DELPHI: Pre-trained antibody developability models (PSR and SEC)

This repository accompanies the DELPHI (Deep End-to-end Learning Platform 
for antibody developability with High Interpretability) platform from the 
Institute for Protein Innovation (IPI).

DELPHI platform: https://github.com/proteininnovation/delphi

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
PRETRAINED MODELS (trained on IPI internal datasets)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Models trained on proprietary IPI antibody developability datasets.
Training sequences cannot be shared due to confidentiality.

  PSR filter models  (polyreactivity):
    FINAL_psr_filter_*.pt / .pkl

  SEC filter models  (size exclusion chromatography):
    FINAL_sec_filter_*.pt / .pkl

  Supported architectures: TransformerOneHot, TransformerLM 
  (ABlang2, AntiBERTy, AntiBERTa2, IgBERT), Random Forest, XGBoost

  Usage:
    python delphi.py --list-models
    python delphi.py --predict my_antibodies.xlsx \
        --model_id FINAL_psr_filter_onehot_transformer_onehot_ipi_psr.pt

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
DS1 — PRE-PROCESSED PUBLIC DATASET
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Pre-processed version of the human antibody polyreactivity dataset
from Chen et al. 2024 (Cell Reports, PMC11564698, MIT License).

Processing applied to the original dataset:
  - HCDR3 extracted via ANARCI (IMGT numbering, positions 105-117)
  - Columns standardised to DELPHI format:
      BARCODE     — unique antibody identifier
      HSEQ        — VH amino acid sequence
      LSEQ        — VL amino acid sequence
      CDR3        — HCDR3 sequence (ANARCI IMGT 105-117)
      psr_filter  — 1=PASS (low polyreactivity), 0=FAIL (high)

  DS1.xlsx                         — full dataset (246,293 antibodies)
  DS1.xlsx.ablang.emb.csv          — ABlang2 embeddings (480-dim)
  DS1.xlsx.antiberta2.emb.csv      — AntiBERTa2 embeddings (1024-dim)
  DS1.xlsx.antiberta2-cssp.emb.csv — AntiBERTa2-CSSP embeddings (1024-dim)
  DS1.xlsx.antiberty.emb.csv       — AntiBERTy embeddings (512-dim)
  DS1.xlsx.igbert.emb.csv          — IgBERT embeddings (1024-dim)

Source: Chen HT, Zhang Y, Huang J et al. Human antibody polyreactivity is governed primarily by the heavy-chain complementarity-determining regions.Cell Reports, 2024.
10.1016/j.celrep.2024.114801
Original data: https://zenodo.org/records/14735846 (MIT License)