installed successfully"
 fi
 
@@ -173,6 +175,24 @@ else
     echo "WarpDemuX installed successfully"
 fi
 
+# ============================================================================
+# Leech Setup (via uv, from submodule)
+# ============================================================================
+echo "=== Checking leech ==="
+if python -c "import leech" 2>/dev/null; then
+    echo "Leech already installed"
+else
+    if [ -d "${REPO_ROOT}/resources/leech" ]; then
+        echo "Installing leech from submodule (with dependencies)..."
+        uv pip install -e "${REPO_ROOT}/resources/leech"
+        echo "Leech installed successfully"
+        echo "NOTE: pyarrow will be reconciled with conda in the next step"
+    else
+        echo "Leech submodule not found at resources/leech"
+        echo "Run 'git submodule update --init resources/leech' to clone it"
+    fi
+fi
+
 # ============================================================================
 # Reconcile pyarrow: ensure pip hasn't overridden conda's version
 # ============================================================================

scripts/setup-tools.sh

@@ -57,6 +57,9 @@ onstart:
     if config.get("warpdemux", {}).get("enabled", False):
         tools_to_check.append(("warpdemux", [conda_prefix]))
 
+    if config.get("classifier", "remora") == "leech":
+        tools_to_check.append(("leech", [conda_prefix, os.path.join(repo_root, "resources")]))
+
     for tool_name, allowed_prefixes in tools_to_check:
         tool_path = shutil.which(tool_name)
         if tool_path is None:
@@ -103,6 +106,16 @@ localrules:
     generate_squiggy_session,
 
 
+# Route charging classifier: prefer leech or remora based on config
+if config.get("classifier", "remora") == "leech":
+
+    ruleorder: classify_charging_leech > classify_charging
+
+else:
+
+    ruleorder: classify_charging > classify_charging_leech
+
+
 # Conditionally include WarpDemuX demultiplexing rules
 if config.get("warpdemux", {}).get("enabled", False):
 

workflow/Snakefile

@@ -214,6 +214,61 @@ rule classify_charging:
     """
 
 
+rule classify_charging_leech:
+    """
+  run leech trained model to classify charged and uncharged reads
+  GPU-accelerated alternative to remora (requires leech installed from resources/leech)
+
+  For EDX samples, uses the EDX-filtered POD5 to match the filtered BAM.
+  """
+    input:
+        pod5=get_classification_pod5,
+        bam=rules.inject_ubam_tags.output.bam,
+    output:
+        charging_bam=os.path.join(
+            outdir, "bam", "charging", "{sample}", "{sample}.charging.bam"
+        ),
+        charging_bam_bai=os.path.join(
+            outdir, "bam", "charging", "{sample}", "{sample}.charging.bam.bai"
+        ),
+        temp_sorted_bam=temp(
+            os.path.join(
+                outdir, "bam", "charging", "{sample}", "{sample}.charging.bam.tmp"
+            )
+        ),
+    log:
+        os.path.join(outdir, "logs", "classify_charging_leech", "{sample}"),
+    threads: 4
+    params:
+        model=config["remora_cca_classifier"],
+    shell:
+        """
+    if [[ "${{CUDA_VISIBLE_DEVICES:-}}" ]]; then
+      echo "CUDA_VISIBLE_DEVICES $CUDA_VISIBLE_DEVICES"
+      export CUDA_VISIBLE_DEVICES
+    fi
+
+    leech predict \
+      --model {params.model} \
+      --pod5 {input.pod5} \
+      --bam {input.bam} \
+      --output {output.charging_bam} \
+      --device cuda \
+      --motif CCAGGC \
+      --motif-offset 3 \
+      --reference-anchored \
+      --workers 4 \
+      --batch-size 512 \
+      2>&1 | tee {log}
+
+    # sort the result
+    samtools sort -@ {threads} {output.charging_bam} > {output.temp_sorted_bam}
+    cp {output.temp_sorted_bam} {output.charging_bam}
+
+    samtools index {output.charging_bam}
+    """
+
+
 rule transfer_bam_tags:
     """
   creates classified bam with MM and ML tags transferred to CM/CL