IMPRINT (immunological fingerprinting): Decoding TCR Recognition via Geometric Deep Learning of Immunological Fingerprints

Overview

IMPRINT is a surface-based geometric deep learning framework that learns immunological fingerprints—multimodal geometric and physicochemical features from pMHC interfaces—to model and predict T cell receptor (TCR) binding preferences.

---

Requirements

IMPRINT builds upon the MaSIF framework and adapts it to immunological contexts. Please refer to the Software prerequisites section of MaSIF for detailed library and environment setup requirements.

---

Datasets

HLA-A02 Dataset

Trained and validated on a curated HLA-A*02 dataset containing 40 TCR–pMHC complexes from the Protein Data Bank (PDB), spanning 7 TCR categories.

## HLA-A02 dataset
./source/data_preparation_pmhc/00-raw_pdbs/HLA-A02_40-7/

HLA-B27 Dataset

Used for zero-shot inference to assess cross-allele generalization.

## HLA-B27 dataset
./source/data_preparation_pmhc/00-raw_pdbs/HLA-B27/

User-defined Dataset

Custom pMHC structure datasets can be provided as .pdb files, either directly or extracted from TCR–pMHC complex structures.

---

Immunological Fingerprints

pMHC Surface Featurization

Four main steps are involved:

1. The pMHC surface is triangulated into a discrete mesh.
2. Around each mesh vertex, a radial patch is extracted with geodesic radius r = 12 Å.
3. Within each patch, two geometric and three physicochemical features are computed and interpolated onto surface points.
4. Geodesic polar coordinates are mapped to represent the relative spatial positions of features within the patch.

# cd ./source
bash ./data_preparation_pmhc/pmhc_batch_prepare.sh

Definition of Interface Regions

Interface patches are defined by a distance cutoff (typically 4 Å) from any peptide atom, representing the peptide-centric recognition interface.

---

Deep Discriminator

Cross-validation on HLA-A02 Dataset

Training and evaluation using All-test cross-validation:

# cd ./source
bash ./deep_discriminator/workflow_40-7_pepcut4_all-test.sh

Explainability: Patch-level Scoring Analysis

# cd ./source
bash ./deep_discriminator/workflow_trace_40-7_pepcut4_all-test.sh

## pMHC Interface scoring
bash ./trace_pmhc/batch_trace-all-test_40-7_pepcut4.sh

## Peptide scoring
bash ./trace_peptide/batch_tracepep-all-test_40-7_pepcut4.sh

---

Inference

Zero-shot Inference on HLA-B27 Dataset

## Retraining and inference
bash ./deep_discriminator/workflow_40-7_pepcut4_b27-rigid.sh

## Explainability
bash ./trace_pmhc/batch_trace-b27-rigid_40-7_pepcut4.sh
bash ./trace_peptide/batch_tracepep-b27-rigid_40-7_pepcut4.sh

Inference on User-defined Dataset

## Model weights: 7-D discriminator trained on 40 HLA-A02 structures
./model/

---

Contact

If you have any questions or requests, please contact us at: shangchun@zju.edu.cn:)