Mining microbial dark matter -- a new idea

Gu He health

An in-depth study of the microbiome , It is expected to provide a new solution to our health problems .

However, most microbial genomes have not been cultured yet , Even in the genome sequences that have been found , There are a lot of Unable to comment function . Therefore, the functional diversity of microbial phylogenetic tree cannot be fully captured , this This limits our ability to model advanced features of biological sequences . The construction of the model is an important part of the microbiome research .

Here we introduce a The newly released deep learning model , It provides some new ideas for the microbial research .

What can this model do ？

The researchers built LookingGlass Program , application RNN Cyclic neural networks and LSTM Learning method of short - and long-term memory neural network , Learn each nucleotide character in the sequence , In order to achieve Be able to predict different functions of classification 、 Homology and environmental origin reads Purpose .

LookingGlass Model also Have the ability of transfer learning , After fine tuning, you can perform a series of different tasks ： For example, identifying new Oxidoreductases , Predict the optimum temperature of the enzyme , And identifying amino acid sequences .

LookingGlass The model can represent other unknown and uncommented sequences in terms of function , thus Mining microbial dark matter .

Fang Law

1► LookingGlass The training set of the model 、 Verification set And test set selection

The genome sequences of representative bacteria and archaea are classified by GTDB51（89.0 edition ） determine .

The complete genome sequence passed  NCBI Genbank ftp download . That's what happened  24,706 Genome , Include 23,458 Bacterial genomes and 1248 Archaea genome .

To determine their actual sequence length , Use MetaSeek API Downloaded their sequencing metadata . Remove the length ＜60bp or ＞300bp The sample of , Finally, the average sequence length is 136bp Total of 7909 Samples .

LookingGlass The training set of the model 、 Both the verification set and the test set are in Class level Divided on , Between the three at this classification level No overlapping parts .

• The verification set is Actinobacteria, Alphaproteobacteria, Thermoplasmata, Bathyarchaeia Under each species 8 Genome , total 32 Genome ;
• The test set is Bacteroidia, Clostridia, Methanosarcinia, Nitrososphaeria Total under species 32 Genome ;
• The training set is for each of the remaining categories 1 Genome , total 32 Archaea genomes and 298 Bacterial genomes .

2►LookingGlass Model architecture and training

Main application RNN loop Neural networks and LSTM Long and short term memory neural network .

LookingGlass Use Three layers LSTM Encoder model , Each hidden layer has 1152 A unit , According to the result of super parameter adjustment ,embedding The size is 104.

LookingGlass Train in a self supervised way , According to the context of the preceding nucleotides in the sequence , Predict a masked nucleotide .

For each... In the training set sequence reads, Consider multiple training inputs , Move the masked nucleotide from the second position to the last position along the length of the sequence . Because it is a character level model , Linear decoder from possible vocabulary “A”、“C”、“G” and “T” The next nucleotide in the predicted sequence , with “ Start reading ” Special mark of 、“ Unknown nucleotides ”（ In the case of ambiguous sequences ）、“ End of read ”（ stay LookingGlass Only during training “ Read start ” marked ） and “ fill ” Mark （ For classification purposes only ）.

LSTM Regularization and optimization of utilize dropout And gradient descent method for optimal performance , Use fastai Library training .

Hardware aspect ,LookingGlass stay Microsoft Azure On , Memory is 16GB Of Pascal P100 GPU Training . In total, I trained 12 God , common 75 individual epoch, The learning rate is gradually reduced according to the result of hyperparametric optimization ：15 individual epoch, The learning rate is 1e-2,15 individual epoch, The learning rate is 2e-3, And 1e-3 The learning rate of 45 individual epoch.

3► Superparameter optimization

Adjust the super parameters by random search , Major adjustments ：

• kmer size
• stride
• number of LSTM layers
• number of hidden nodes per layer
• dropout rate
• weight decay
• momentum
• embedding size
• bptt size
• learning rate
• batch size

The main result

LookingGlass Be able to capture the functional characteristics of the sequence

Multi classification confusion matrix for functional annotation prediction . The horizontal axis represents the true value , The vertical axis represents the predicted value . The value in the box is the normalized prediction percentage , On the left is the correct prediction , On the right is the wrong prediction .

chart a Represents the validation set EC Prediction of the first position of the function number , chart b Indicates the prediction of the second position , The accuracy of the display is 80% above .

LookingGlass Able to capture phylogenetic features of sequences

LookingGlass Identify homologous sequence pairs at the gate level . Blue is homologous (Homologous), Red is non homologous (Nonhomologous).

chart a by embedding Degree of similarity Comparison between groups , Homologous group was significantly higher than non homologous group ,embedding The similarity is calculated as embedding Cosine similarity between vectors .

chart b by Accuracy 、 accuracy 、 Recall rate and F1 Changes in scores , Visible in embedding The similarity threshold is 0.62 Its accuracy (accuracy) The highest , Reached 66.4%, This means Door level Of . In the article, it shows that it has reached the level of 68.3%, At the target level 73.2%, It has reached the level of 76.6%, stay Genus level Up to 78.9%.LookingGlass Use embedding Methods to distinguish homologous and non homologous sequences , Without depending on their sequence similarity (Smith-Waterman comparison ).

chart c The results of searching homologues by these two methods are compared , The black box in the figure indicates that LookingGlass Correctly identified homologous sequences , But when using the comparison, we missed . It can be seen that many homologues have very low sequence similarity (bit score＜50), Cannot be captured by comparison based methods , but LookingGlass Sure .LookingGlass High accuracy in identifying homologous genes , It has nothing to do with their sequence similarity , It shows that it captures a high level of features , It may reflect the phylogenetic relationship between sequences .

LookingGlass Distinguish the sequence of different environmental samples

come from 100 Metagenomic functional annotation sets of different environmental samples are used as validation sets . From there, for each environment group Random sampling 20000 Sequence calculation embedding Similarity degree . Find out Intergroup Of embedding similarity It is usually lower than that in the group Of , That is, sequences from the same environmental background are usually clustered together .

LookingGlass Support the transfer of learning tasks

With LookingGlass As a starting point , Fine tune the pre training model , To perform different tasks .

1. Predict taxonomic oxidoreductase

Yes LookingGlass After fine tuning the function annotation classification model , Perform the prediction and classification task of oxidoreductase .

The above figure shows the sequence similarity (bit score＜50) Sequence , Predict whether the classification is the coding gene of oxidoreductase Accuracy 、 accuracy 、 Recall rate and F1 Changes in scores , The result indicates that the default threshold is 0.5 when , Its accuracy (accuracy) The highest , by 82.3%.

be based on LookingGlass The model can Distinguish between different environmental backgrounds Sequence under This advantage , The researchers used data from 16 Marine metagenomes as test sets , The sample coverage ranges from latitude （ from -62 C to 76 degree ）, Ocean depth ( From the surface ~5 Meters to the middle ~ 200-1000 rice ) And oxygen concentration ( Including... From the lowest oxygen zone 4 Middle and upper layer samples ), From which each macrogenome is randomly selected 2000 ten thousand reads.

Mining the sequence of oxidoreductase , And prove that LookingGlass The classification of oxidoreductases is superior to the traditional homology based methods .

chart a by LookingGlass In the surface ocean (surface)、 Middle zone (mesopelagic) And the area with the lowest oxygen concentration (OMZ) Predict the proportion classified as oxidoreductase sequence .

chart b In the ocean surface zone group , Correlation between latitude and oxidoreductase , The results showed that there was a significant correlation (R2 = 0.79,P = 0.04).

chart c For separate use LookingGlass、MG-RAST and  mi-faser Tool search oxidoreductase sequence , It was predicted to be oxidoreductase (oxidoreductases)、 Non oxidoreductase (not oxidoreductases) And uncommented (unannotated) Sequence proportion of . The result is MG-RAST The comments 26.7-50.3% Of reads, among 0.01-4.0% Identified as oxidoreductase .Mi-faser The comments 0.17-2.9% Of reads, among 0.04-0.59% Identified as oxidoreductase . so ,LookingGlass More advantage .

2. Use LookingGlass Identify amino acid sequences

LookingGlass Directly from CDS Predict the starting position of the translation frame （1、2、3、-1、-2 or -3）, The accuracy is up to 97.8%, But it is only used for non coding at present DNA A lower proportion of prokaryotic sequences .

3. from DNA The sequence fragment predicts the optimal temperature of the enzyme

The optimum temperature of the enzyme depends in part on DNA Sequence characteristics , But it's hard to predict , Especially short reading and long reading . Divide the temperature into psychrophilic （<15°C）、 Thermophilic （20-40°C） Or thermophilic （>50°C）, fine-tuning LookingGlass After procedure , The input sequence predicts the best temperature category , Accuracy up to 70.1%.

junction On

LookingGlass The program can Not by comparison Refer to the database to pre Measure and characterize DNA Sequence , thus Obtain functional notes and phylogenetic information . meanwhile ,LookingGlass The transfer learning framework can quickly learn 、 Training and convergence for different classification tasks , This provides some contributions to the modeling of complex biological systems in the future .

It is predicted that taxonomic oxidoreductase can Mining the potential functions of location sequences , Predictable enzymes will be expanded in the future . Predicting the optimal temperature of enzyme can be used to guide the function and optimal temperature of protein design .

To make a long story short , This is a good exploration . The author has encapsulated the model functions used in this article as python library —fastBio, The data set used in this paper can be directly used for model training , Available at ：github.com/ahoarfrost/fastBio/

Main references ：Hoarfrost A, Aptekmann A, Farfañuk G, Bromberg Y. Deep learning of a bacterial and archaeal universal language of life enables transfer learning and illuminates microbial dark matter. Nat Commun. 2022 May 11;13(1):2606. doi: 10.1038/s41467-022-30070-8. PMID: 35545619; PMCID: PMC9095714.