LucaOne APP

TimeLine

• 2024/10/01: optimized embedding inference code: src/llm/lucagplm/get_embedding.py
• 2024/08/01: add checkpoint=17600000, location: checkpoint-step17600000
  This project will download the checkpoint automatically according to the value of parameter --llm_step.

1. Embedding

Two embedding methods for nucleic acid or protein sequence: matrix or vector.
suggestion: If matrix is applied, it can be converted to a vector in the downstream networks as follows when using the embedding matrix:

• [CLS] Vector(matrix[0, :])
• Avg Pooling
• Max Pooling(recommend)
• Value-Level Attention Pooling(recommend) (ref: https://arxiv.org/abs/2210.03970)
  ...
  Pooling: transform the embedding matrix into a vector.
  Recommend: Use pooling in downstream networks rather than in data pre-processing.
  Notice: If your task is sequence, use the pooling operation; otherwise, it is not required.

2. Environment Installation

step1: update git

1) centos

sudo yum update
sudo yum install git-all

2) ubuntu

sudo apt-get update
sudo apt install git-all

step2: install python 3.9

1) download anaconda3

wget https://repo.anaconda.com/archive/Anaconda3-2022.05-Linux-x86_64.sh

2) install conda

sh Anaconda3-2022.05-Linux-x86_64.sh

Notice: Select Yes to update ~/.bashrc

source ~/.bashrc

3) create a virtual environment: python=3.9.13

conda create -n lucaone_app python=3.9.13

4) activate lucaone_app

conda activate lucaone_app

step3: install other requirements

pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple

3. Preparation

Trained LucaOne Checkpoint FTP: TrainedCheckPoint for LucaOne

Notice:
The project will download automatically LucaOne Trained-CheckPoint from FTP.

When downloading automatically failed, you can manually download:

Copy the TrainedCheckPoint Files(models/ + logs/) from http://47.93.21.181/lucaone/TrainedCheckPoint/* into the directory: ./models/llm/

4. Inference

Scripts in algorithms/
inference_embedding_lucaone.py: embedding using LucaOne(for nucleic acid or protein).
inference_embedding_dnabert2.py: embedding using DNABert2(only for nucleic acid).
inference_embedding_esm.py: embedding using ESM2(only for protein).

建议与说明:
1）尽量使用显存大进行embedding 推理，如：A100，H100，H200等，这样一次性能够处理较长的序列，LucaOne在A100下可以一次性处理2800左右长度的序列；
2）对于超长序列，LucaOne会进行Overlap分片进行embedding，最后合并成完整的embedding，请设置--embedding_complete与--embedding_complete_seg_overlap；
3）如果显卡不足以处理输入的序列长度，会调用CPU进行处理，这样速度会变慢，如果你的数据集中长序列不是很多，那么可以使用这种方式: --gpu_id -1；
4）如果你的数据集中长序列很多，比如: 万条以上，那么再设置--embedding_complete与--embedding_complete_seg_overlap之外，再加上设置--embedding_fixed_len_a_time，表示一次性embedding的最大长度。 如果序列长度大于这个长度，基于这个长度进行分片embedding，最后进行合并。否则根据序列的实际长度；
5）如果不设置--embedding_complete，那么根据设置的--truncation_seq_length的值对序列进行截断embedding；
6）对于蛋白，因为绝大部分蛋白长度在1000以下，因此超长蛋白序列不会很多，因此可以将--embedding_fixed_len_a_time设置长一点或者不设置；
7）对于DNA，因为很多任务的DNA序列很长，那么请设置--embedding_fixed_len_a_time。
如果数据集中超长序列数据量越多，该值设置越小一点，比如在A100下设置为2800，否则设置大一点，如果GPU根据这个长度embedding失败，则会调用CPU。如果数据集数不大，则时间不会很久；
8）对于RNA，因为大部分RNA不会很长，因此与蛋白处理方式一致，因此可以将--embedding_fixed_len_a_time设置长一点或者不设置；

Suggestions and Instructions:

1. Try to use a large GPU-memory machine for embedding reasoning, such as A100, H100, H200, etc., so that long sequences can be processed once.
   LucaOne can process sequences of about 2800 in length at one time under A100;
2. For long sequences, LucaOne will do overlapped fragments in the sequence for embedding and finally merge them into a completed embedding matrix.
   Please set --embedding_complete and --embedding_complete_seg_overlap;
3. If the GPU memory is not enough to process the longer sequence, it will use the CPU for embedding, so the speed will be reduced.
   If your dataset is small, then you can set: --gpu_id -1;
4. If your dataset includes a lot of long sequences (more than 10,000 sequences), please set: --embedding_complete, --embedding_complete_seg_overlap, and --embedding_fixed_len_a_time (represent the maximum length for embedding at one-time).
   If the sequence length is greater than the value of --embedding_fixed_len_a_time, fragment embedding is performed based on this value, and finally, the merge is performed; otherwise, according to the actual length of the sequence;
5. If --embedding_complete is not set, the code will truncate the sequence embedding according to the value of --truncation_seq_length;
6. For proteins, the length of most proteins is less than 1000; there are not many ultra-long protein sequences, so the value of --embedding_fixed_len_a_time can be set a large value or not be set;
7. For DNA, the DNA sequence of many tasks is very long; please set --embedding_fixed_len_a_time.
   The larger the amount of ultra-long sequence, the smaller value should be set, such as 2800 under A100.
   If the GPU embedding fails to process the longer sequence, the CPU will be called.
   When the amount of dataset is not large, the spent time will not be long;
8. For RNA, most RNA is not very long, so the processing method can be consistent with the protein, so the --embedding_fixed_len_a_time can be set a larger value or not be set.

Parameters

1. LucaOne checkpoint parameters:

   • llm_dir: the path for storing the checkpoint LucaOne model，default: ../models/
   • llm_type: the type of LucaOne, default: lucagplm
   • llm_version: the version of LucaOne, default: v2.0
   • llm_task_level: the pretrained tasks of LucaOne, default: token_level,span_level,seq_level,structure_level
   • llm_time_str: the trained time str of LucaOne, default: 20231125113045
   • llm_step: the trained checkpoint of LucaOne, default: 5600000, choices=[5600000, 17600000]

2. Important parameters:

   • embedding_type: matrix or vector, output the embedding matrix or [CLS] vector for the entire sequence, recommend: matrix.
   • trunc_type: truncation type: right or left, truncation when the sequence exceeds the maximum length.
   • truncation_seq_length: the maximum length for embedding(not including [CLS] and [SEP]), itself does not limit the length, depending on the capacity of GPU.
   • matrix_add_special_token: if the embedding is matrix, whether the matrix includes [CLS] and [SEP] vectors.
   • seq_type: type of input sequence: gene or prot, gene for nucleic acid(DNA or RNA), prot for protein.
   • input_file: the input file path for embedding(format: csv or fasta). The seq_id in the file must be unique and cannot contain special characters.
   • save_path: the saving dir for storing the embedding file.
   • embedding_complete: When embedding_complete is set, truncation_seq_length is invalid. If the GPU memory is not enough to infer the entire sequence at once, it is used to determine whether to perform segmented completion (if this parameter is not used, 0.95*len is truncated each time until the CPU can process the length).
   • embedding_complete_seg_overlap: When embedding_complete is set, whether the method of overlap is applicable to segmentation(overlap sliding window)
   • embedding_fixed_len_a_time: When the input sequence is too long for your GPU to complete the inference at once, you can specify the fixed length of the inference at once(default: None)
   • gpu: the gpu id to use(-1 for cpu).

3. Optional parameters:

   • id_idx & seq_idx: when the input file format is csv file, need to use id_idx and seq_idx to specify the column index in the csv (starting with 0).

Notice:
A sequence outputs one embedding file named with seq_id, so the seq_id must be unique and must not contain special characters, such as Spaces, "/", and so on.

Examples:

1) the csv file format of input

Notice:

1. need to specify the column index of the sequence id(id_idx*) and sequence(seq_idx), starting index: 0 .
2. The sequence id must be globally unique in the input file and cannot contain special characters (because the embedding file stored is named by the sequence id, e.g. matrix_seq_1000.pt).

# for protein
cd ./algorithms/
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7,8"
python inference_embedding_lucaone.py \
    --llm_dir ../models  \
    --llm_type lucaone_gplm \
    --llm_version v2.0 \
    --llm_task_level token_level,span_level,seq_level,structure_level \
    --llm_time_str 20231125113045 \
    --llm_step 5600000 \
    --truncation_seq_length 100000 \
    --trunc_type right \
    --seq_type prot \
    --input_file ../data/test_data/prot/test_prot.csv \
    --id_idx 2 \
    --seq_idx 3 \
    --save_path ../embedding/lucaone/test_data/prot/test_prot \
    --embedding_type matrix \
    --matrix_add_special_token \
    --embedding_complete \
    --embedding_complete_seg_overlap \
    --gpu 0   
 
# for DNA or RNA
cd ./algorithms/
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7,8"
python inference_embedding_lucaone.py \
    --llm_dir ../models \
    --llm_type lucaone_gplm \
    --llm_version v2.0 \
    --llm_task_level token_level,span_level,seq_level,structure_level \
    --llm_time_str 20231125113045 \
    --llm_step 5600000 \
    --truncation_seq_length 100000 \
    --trunc_type right \
    --seq_type gene \
    --input_file ../data/test_data/gene/test_gene.csv \
    --id_idx 0 \
    --seq_idx 1 \
    --save_path ../embedding/lucaone/test_data/gene/test_gene \
    --embedding_type matrix \
    --matrix_add_special_token \
    --embedding_complete \
    --embedding_complete_seg_overlap \
    --gpu 0   

2) the fasta file format of input

Notice:

1. The sequence id must be globally unique in the input file and cannot contain special characters (because the embedding file stored is named by the sequence id).

# for protein
cd ./algorithms/
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7,8"
python inference_embedding_lucaone.py \
    --llm_dir ../models \
    --llm_type lucaone_gplm \
    --llm_version v2.0 \
    --llm_task_level token_level,span_level,seq_level,structure_level \
    --llm_time_str 20231125113045 \
    --llm_step 5600000 \
    --truncation_seq_length 100000 \
    --trunc_type right \
    --seq_type prot \
    --input_file ../data/test_data/prot/test_prot.fasta \
    --save_path ../embedding/lucaone/test_data/prot/test_prot \
    --embedding_type matrix \
    --matrix_add_special_token \
    --embedding_complete \
    --embedding_complete_seg_overlap \
    --gpu 0   

# for DNA or RNA
cd ./algorithms/
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7,8"
python inference_embedding_lucaone.py \
    --llm_dir ../models \
    --llm_type lucaone_gplm \
    --llm_version v2.0 \
    --llm_task_level token_level,span_level,seq_level,structure_level \
    --llm_time_str 20231125113045 \
    --llm_step 5600000 \
    --truncation_seq_length 100000 \
    --trunc_type right \
    --seq_type gene \
    --input_file ../data/test_data/gene/test_gene.fasta \
    --save_path ../embedding/lucaone/test_data/gene/test_gene \
    --embedding_type matrix \
    --matrix_add_special_token \
    --embedding_complete \
    --embedding_complete_seg_overlap \
    --gpu 0   

5. Data and Code Availability

FTP:
Pre-training data, code, and trained checkpoint of LucaOne, embedding inference code, downstream validation tasks data & code, and other materials are available: FTP.

Details:

The LucaOne's model code is available at: LucaOne Github or LucaOne.

The trained-checkpoint files are available at: TrainedCheckPoint.

LucaOne's representational inference code is available at: LucaOneApp Github or LucaOneApp.

The project of 8 downstream tasks is available at: LucaOneTasks Github or LucaOneTasks.

The pre-training dataset of LucaOne is opened at: PreTrainingDataset.

The datasets of downstream tasks are available at: DownstreamTasksDataset .

Other supplementary materials are available at: Others .

6. Contributor

Yong He, Zhaorong Li, Yongtao Shan, Yanhong Wei, Yuan-Fei Pan, Mang Shi

7. Zenodo

We have uploaded the model code, training scripts, and embedding inference scripts of LucaOne;
The mode code, training and evaluation scripts, datasets, and trained models for downstream tasks,
and additional supplementary materials to Zenodo (10.5281/zenodo.14604463).
However, due to the substantial size of the pretraining dataset of LucaOne, it has not been included on Zenodo.
Instead, it remains accessible via our publicly available FTP server (LucaOne Pretraining dataset).
We are actively seeking an open FTP platform with sufficient storage capacity to host our pretraining dataset.

LucaOne Zenodo

8. Citation

LucaOne Biorxiv

@article {LucaOne,
author = {Yong He and Pan Fang and Yongtao Shan and Yuanfei Pan and Yanhong Wei and Yichang Chen and Yihao Chen and Yi Liu and Zhenyu Zeng and Zhan Zhou and Feng Zhu and Edward C. Holmes and Jieping Ye and Jun Li and Yuelong Shu and Mang Shi and Zhaorong Li},
title = {LucaOne: Generalized Biological Foundation Model with Unified Nucleic Acid and Protein Language},
elocation-id = {2024.05.10.592927},
year = {2024},
doi = {10.1101/2024.05.10.592927},
publisher = {Cold Spring Harbor Laboratory},
URL = {https://www.biorxiv.org/content/early/2024/05/14/2024.05.10.592927 },
eprint = {https://www.biorxiv.org/content/early/2024/05/14/2024.05.10.592927.full.pdf },
journal = {bioRxiv}
}

9. LucaTeam

Fig. 5 LucaTeam at the West Lake in Hangzhou.