README.md 7.45 KB
Newer Older
zhuwenwen's avatar
zhuwenwen committed
1
2
3
4
<!--
 * @Author: zhuww
 * @email: zhuww@sugon.com
 * @Date: 2023-03-31 17:09:07
zhuwenwen's avatar
zhuwenwen committed
5
 * @LastEditTime: 2023-08-24 09:07:01
zhuwenwen's avatar
zhuwenwen committed
6
7
-->
# FastFold
zhuwenwen's avatar
zhuwenwen committed
8
9
10
## 论文
- [https://arxiv.org/abs/2203.00854](https://arxiv.org/abs/2203.00854)

zhuwenwen's avatar
zhuwenwen committed
11
12
13
## 模型结构
模型基于Transformer架构,主要结构包括Evofomer(48 blocks)和Struture module(8 blocks)两个模块。

zhuwenwen's avatar
zhuwenwen committed
14
15
![img](./docs/alphafold2.png)

zhuwenwen's avatar
zhuwenwen committed
16
17
## 算法原理
FastFold通过搜索同源序列和模板进行特征构造,基于蛋白质结构预测模型,进行推理的性能优化,预测蛋白质的结构。
zhuwenwen's avatar
zhuwenwen committed
18

zhuwenwen's avatar
zhuwenwen committed
19
## 环境配置
zhuwenwen's avatar
zhuwenwen committed
20
提供[光源](https://www.sourcefind.cn/#/service-details)拉取推理的docker镜像:
zhuwenwen's avatar
zhuwenwen committed
21
22
23
24
25
26
27
28
29
30
```
docker pull image.sourcefind.cn:5000/dcu/admin/base/custom:fastfold-0.2.1-centos7.6-dtk-22.10-patch4-py38-latest
docker run -it --name fastfold --shm-size=32G  --device=/dev/kfd --device=/dev/dri/ --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video image.sourcefind.cn:5000/dcu/admin/base/custom:fastfold-0.2.1-centos7.6-dtk-22.10-patch4-py38-latest /bin/bash
```

镜像版本依赖:
* DTK驱动:dtk22.10
* Pytorch: 1.10
* fastfold: 0.2.1
* python: python3.8
zhuwenwen's avatar
zhuwenwen committed
31
32
33
34
35
36
37
38

激活镜像环境:
`source /opt/dtk-22.10/env.sh`
`source /opt/openmm-dtk-22.10/env.sh`

测试目录:
`/opt/docker/test`

zhuwenwen's avatar
zhuwenwen committed
39
## 数据集
zhuwenwen's avatar
zhuwenwen committed
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
推荐使用AlphaFold2中的开源数据集,包括BFD、MGnify、PDB70、Uniclust、Uniref90等,数据集大小约3TB。数据集格式如下:
```
$DOWNLOAD_DIR/                             
    bfd/  
        bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_hhm.ffindex
        bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_hhm.ffdata 
        bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_cs219.ffindex                           
        ...
    mgnify/                                
        mgy_clusters_2018_12.fa
    params/                                
        params_model_1.npz
        params_model_2.npz
        params_model_3.npz
        ...
    pdb70/                                
        pdb_filter.dat
        pdb70_hhm.ffindex
        pdb70_hhm.ffdata
        ...
    pdb_mmcif/                            
        mmcif_files/
            100d.cif
            101d.cif
            101m.cif
            ...
        obsolete.dat
    pdb_seqres/                            
        pdb_seqres.txt
    small_bfd/                           
        bfd-first_non_consensus_sequences.fasta
    uniclust30/                            
        uniclust30_2018_08/
            uniclust30_2018_08_md5sum
            uniclust30_2018_08_hhm_db.index
            uniclust30_2018_08_hhm_db
            ...
    uniprot/                               
        uniprot.fasta
    uniref90/                             
        uniref90.fasta
```
zhuwenwen's avatar
zhuwenwen committed
82
83
84
85
86
87
88

我们提供了一个脚本download_all_data.sh用于下载使用的数据集和模型文件:

    ./scripts/download_all_data.sh 数据集下载目录

## 推理
我们分别提供了基于Pytorch的单体和多体的推理脚本。
zhuwenwen's avatar
zhuwenwen committed
89

zhuwenwen's avatar
zhuwenwen committed
90
### 单体
zhuwenwen's avatar
zhuwenwen committed
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110

    python inference.py T1024.fasta data/pdb_mmcif/mmcif_files/ \
    --output_dir ./ \
    --gpus 2 \
    --use_precomputed_alignments alignments/ \
    --param_path /data/params/params_model_1.npz  \
    --uniref90_database_path data/uniref90/uniref90.fasta \
    --mgnify_database_path data/mgnify/mgy_clusters_2018_12.fa \
    --pdb70_database_path data/pdb70/pdb70 \
    --uniclust30_database_path data/uniclust30/uniclust30_2018_08/uniclust30_2018_08 \
    --bfd_database_path data/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \
    --jackhmmer_binary_path `which jackhmmer` \
    --hhblits_binary_path `which hhblits` \
    --hhsearch_binary_path `which hhsearch` \
    --kalign_binary_path `which kalign` \
    --chunk_size 4 \
    --inplace

或者使用`./inference.sh`

zhuwenwen's avatar
zhuwenwen committed
111
#### 单体推理参数说明
zhuwenwen's avatar
zhuwenwen committed
112
T1024.fasta为推理的单体序列;data修改为数据集下载目录;
zhuwenwen's avatar
zhuwenwen committed
113
114
115
`--output_dir`为输出目录;`--gpus`为使用的gpu数量;`--use_precomputed_alignments`为搜索对齐目录,可以加载已经搜索对齐的序列,若不添加则进行搜索对齐;
`--param_path`为加载单体模型路径,需要和`--model_name`保持一致,默认为model_1;`--chunk_size`为分块数量,设置为4,并且使用`--inplace`来降低显存占用;
默认进行relax操作,若不需要,添加`--relaxation`;默认不保存输出的.pkl文件,若需要,添加`--save_outputs`.
zhuwenwen's avatar
zhuwenwen committed
116
117


zhuwenwen's avatar
zhuwenwen committed
118
Alphafold的数据预处理需要花费大量时间,因此我们通过[ray](https://docs.ray.io/en/latest/workflows/concepts.html)加快了数据预处理工作流程。
zhuwenwen's avatar
zhuwenwen committed
119
120
要使用ray工作流运行推理,应将参数--enable_workflow添加到cmdline或`./inference.sh`脚本中。

zhuwenwen's avatar
zhuwenwen committed
121
### 多体
zhuwenwen's avatar
zhuwenwen committed
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
    python inference.py SUGP1.fasta data/pdb_mmcif/mmcif_files/ \
    --output_dir ./ \
    --gpus 2 \
    --use_precomputed_alignments alignments/ \
    --model_preset multimer \
    --uniref90_database_path data/uniref90/uniref90.fasta \
    --mgnify_database_path data/mgnify/mgy_clusters_2018_12.fa \
    --pdb70_database_path data/pdb70/pdb70 \
    --uniclust30_database_path data/uniclust30/uniclust30_2018_08/uniclust30_2018_08 \
    --bfd_database_path data/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \
    --uniprot_database_path data/uniprot/uniprot_sprot.fasta \
    --pdb_seqres_database_path data/pdb_seqres/pdb_seqres.txt  \
    --param_path data/params/params_model_1_multimer.npz \
    --model_name model_1_multimer \
    --jackhmmer_binary_path `which jackhmmer` \
    --hhblits_binary_path `which hhblits` \
    --hhsearch_binary_path `which hhsearch` \
    --kalign_binary_path `which kalign` \
    --chunk_size 4 \
    --inplace 

或者使用`./inference_multimer.sh`

zhuwenwen's avatar
zhuwenwen committed
145
#### 多体推理参数说明
zhuwenwen's avatar
zhuwenwen committed
146
SUGP1.fasta为推理的多体序列;`--param_path`为加载多体模型路径,需要和`--model_name`保持一致,其他参数同单体推理参数说明一致.
zhuwenwen's avatar
zhuwenwen committed
147

zhuwenwen's avatar
zhuwenwen committed
148
149
150
151
152
153
154
155
156
157
158
159
160
161
## result
`--output_dir`目录结构如下:
```
alignments/
    <target_name>/
        bfd_uniclust_hits.a3m
        mgnify_hits.sto
        uniref90_hits.sto
        ...
{target_name}_{model_name}_output_dict.pkl
{target_name}_{model_name}_unrelaxed.pdb
{target_name}_{model_name}_relaxed.pdb
```

zhuwenwen's avatar
zhuwenwen committed
162
163
164
查看蛋白质3D结构:[https://www.pdbus.org/3d-view](https://www.pdbus.org/3d-view)
![img](./docs/result_pdb.png)

zhuwenwen's avatar
zhuwenwen committed
165
## 精度
zhuwenwen's avatar
zhuwenwen committed
166
167
测试数据:[casp14](https://www.predictioncenter.org/casp14/targetlist.cgi)[uniprot](https://www.uniprot.org/),使用的加速卡:4张 DCU 1代-16G

zhuwenwen's avatar
zhuwenwen committed
168
169
170
171
172
173
174
175
1、计算lddt的值

    python3 pkl2plddt.py
    其中,data_path为推理生成的pkl文件路径。


2、其它精度值计算:[https://zhanggroup.org/TM-score/](https://zhanggroup.org/TM-score/)

zhuwenwen's avatar
zhuwenwen committed
176
177
178
179
180
181
182
准确性数据:
| 数据类型 | 序列类型 | 序列标签 | 序列长度 | GDT-TS | GDT-HA | LDDT | TM score | MaxSub | RMSD |
| :------: | :------: | :------: | :------: |:------: |:------: | :------: | :------: | :------: |:------: |
| fp32 | 单体 | T1026  | 172  | 0.914 | 0.765 | 79.634 | 0.941 | 0.907 | 1.289 |
| fp32 | 单体 | T1053  | 580  | 0.937 | 0.782 | 92.284 | 0.984 | 0.929 | 1.105 |
| fp32 | 单体 | Q9NYK1 | 1046 | 0.907 | 0.744 | 86.642 | 0.962 | 0.905 | 5.757 |

zhuwenwen's avatar
zhuwenwen committed
183
184
185
186
187
## 应用场景

### 算法类别
NLP

zhuwenwen's avatar
zhuwenwen committed
188
### 热点应用行业
zhuwenwen's avatar
zhuwenwen committed
189
医疗,科研,教育
zhuwenwen's avatar
zhuwenwen committed
190
191


zhuwenwen's avatar
zhuwenwen committed
192
## 源码仓库及问题反馈
zhuwenwen's avatar
zhuwenwen committed
193
* https://developer.hpccube.com/codes/modelzoo/FastFold
zhuwenwen's avatar
zhuwenwen committed
194

zhuwenwen's avatar
zhuwenwen committed
195
196
197
## 参考
* [https://github.com/deepmind/alphafold](https://github.com/deepmind/alphafold)
* [https://github.com/hpcaitech/FastFold](https://github.com/hpcaitech/FastFold)