TriStageHLA-PRE

TriStageHLA-PRE

A minimal Hugging Face-compatible PyTorch model for peptide presentation using ESM with optional LoRA and cross-attention. There is no custom predict API; inference follows the training path: tokenize peptide and HLA pseudosequence with the ESM tokenizer, pad or truncate to fixed lengths (default peptide=16, HLA=36), run a forward pass as logits, features = model(epitope_ids, hla_ids), then apply softmax to get the binding probability.

Quick Start

Requirements:

• Python >= 3.8
• torch >= 2.0
• transformers >= 4.40
• peft (only if you use LoRA/PEFT adapters)

Install:

pip install torch transformers peft

Usage (Transformers)

import torch
import torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer
model_id = "SkywalkerLu/TriStageHLA-PRE"
model = AutoModel.from_pretrained(model_id, trust_remote_code=True).to(device).eval()

## How to use TriStageHLA-PRE
```python
import torch
import torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer

# Device
device = "cuda" if torch.cuda.is_available() else "cpu"

model_id = "SkywalkerLu/TriStageHLA-PRE"
model = AutoModel.from_pretrained(model_id, trust_remote_code=True).to(device).eval()

# Load tokenizer used in training (ESM2 650M)
tok = AutoTokenizer.from_pretrained("facebook/esm2_t33_650M_UR50D")
peptide = "GILGFVFTL"
PEP_LEN = 16
PAD_ID = tok.pad_token_id if tok.pad_token_id is not None else 1

def pad_to_len(ids_list, target_len, pad_id):
    return ids_list + [pad_id] * (target_len - len(ids_list)) if len(ids_list) < target_len else ids_list[:target_len]
pep_ids = tok(peptide, add_special_tokens=True)["input_ids"]
pep_ids = pad_to_len(pep_ids, PEP_LEN, PAD_ID)
pep_tensor = torch.tensor([pep_ids], dtype=torch.long, device=device)
with torch.no_grad():
    logits, features = model(pep_tensor)
    prob_bind = F.softmax(logits, dim=1)[0, 1].item()
    pred = int(prob_bind >= 0.5)

print({"peptide": peptide, "pre_prob": round(prob_bind, 6), "label": pred})

Batch Inference (Python)

import torch
import torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer

# Device
device = "cuda" if torch.cuda.is_available() else "cpu"

# 加载 PRE 模型（注意：确保该模型确实支持 logits 输出）
model_id = "SkywalkerLu/TriStageHLA-PRE"
model = AutoModel.from_pretrained(model_id, trust_remote_code=True).to(device).eval()

# 加载与训练一致的 ESM2 tokenizer
tok = AutoTokenizer.from_pretrained("facebook/esm2_t33_650M_UR50D")

# 固定长度（需与训练一致）
PEP_LEN = 16
PAD_ID = tok.pad_token_id if tok.pad_token_id is not None else 1

def pad_to_len(ids_list, target_len, pad_id):
    if len(ids_list) < target_len:
        return ids_list + [pad_id] * (target_len - len(ids_list))
    return ids_list[:target_len]

# 示例批次（替换为你的真实肽序列列表）
batch = [
    {"peptide": "GILGFVFTL"},
    {"peptide": "NLVPMVATV"},
    {"peptide": "SIINFEKL"},
    {"peptide": "GLCTLVAML"},
]

# 批量分词与填充
pep_ids_batch = []
for item in batch:
    pep = item["peptide"]
    ids = tok(pep, add_special_tokens=True)["input_ids"]
    ids = pad_to_len(ids, PEP_LEN, PAD_ID)
    pep_ids_batch.append(ids)

# 转为张量 [B, PEP_LEN]
pep_tensor = torch.tensor(pep_ids_batch, dtype=torch.long, device=device)

# 前向计算
with torch.no_grad():
    logits, features = model(pep_tensor)   # 期望 logits 形状: [B, 2]
    probs = F.softmax(logits, dim=1)[:, 1] # 取类别1（bind）的概率

# 二分类标签（0/1）
labels = (probs >= 0.5).long().tolist()

# 打印结果
for i, item in enumerate(batch):
    print({
        "peptide": item["peptide"],
        "pre_prob": float(probs[i].item()),
        "label": labels[i]
    })