Add Corrected SaProt VH+VL baseline

models/saprot_vh_vl/README.md

@@ -0,0 +1,97 @@
+# Saprot_VH_VL Baseline
+
+Ridge regression on embeddings from the **SaProt** protein language model on VH (variable heavy) and VL (variable light) sequences with two-chain encoding.
+
+## Description
+
+SaProt (Structure-aware Protein Language Model) generates predictions for protein properties using sequence and structure information. 
+
+This baseline uses locally computed SaProt embeddings(SaProt_35M_AF2) to generate fixed-length embeddings for VH and VL chains using their sequences and structures, then concatenates these embeddings(VH + VL) and trains simple Ridge regression models on top to predict antibody developability properties.
+
+The rationale behind this joint representation is same as that of the ESM2 case(no token contamination and learning from features independently).
+
+Note: At the time of writing, there were two choices to fetch structures from - MOE and ABB3. This implementation concerns itself only with the MOE structures
+
+## Method
+
+### 1. Separate Chain Embedding
+
+For each antibody, we embed the heavy and light chains independently:
+
+**VH Embedding:**
+```
+Complexed .pdb files from MOE → Extract VH pdb  → FoldSeek 3di Descriptors → Interleaved with VH_seq → SaProt Tokenizer → Last Hidden State → Mean Pool → vh_embedding
+```
+
+**VL Embedding:**
+```
+Complexed .pdb files from MOE → Extract VL pdb   → FoldSeek 3di Descriptors → Interleaved with VL_seq → SaProt Tokenizer → Last Hidden State → Mean Pool → vl_embedding
+```
+### 2. Feature Concatenation
+
+After generating embeddings for both chains, we concatenate them:
+
+```
+combined_embedding = np.concatenate([vh_embed, vl_embed])
+```
+
+For SaProt_35M_AF2, the embedding dimension is 480, so:
+- VH embedding: 480D
+- VL embedding: 480D  
+- Combined: 960D
+
+
+## Requirements
+
+- The Complexed(VH+VL) PDB structures for training are in `../../data/structures/MOE_structures/GDPa1/` and in the format of `{antibody_name}.csv`
+- The Complexed(VH+VL) PDB structures for the heldout data is in `../../data/structures/MOE_structures/heldout_test/` and in the format of `{antibody_name}.csv`
+- The Heavy chains are labelled by 'B' and the light chains are labelled by 'A'
+- foldseek installed
+- BioPython installed
+
+Note: While SaProt embeddings can be calculated from the sequence and structure information, in the absence of structure information, it defaults to calculating embeddings with sequence information only.
+
+Also Note: Current implementation works around abdev-core via hard-coding the size of heldout data. This is not good practice, and is only a temporary fix
+
+### Train
+
+From the repository root:
+
+```bash
+cd model/saprot_vh_vl
+pixi install
+
+# Train on GDPa1 dataset
+pixi run python -m saprot_vh_vl train \
+  --data ../../data/GDPa1_v1.2_20250814.csv \
+  --run-dir ./runs/my_run
+```
+
+### Predict
+
+```bash
+# Predict on training data
+pixi run python -m saprot_vh_vl predict \
+  --data ../../data/GDPa1_v1.2_20250814.csv \
+  --run-dir ./runs/my_run
+```
+
+### Full Workflow via Orchestrator
+
+From repository root:
+
+```bash
+pixi run all
+```
+
+This automatically discovers and runs all models, including SaProt_VH_VL, with 5-fold cross-validation.
+
+## Citation
+
+Saprot: Su J, et al. (2023). "SaProt: Protein Language Modeling with Structure-aware Vocabulary." bioRxiv.
+
+# Code References
+
+SaProt - https://github.com/westlake-repl/SaProt
+Foldseek - https://github.com/steineggerlab/foldseek
+

models/saprot_vh_vl/pixi.toml

@@ -0,0 +1,35 @@
+[workspace]
+name = "saprot_vh_vl"
+version = "0.1.0"
+description = "Saprot_VH_VL baseline - protein language model predictions on VH and VL sequences and Structures"
+channels = ["conda-forge", "bioconda", "pytorch"]
+platforms = ["linux-64", "osx-64", "osx-arm64"]
+
+[dependencies]
+python = "3.11.*"
+numpy = ">=1.24"
+pandas = ">=2.0"
+scikit-learn = ">=1.3"
+typer = ">=0.9"
+foldseek = "*"  
+biopython = ">=1.81"
+
+
+[pypi-dependencies]
+abdev-core = { path = "../../libs/abdev_core", editable = true }
+saprot_vh_vl = { path = ".", editable = true }
+transformers = ">=4.30"
+torch = ">=2.0"
+
+[environments]
+default = []
+dev = ["dev"]
+
+[feature.dev.dependencies]
+pytest = ">=7.0"
+ruff = ">=0.1"
+
+[feature.dev.tasks]
+lint = "ruff check src && ruff format --check src"
+test = "pytest tests -v"
+

models/saprot_vh_vl/pyproject.toml

@@ -0,0 +1,20 @@
+[build-system]
+requires = ["setuptools>=64", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "saprot_vh_vl"
+version = "0.1.0"
+description = "Saprot_VH_VL baseline - protein language model predictions on VH and VL sequences and structures"
+requires-python = ">=3.11"
+dependencies = [
+    "abdev-core",
+    "pandas>=2.0",
+]
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.setuptools.package-dir]
+"" = "src"
+

models/saprot_vh_vl/src/saprot_vh_vl/__init__.py

@@ -0,0 +1,4 @@
+"""Saprot_VH_VL baseline - protein language model predictions."""
+
+__version__ = "0.1.0"
+

models/saprot_vh_vl/src/saprot_vh_vl/__main__.py

@@ -0,0 +1,7 @@
+"""Entry point for model CLI."""
+
+from .run import app
+
+if __name__ == "__main__":
+    app()
+