Update README.md

README.md

@@ -1,14 +1,67 @@
-# TransHLA2.0
-
-**TransHLA2.0** 是一个基于 ESM2 + LoRA + CNN-Transformer 融合的蛋白质序列建模模型。  
-支持 Hugging Face Hub 上的 `trust_remote_code` 加载和自定义微调。
-
-## 快速加载
-
-```python
-from transformers import AutoModel
-
-model = AutoModel.from_pretrained(
-    "SkywalkerLu/TransHLA2.0",
-    trust_remote_code=True
-)
+# TransHLA2.0
+
+A minimal Hugging Face-compatible PyTorch model for peptide–HLA binding classification 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:
+```bash
+pip install torch transformers peft
+
+
+import torch
+import torch.nn.functional as F
+from transformers import AutoModel, AutoTokenizer
+
+# Device
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+# Load model (replace with your model id if different)
+model_id = "SkywalkerLu/TransHLA2.0"
+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")
+
+# Example inputs
+peptide = "GILGFVFTL"  # 9-mer example
+# Fake placeholder pseudosequence for demo; replace with a real one from your mapping/data
+hla_pseudoseq = (
+    "GSHSMRYFYTAVSRPGRGEPRFIAVGYVDDTQFVRFDSDAASPRMEPRAPWIEQEGPEYWERETRNVK"
+    "AQSQTDRVDLRTLLRYNQSEAGSHTVQRMYGCDVGSDWRFLRGYHQYAYDGKDYIALNEDLRSWTAAD"
+    "MAAQTTKHKWEQAGAAER"
+)
+
+# Fixed lengths (must match training)
+PEP_LEN = 16
+HLA_LEN = 36
+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]
+
+# Tokenize
+pep_ids = tok(peptide, add_special_tokens=True)["input_ids"]
+hla_ids = tok(hla_pseudoseq, add_special_tokens=True)["input_ids"]
+
+# Pad/truncate
+pep_ids = pad_to_len(pep_ids, PEP_LEN, PAD_ID)
+hla_ids = pad_to_len(hla_ids, HLA_LEN, PAD_ID)
+
+# Tensors (batch=1)
+pep_tensor = torch.tensor([pep_ids], dtype=torch.long, device=device)
+hla_tensor = torch.tensor([hla_ids], dtype=torch.long, device=device)
+
+# Forward + probability
+with torch.no_grad():
+    logits, features = model(pep_tensor, hla_tensor)
+    prob_bind = F.softmax(logits, dim=1)[0, 1].item()
+    pred = int(prob_bind >= 0.5)
+
+print({"peptide": peptide, "bind_prob": round(prob_bind, 6), "label": pred})