[NEW SUBMISSION] Confidence-Calibrated Delta scoring function

config.json

@@ -491,6 +491,13 @@
             "key": "mutant",
             "model_type": "Single sequence & Structure"
         },
+        "ProSST-2048-Single-Pass-Delta": {
+            "input_score_name": "ProSST_2048_single_pass_delta_score",
+            "location": "ProSST/ProSST-2048-Single-Pass-Delta",
+            "directionality": 1,
+            "key": "mutant",
+            "model_type": "Single sequence & Structure"
+        },
         "ProSST-4096": {
             "input_score_name": "ProSST-4096",
             "location": "ProSST/ProSST-4096",
@@ -1302,4 +1309,4 @@
             "model_type": "Sequence Embedding"
         }
     }
-}
+}

proteingym/baselines/prosst_single_pass_delta/README.md

@@ -0,0 +1,27 @@
+# ProSST Single-Pass Delta
+
+This baseline uses `AI4Protein/ProSST-2048` with a single wild-type forward pass per assay and an additive mutation readout.
+
+Scoring details:
+- one forward pass on the wild-type sequence plus quantized structure tokens
+- mutation scores are summed across substitutions
+- the per-position readout uses a delta that combines:
+  - mutant log-probability
+  - a background correction
+  - an expected log-probability penalty
+  - a wild-type confidence bonus
+
+Required benchmark layout:
+- `residue_sequence/*.fasta`
+- `structure_sequence/2048/*.fasta`
+- `substitutions/*.csv`
+
+Example usage:
+
+```bash
+python ../../proteingym/baselines/prosst_single_pass_delta/compute_fitness.py \
+  --model_name AI4Protein/ProSST-2048 \
+  --base_dir /path/to/proteingym_benchmark \
+  --reference_file_path ../../reference_files/DMS_substitutions.csv \
+  --output_scores_folder /path/to/output_scores/ProSST/ProSST-2048-Single-Pass-Delta
+```