BeanovaLi commited on
Commit
8120729
·
verified ·
1 Parent(s): c9df6d3

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +170 -33
README.md CHANGED
@@ -1,33 +1,170 @@
1
- ---
2
- license: bsd-3-clause
3
- ---
4
- Mirror of the base ProGen2-small model (with slightly modified configuration and forward pass) by [Nijkamp, et al.](https://arxiv.org/abs/2206.13517).
5
-
6
- See also my github [repo](https://github.com/hugohrban/ProGen2-finetuning/tree/main) for an example of finetuning this model.
7
-
8
- Example usage:
9
-
10
- ```python
11
- from transformers import AutoModelForCausalLM
12
- from tokenizers import Tokenizer
13
- import torch
14
- import torch.nn.functional as F
15
-
16
- # load model and tokenizer
17
- model = AutoModelForCausalLM.from_pretrained("hugohrban/progen2-small", trust_remote_code=True)
18
- tokenizer = Tokenizer.from_pretrained("hugohrban/progen2-small")
19
- tokenizer.no_padding()
20
-
21
- # prepare input
22
- prompt = "1MEVVIVTGMSGAGK"
23
- input_ids = torch.tensor(tokenizer.encode(prompt).ids).to(model.device)
24
-
25
- # forward pass
26
- logits = model(input_ids).logits
27
-
28
- # print output probabilities
29
- next_token_logits = logits[-1, :]
30
- next_token_probs = F.softmax(next_token_logits, dim=-1)
31
- for i in range(tokenizer.get_vocab_size(with_added_tokens=False)):
32
- print(f"{tokenizer.id_to_token(i)}: {100 * next_token_probs[i].item():.2f} %")
33
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # TargetAMP-Designer Finetuning
2
+
3
+ **TargetAMP-Designer** is an end-to-end pipeline for **target-specific peptide design and in silico screening**, accompanying my bachelor thesis and paper. It integrates (i) **LLM-based peptide generation via fine-tuning**, and (ii) a **multi-stage structural evaluation workflow** (structure prediction → docking → MD/MetaD → consistency validation).
4
+
5
+
6
+
7
+ ## Highlights
8
+
9
+ - **Fine-tune** a ProGen2-style autoregressive model with labeled peptide data and generate new candidates by controllable sampling.
10
+ - **Multi-stage screening** using ZDOCK (rigid docking) → Rosetta (flexible refinement) → MD/Metadynamics validation.
11
+ - **Consistency validation** scripts for dataset similarity/identity and physicochemical property profiling.
12
+
13
+
14
+
15
+ ## Repository layout (recommended)
16
+
17
+ Adjust the paths below to match your local clone directory.
18
+
19
+ - `src/`
20
+
21
+ `data/` – training / test data
22
+
23
+ `models/weights/` – base model weights directory
24
+
25
+ `finetune.py` – model fine-tuning
26
+
27
+ `sample.py` / `sample.sh` – sequence sampling
28
+
29
+ `checkpoints/` – fine-tuned model outputs
30
+
31
+ `requirements.txt`
32
+
33
+ - `src/experiments/` – docking / simulation / analysis workflow
34
+
35
+
36
+
37
+ ## Installation安装
38
+
39
+ Use a virtual environment (conda/venv) to avoid conflicts, then install dependencies:
40
+
41
+ ```bash
42
+ cd src
43
+ pip install -r requirements.txt
44
+ ```
45
+
46
+ > For GPU training/inference, ensure CUDA and a compatible PyTorch are installed in your environment.
47
+ >
48
+ > Docking / MD steps require external tools (ZDOCK, Rosetta, OpenMM, etc.) configured in `src/experiments/`
49
+
50
+
51
+
52
+ ## Quickstart
53
+
54
+ ### 1.Fine-tuning
55
+
56
+ Fine-tune on prepared train/test files (GPU recommended). The base weights are hosted on huggingface [repo](https://huggingface.co/BeanovaLi/TargetAMP-Designer). After finetuning, the model binary, config file and tokenizer are saved into the `./checkpoints/` directory.
57
+
58
+ ```python
59
+ python finetune.py \
60
+ --model=/home/TargetAMP-Designer/src/models/weights \
61
+ --train_file=/home/TargetAMP-Designer/src/data/train.txt \
62
+ --test_file=/home/TargetAMP-Designer/src/data/test.txt \
63
+ --device=cuda \
64
+ --epochs=20 \
65
+ --batch_size=16 \
66
+ --accumulation_steps=4 \
67
+ --lr=1e-4 \
68
+ --decay=cosine \
69
+ --warmup_steps=200 \
70
+ --eval_before_train
71
+ ```
72
+
73
+ Run `python3 finetune.py --help` to see the full list of available options and their descriptions.
74
+
75
+ ### 2.Sampling
76
+
77
+ We provide **two equivalent ways** to run sequence generation:
78
+
79
+ **Option A: run the Python script with CLI arguments.**This is convenient for quick experiments and parameter sweeps.Use the `sample.py` script to generate new sequences from a model using top-k sampling with temperature.
80
+
81
+ ```python
82
+ python sample.py
83
+ --model=/home/TargetAMP-Designer/src/models/weights/ \
84
+ --device=cuda \
85
+ --batch_size=100 \
86
+ --iters=1 \
87
+ --max_length=50 \
88
+ --t=1.0 \
89
+ --k=10 \
90
+ --prompt="1<|label|>"
91
+ ```
92
+
93
+ Use the `--help` or `-h` option to see the full list of available options and their descriptions.
94
+
95
+ **Option B: run a bash script with parameters written inside.**For better reproducibility, you can also run a `.sh` script (e.g., `sample.sh`) where all arguments are predefined and version-controlled.
96
+
97
+ ```bash
98
+ bash sample.sh
99
+ ```
100
+
101
+ > Tip: We recommend using the `.sh` entrypoint for paper-level experiments, as it keeps sampling hyperparameters (temperature, top-k, prompt format, max length, batch size, random seed if applicable) explicitly recorded.
102
+
103
+
104
+
105
+ ## Structural screening workflow(experiments)
106
+
107
+ All downstream experiments are under `./src/experiments/`.
108
+
109
+ ### 3.Molecular Docking
110
+
111
+ Before molecular docking, we optionally run **Boltz-2** to predict peptide (or peptide–target) 3D structures as docking-ready inputs.
112
+
113
+ - Run ZDOCK with:
114
+
115
+ ```bash
116
+ bash zdock_bash.sh
117
+ ```
118
+
119
+ The docking results are saved as PDB files; you can change `result_base_dir` inside the script to control the output directory.
120
+
121
+ - Extract docking complexes:
122
+
123
+ ```python
124
+ python createpl_complex.py
125
+ ```
126
+
127
+ - flexible refinement:
128
+
129
+ Relax receptor/ligand first, then run flexible docking using the provided flags and scripts.
130
+
131
+ ```python
132
+ python relax_ligands.py
133
+ python relax_receptor.py
134
+ python dock.py
135
+ ```
136
+
137
+ > We perform Rosetta-based refinement following a standard *Relax → Prepack → Dock* workflow. First, the receptor and ligand are relaxed separately using the flag files `relax_flag_ChainR` (receptor) and `relax_flag_ChainL` (ligand), executed via `relax_receptor.py` and `relax_ligands.py`, respectively. Next, the docking prepack stage is carried out using the prepack configuration `docking_prepack_flag` to optimize side-chain packing at the interface before docking. Finally, Rosetta docking is performed with the docking configuration `docking_flag` using `dock.py` to generate refined complex models and corresponding scores.
138
+
139
+ ### 4.Molecular dynamics
140
+
141
+ Run the MD pipeline:
142
+
143
+ ```python
144
+ python MD_simulation.py
145
+ ```
146
+
147
+ ### 5.Consistency Vailidation & property profiling
148
+
149
+ - Dataset similarity/identity:
150
+
151
+ ```bash
152
+ python needle_comp_similarity_identity.py
153
+ ```
154
+
155
+ - Physicochemical properties:
156
+
157
+ ```python
158
+ python physical_chemical_properties.py
159
+ ```
160
+
161
+
162
+
163
+ ## Citation
164
+
165
+ If you found this code useful, please cite:
166
+
167
+ ```bash
168
+
169
+ ```
170
+