imspy_predictors RT: default to Chronologer

packages/imspy-predictors/src/imspy_predictors/pretrained/hub.py

@@ -29,7 +29,9 @@
 )
 
 # Maps the *package-relative* path (as used by callers of get_model_path)
-# to its download URL and expected SHA-256 hash.
+# to its download URL and expected SHA-256 hash. Entries default to the
+# rustims models GitHub Release; an explicit "url" overrides that for
+# vendor-free upstream redistributions (e.g. Chronologer).
 MODELS = {
     "ccs/best_model.pt": {
         "filename": "ccs-best_model.pt",
@@ -51,6 +53,15 @@
         "filename": "pretrained_encoder.pt",
         "sha256": "43ccc2f836bf3d81943ddce353ade9628e7d036421ba5b5c182bf163e496385e",
     },
+    # Chronologer base weights — fetched directly from upstream Searle Lab
+    # (Apache-2.0, attribution preserved in imspy_predictors.rt.chronologer).
+    # We deliberately do not re-host: pulls from the upstream repo so any
+    # future fix or retraining propagates automatically.
+    "rt/chronologer_base.pt": {
+        "filename": "chronologer_base.pt",
+        "url": "https://github.com/searlelab/chronologer/raw/main/models/Chronologer_20220601193755.pt",
+        "sha256": "1a500c246b49a1a23643bce7f2df86d5a107359bf0ec34365531c73431b6c0b3",
+    },
 }
 
 
@@ -153,7 +164,7 @@ def ensure_model(model_name: str) -> Path:
 
     # 3. Download into a temp file in the same filesystem, then atomic-rename.
     cached_path.parent.mkdir(parents=True, exist_ok=True)
-    url = f"{_RELEASE_BASE}/{meta['filename']}"
+    url = meta.get("url") or f"{_RELEASE_BASE}/{meta['filename']}"
     logger.info("Downloading model '%s' from %s ...", model_name, url)
 
     tmp_fd, tmp_path = tempfile.mkstemp(

packages/imspy-predictors/src/imspy_predictors/rt/predictors.py

@@ -228,18 +228,45 @@ def load_deep_retention_time_predictor(backend: Optional[str] = None):
     """
     Load a pretrained retention time predictor model.
 
+    Defaults to **Chronologer** (Searle Lab, Apache-2.0) — a residual-CNN that
+    reaches roughly 4× tighter median residual than the imspy transformer
+    baseline on timsTOF data. Pass ``backend="transformer"`` to opt back into
+    the legacy imspy ``UnifiedPeptideModel`` RT predictor.
+
     Args:
-        backend: Kept for backward compatibility, ignored (always uses PyTorch)
+        backend: ``"chronologer"`` (default, or pass ``None``) loads the
+            Chronologer wrapper. ``"transformer"`` loads the legacy imspy
+            transformer RT model. Any other value is treated as the default.
 
     Returns:
-        Loaded PyTorch model
+        Either a :class:`Chronologer` wrapper or the legacy transformer
+        PyTorch model, depending on ``backend``.
     """
     if not TORCH_AVAILABLE:
         raise ImportError(
             "PyTorch is required for retention time prediction. "
             "Install with: pip install torch"
         )
 
+    backend = (backend or "chronologer").lower()
+
+    if backend == "chronologer":
+        # Default: Chronologer is the goto RT predictor on timsTOF data.
+        # Falls back to the transformer path only if Chronologer can't be
+        # constructed (e.g. upstream `chronologer` package not installed, or
+        # the base-weights download is unreachable).
+        try:
+            from imspy_predictors.rt.chronologer import Chronologer
+            base_path = get_model_path('rt/chronologer_base.pt')
+            return Chronologer.from_base(str(base_path))
+        except (ImportError, FileNotFoundError, RuntimeError) as e:
+            import logging
+            logging.getLogger(__name__).warning(
+                "Chronologer load failed (%s); falling back to transformer.", e,
+            )
+            backend = "transformer"
+
+    # Transformer path (legacy / fallback).
     # Try to load UnifiedPeptideModel first (new architecture)
     try:
         from imspy_predictors.models import UnifiedPeptideModel