PLL Scoring for Protein Sequences (ESM2)

This tool calculates the Pseudo Log-Likelihood (PLL) and Pseudo-Perplexity (PPPL) of protein sequences using Meta AI's ESM2 model. These scores help quantify how "natural" a protein looks to a protein language model. Making it useful for evaluating designed sequences like antibodies or binders.

---

Usage

Score a FASTA file containing a single protein sequence (binder only):

python pll_score.py --fasta example_binder.fasta

Example Output

Reading sequences from example_binder.fasta
Scoring: Cradle_EGFR_241aa (length: 241)
PLL: -301.2522
Avg Log-Likelihood: -1.2500
PPPL: 3.4904

---

Installation

1. Create a Conda Environment (Recommended)

conda create -n pll_env python=3.8
conda activate pll_env

2. Install Dependencies

pip install fair-esm torch biopython

Requires Python 3.8+. A GPU is recommended (but CPU works for short sequences).

Files

• pll_score.py – script to score PLL and PPPL using a sequence from a FASTA file
• example_binder.fasta – example input binder sequence from Adaptyv Bio challenge

---

Citation

If you use this tool, please cite Meta AI's ESM2 models.

---