Authors

• Yishak Bililign - yishakbililign

• Luis Hernandez - lhmag89

• Sheyda Nazarian - SheydaNazarian

• Natalia Rivera - nataliariverax

Overview

This project is focused on analyzing and refining filtering strategies for de novo protein binders using model confidence metrics, structural interface descriptors, and sequence-based biophysical characterization with the goal of improving target and design class generalization. We explore filtering for both protein expression & target binding using open-source de novo design datasets from Proteinbase. We propose independent LDA scoring approaches for both filtering tasks, which together create a 6-fold improvement in binder success rate on a new target dataset unincluded in model development:

Repository Map

root/
├── boltz/
├── data/
├── figures/
├── notebooks/
├── scripts/
├── utils/
├── data/
├── environment.yml
└── README.md

Data

Datasets & protein structures used for statistical analysis & modeling.

root/
└── data/
    ├── processed/
    |   ├── full/
    |       ├── full_binder_confidence_metrics.csv
    |       ├── full_binder_dataset.csv
    |       ├── full_binder_lda_scores.csv
    |       └── full_binder_structural_features.csv
    |   ├── nipah/
    |       ├── nipah_binder_dataset.csv
    |       └── nipah_binder_structural_features.csv
    |   └── rbx1/
    |       ├── rbx1_test_dataset.csv
    |       └── rbx1_test_predictions.csv
    ├── structures/
    |   ├── nipah/
    |       ├── <id>_boltz2.cif
    |       └── <id>_esmfold.cif
    |   └── full/
    |       └── <id>_boltz2.cif
    ├── proteinbase_collection_full_28_01_2026.csv
    ├── proteinbase_collection_nipah-binder-competition-results.csv
    └── target_sequences.csv

Notebooks

Notebooks used for EDA, statistical analysis, & modeling.

root/
└── notebooks/
    ├── binding/
    |   ├── ANOVA.ipynb
    |   ├── GAM.ipynb
    |   ├── Thresholding.ipynb
    |   ├── Full_Binder_LDA_Analysis.ipynb
    |   └── Nipah_Binder_LDA_Analysis.ipynb            
    └── expression/
        ├── Full_Expression_LDA_Analysis.ipynb
        └── Nipah_Expression_LDA_Analysis.ipynb

Scripts

Python scripts with CLI for data processing & feature extraction.

root/
└── scripts/
    ├── analyze_structure.py
    ├── parse_proteinbase.py
    └── read_boltz.py

Boltz

Pipeline for generating structure predictions using Boltz2. See boltz/README.md for more details.

Configuration

Create a conda environment with all required package dependencies for Python scripts in this repository with the following command:

conda env create -f environment.yml -n <env_name>

The Boltz2 structure prediction pipeline uses a Docker container which can be built from boltz/Dockerfile with the following command:

cd boltz
docker build -t boltz-image .

Usage

Parsing Proteinbase Datasets

usage: parse_proteinbase.py [-h] --input FILE --output FILE [--target NAME] [--download DIR]

Flag	Type	Required	Default	Description
-i, --input	str	✅	—	Input csv path
-o, --output	str	✅	—	Output csv path
-t, --target	str	❌	None	Binder target
-d, --download	str	❌	—	Directory to download structures

scripts/parse_proteinbase.py can be used to parse the Nipah competition dataset from Proteinbase saved in data/proteinbase_collection_nipah-binder-competition-results.csv to produce a new csv in data/processed/nipah/nipah_binder_dataset.csv with the following command:

python scripts/parse_proteinbase.py -i data/proteinbase_collection_nipah-binder-competition-results.csv -o data/processed/nipah/nipah_binder_dataset.csv -t nipah-glycoprotein-g

The full Proteinbase collection saved in data/proteinbase_collection_full_28_01_2026.csv can similarly be processed to produce a new csv in data/processed/full/full_binder_dataset.csv with the following command:

python scripts/parse_proteinbase.py -i data/proteinbase_collection_full_28_01_2026.csv -o data/processed/full/full_binder_dataset.csv

Extracting Structural Features

usage: analyze_structures.py [-h] --input DIR --output FILE [--source NAME] [--target TARGET_CHAIN_ID] [--binder BINDER_CHAIN_ID]

Flag	Type	Required	Default	Description
-i, --input	str	✅	—	Input directory of CIF files
-o, --output	str	✅	—	Output csv path
-s, --source	str	❌	boltz2	Structure prediction source
-t, --target	str	❌	A	Chain id of target
-b, --binder	str	❌	B	Chain id of binder

scripts/analyze_structure.py can be used to extract features from bound structures in CIF files. Files should be named in the format <id>_<source>.cif where id is a unique identifier for the binder submission. This can be used to extract features from the Nipah competition Boltz structure predictions in data/structures/nipah/ with the following command:

python scripts/analyze_structures.py -i data/structures/nipah/ -o data/processed/nipah/nipah_structural_features.csv

Compiling Boltz2 Prediction Results

usage: read_boltz.py [-h] --input DIR --output FILE [--copy DIR]

Flag	Type	Required	Default	Description
-i, --input	str	✅	—	Directory containing Boltz prediction output
-o, --output	str	✅	—	Output path for compiled CSV
-c, --copy	str	❌	—	Directory to copy output CIF files

scripts/read_boltz.py can be used to compile & calculate confidence metrics from Boltz2 predictions. This uses code from the Adaptyv Bio Nipah competition for calculating ipSAE & LIS (see: utils/ipsae.py). Example results can be compiled to data/processed/full/full_binder_confidence_metrics.csv with the structures copied to data/structures/full/ using the following command:

python scripts/read_boltz.py -i boltz/results -o data/processed/full/full_binder_confidence_metrics.csv -c data/structures/full/

Full results for the Boltz2 predictions used in this project can be accessed here.