Degformer

A transformer mode for prediction of protein stability index for 28-residue peptides. Performs better than current composition or motif-search based methods and is applicable for protein degron mapping (internal + C-degrons only), in silico scanning/saturation mutagenesis for degron motif detection, and generally stability prediction.

Training:

• Uses S1, S4, S7 data from the Elledge lab (doi.org/10.1016/j.molcel.2023.08.022)

  • S1: 260K peptides from the human proteome

  • S4: Scanning mutagenesis on a subset of degrons

  • S7: Saturation mutagenesis on a subset of degrons

  • For S7, only degrons with >=60 reads were kept

• 90% of data from each table were selected at random and merged into data.csv (569270 sequences)

• The other 10% was kept for final blind testing

• Checkpoints for epoch 30 and 50 saved

Performance:

Degformer was tested on data from S1, S4, and S7 that was excluded from training/evaluation. Predictions from the Elledge lab paper were only available for S1.

	Dataset	Correlation (ρ)	Mean ΔPSI	Median ΔPSI	σ(ΔPSI)
Elledge lab*	S1 S4 S7	0.8955 - -	0.4174 - -	0.3179 - -	0.5584 - -
Degformer (e30)	S1 S4 S7	0.9409 0.9441 0.9089	0.3178 0.3065 0.3131	0.2475 0.2476 0.2240	0.2873 0.2625 0.3010
Degformer (e50)	S1 S4 S7	0.9420 0.9463 0.9162	0.3148 0.2932 0.3014	0.2397 0.2322 0.2150	0.2905 0.2607 0.2923

* Statistics determined from entire 260K dataset for S1 only

Figure 1: Performance of Degformer e50 on S1 blind set.

Figure 2: in silico saturation mutagenesis recovers diverse degron motifs.

Figure 3: C-degrons are detected by Degformer. Addition of any residues to the ends of C-degrons stabilizes PDGFC (G-end), SNF8 (P-end), TRPC4AP (EE-end), GNMT (R-3) C-terminals (clockwise from top left).

Figure 4: Degron scanning of human GNMT.

Setting up (VSCode):

1. Download repository

2. In VSCode terminal, cd into project folder

3. Create virtual environment (python -m venv degron-env)

4. Activate venv (degron-env/scripts/activate). Check that python interpreter is using the virtual environment in bottom right corner.

5. Install dependencies (torch, pandas, numpy, argparse, seaborn, matplotlib, etc.)

   1. If using a GPU make sure the torch package is compatible with the architecture (Blackwell => cu128)

   2. For CPU, install CPU-only torch

Scripts:

train_v2.py:

• Main training script using built in transformer model training with Pytorch

• Model weights stored as .pt

predict.py:

• Main inference script using model weights from train_v2.py

• Provide input .csv file where first column is peptide name, and second column is the peptide sequence. Header should be (name, sequence).

• Set desired model checkpoint in python file. Recommended: peptide_model_epoch50.pt

Modes:

• Default: predicts deltaPSI and controlPSI for peptides

• Saturation mutagenesis: predicts PSIs for input peptides, and all possible point mutants

• Scanning mutagenesis: predicts PSIs for input peptides, with scanning point mutation using specified residue (alanine by default)

• Protein scanning: provided protein sequence instead of peptide, will predict PSI for all possible 28 residue fragments in order (overlapping adjacent fragments by 27 residues)

Usage:

• Default prediction

python predict.py --input predict_input.csv

• Saturation mutagenesis

python predict.py --input predict_input.csv --mode sat_mut

• Scanning mutagenesis (to Alanine)

python predict.py --input predict_input.csv --mode scan_mut

• Scanning mutagenesis (to Glycine)

python predict.py --input predict_input.csv --mode scan_mut --residue G

• Protein scanning

python predict.py --input predict_input.csv --mode protein

• Fragment index for protein mode is based on position of rightmost residue in protein (add 13 or 14 to the index for fragment center format)

saturation_mut_heatmap.py:

• Input .csv should be the output from predict.py or formatted in the same way
• Capable of generating heatmaps for multiple saturation mutagenesis results at a time

python saturation_mut_heatmap.py predict_output.csv