TransfoRNA

TransfoRNA is a bioinformatics and machine learning tool based on Transformers to provide annotations for 11 major classes (miRNA, rRNA, tRNA, snoRNA, protein -coding/mRNA, lncRNA, YRNA, piRNA, snRNA, snoRNA and vtRNA) and 1923 sub-classes for human small RNAs and RNA fragments. These are typically detected by RNA-Seq NGS (next generation sequencing) data.

Embeddings of RNAs collected from two different datasets; TCGA and inhouse sequenced LC (lung cancer) data can be visualized on our hosted portal here

TransfoRNA can be trained on just the RNA sequences and optionally on additional information such as secondary structure. The result is a major and sub-class assignment combined with a novelty score (Normalized Levenshtein Distance) that quantifies the difference between the query sequence and the closest match found in the training set. Based on that it decides if the query sequence is novel or familiar. TransfoRNA uses a small curated set of ground truth labels obtained from common knowledge-based bioinformatics tools that map the sequences to transcriptome databases and a reference genome. Using TransfoRNA's framewok, the high confidence annotations in the TCGA dataset can be increased by 3 folds.

Sections:

• Visualize RNAs External link!
• Resources
• Datasets
• Models
• Repo Structure
• Installation
• Inference
• Train
• Novelty Prediction

Resources

The data used for training and the trained models can be downloaded from here.

The downloaded folder contains three subfolders (should be kept on the same folder level as transforna):

• data: Contains three files:
  • TCGA__ngs__miRNA_log2RPM-24.06.0.h5ad anndata with ~75k sequences and var columns containing the knowledge based annotations.
  • HBDXBase.csv containing a list of RNA precursors which are then used for data augmentation.
  • subclass_to_annotation.json holds mappings for every sub-class to major-class.

The models folders contains both benchmark and models trained on tcga.

• models:
  • benchmark : contains benchmark models trained on sncRNA and premiRNA data. (See additional datasets at the bottom)
  • tcga: All models trained on the TCGA data; TransfoRNA_ID (for testing and validation) and TransfoRNA_FULL (the production version) containing higher RNA major and sub-class coverage. Each of the two folders contain all the models trained seperately on major-class and sub-class.
• kba_pipeline: contains mapping reference data required to run the knowledge based pipeline manually

Datasets:

• TCGA: The Cancer Genome Atlas, TCGA offers sequencing data of small RNAs and is used to train and evaluate TransfoRNAs classification performance
  • Sequences are annotated based on a knowledge-based annotation approach that provides annotations for ~2k different sub-classes belonging to 11 major classes.
  • Knowledge-based annotations are divided into three sets of varying confidence levels: a high-confidence (HICO) set, a low-confidence (LOCO) set, and a non-annotated (NA) set for sequences that could not be annotated at all. Only HICO annotations are used for training.
  • HICO RNAs cover ~2k sub-classes and constitute 19.6% of all RNAs found in TCGA. LOCO and NA sets comprise 66.9% and 13.6% of RNAs, respectively.
  • HICO RNAs are further divided into in-distribution, ID (374 sub-classes) and out-of-distribution, OOD (1549 sub-classes) sets.
    • Criteria for ID and OOD: Sub-class containing more than 8 sequences are considered ID, otherwise OOD.
  • An additional putative 5' adapter affixes set contains 294 sequences known to be technical artefacts. The 5’-end perfectly matches the last five or more nucleotides of the 5’-adapter sequence, commonly used in small RNA sequencing.
  • The knowledge-based annotation (KBA) pipline including installation guide is located under kba_pipline
• LC: NGS sequencing data from a clinical lung cancer (LC) study
  • This dataset was only used for validation.
• Additional Datasets:
  • sncRNA, collected from RFam (classification of RNA precursors into 13 classes)
  • premiRNA human miRNAs(classification of true vs pseudo precursors)

Models

There are 5 classifier models currently available, each with different input representation.

• Baseline:
  • Input: (single input) Sequence
  • Model: An embedding layer that converts sequences into vectors followed by a classification feed forward layer.
• Seq:
  • Input: (single input) Sequence
  • Model: A transformer based encoder model.
• Seq-Seq:
  • Input: (dual inputs) Sequence divided into even and odd tokens.
  • Model: A transformer encoder is placed for odd tokens and another for even tokens.
• Seq-Struct:
  • Input: (dual inputs) Sequence + Secondary structure
  • Model: A transformer encoder for the sequence and another for the secondary structure.
• Seq-Rev (best performant):
  • Input: (dual inputs) Sequence
  • Model: A transformer encoder for the sequence and another for the sequence reversed.

Note: These (Transformer) based models show overlapping and distinct capabilities. Consequently, an ensemble model is created to leverage those capabilities.

Repo Structure

conf: Contains the configurations of each model, training and inference settings.

The conf/main_config.yaml file offers options to change the task, the training settings and the logging. The following shows all the options and permitted values for each option.

transforna contains two folders:

• src folder which contains transforna package. View transforna's architecture here.
• bin folder contains all scripts necessary for reproducing manuscript figures.

Installation

Installation could be done from source or using pip:

• Installation from source

  The install.sh is a script that creates an transforna environment in which all the required packages for TransfoRNA are installed. Simply navigate to the root directory and run from terminal:

  #make install script executable
  chmod +x install.sh
  
  
  #run script
  ./install.sh
  

• Installation using pip

  #TODO: Change next line to pip install transforna, once package is on pypi
  pip install git+ssh://git@github.com/gitHBDX/transforna.git
  

Inference

Models could be used for inference in one of three ways:

• TransfoRNA API

  In transforna/src/inference/inference_api.py, all the functionalities of transforna are offered as APIs. There are two functions of interest:

  • predict_transforna : Computes for a set of sequences and for a given model, one of various options; the embeddings, logits, explanatory (similar) sequences, attentions masks or umap coordinates.
  • predict_transforna_all_models: Same as predict_transforna but computes the desired option for all the models as well as aggregates the output of the ensemble model. Both return a pandas dataframe containing the sequence along with the desired computation.

  Check the script at src/test_inference_api.py for a basic demo on how to call the either of the APIs.

• Inference from repo

  For inference, two paths in configs/inference_settings/default.yaml have to be edited:

  • sequences_path: The full path to a csv file containing the sequences for which annotations are to be inferred.
  • model_path: The full path of the model. (currently this points to the Seq model)

  Also in the main_config.yaml, make sure to edit the model_name to match the input expected by the loaded model.

  • model_name: add the name of the model. One of "seq","seq-seq","seq-struct","baseline" or "seq-rev" (see above)

  Then, navigate the repositories' root directory and run the following command:

  python transforna/__main__.py inference=True
  

  After inference, an inference_output folder will be created under outputs/ which will include two files.

  • (model_name)_embedds.csv: contains vector embedding per sequence in the inference set- (could be used for downstream tasks). Note: The embedds of each sequence will only be logged if log_embedds in the main_config is True.
  • (model_name)_inference_results.csv: Contains columns; Net-Label containing predicted label and Is Familiar? boolean column containing the models' novelty predictor output. (True: familiar/ False: Novel) Note: The output will also contain the logits of the model is log_logits in the main_config is True.

• Inference from huggingface

  TransfoRNA Models are uploaded to huggingface 🤗: Seq Seq-Seq Seq-Struct Seq-Rev

  refer to the script: TransfoRNA/scripts/test_huggingface_transforna_model.ipynb for an example.

Train

TransfoRNA can be trained using input data as Anndata, csv or fasta. If the input is anndata, then anndata.var should contains all the sequences. Some changes has to be made (follow configs/train_model_configs/tcga):

In configs/train_model_configs/custom:

• dataset_path_train has to point to the input_data which should contain; a sequence column, a small_RNA_class_annotation coliumn indicating the major class if available (otherwise should be NaN), five_prime_adapter_filter specifies whether the sequence is considered a real sequence or an artifact (True for Real and False for artifact), a subclass_name column containing the sub-class name if available (otherwise should be NaN), and a boolean column hico indicating whether a sequence is high confidence or not.
• If sampling from the precursor is required in order to augment the sub-classes, the precursor_file_path should include precursors. Follow the scheme of the HBDxBase.csv and have a look at PrecursorAugmenter class in transforna/src/processing/augmentation.py
• mapping_dict_path should contain the mapping from sub class to major class. i.e: 'miR-141-5p' to 'miRNA'.
• clf_target sets the classification target of the mopdel and should be either sub_class_hico for training on targets in subclass_name or major_class_hico for training on targets in small_RNA_class_annotation. For both, only high confidence sequences are selected for training (based on hico column).

In configs/main_config, some changes should be made:

• change task to custom or to whatever name the custom.py has been renamed.
• set the model_name as desired.

For training TransfoRNA from the root directory:

python transforna/__main__.py

Using Hydra, any option in the main config can be changed. For instance, to train a Seq-Struct TransfoRNA model without using a validation split:

python transforna/__main__.py train_split=False model_name='seq-struct'

After training, an output folder is automatically created in the root directory where training is logged. The structure of the output folder is chosen by hydra to be /day/time/results folders. Results folders are a set of folders created during training:

• ckpt: (containing the latest checkpoint of the model)
• embedds:
  • Contains a file per each split (train/valid/test/ood/na).
  • Each file is a csv containing the sequences plus their embeddings (obtained by the model and represent numeric representation of a given RNA sequence) as well as the logits. The logits are values the models produce for each sequence, reflecting its confidence of a sequence belonging to a certain class.
• meta: A folder containing a yaml file with all the hyperparameters used for the current run.
• analysis: contains the learned novelty threshold seperating the in-distribution set(Familiar) from the out of distribution set (Novel).
• figures: some figures are saved containing the Normalized Levenstein Distance NLD, distribution per split.

Novelty Prediction

A KNN graph is trained per model taking as an input the embeddings of all the RNAs used for training. Following that, different validation RNA sets (with familiar and novel ground truths) are used to learn a threshold seperating "novel" RNAs from "familiar" RNAs. Familiar meaning RNAs belonging to sub classes, the model has encountered during training while novel is the opposite. This learned threshold is the Normalized Levenshtein Distance threshold, NLD which can be interpreted as the %of nucleotide difference to make the query sequence be identical to the closest match (Explanatory sequence) found by the KNN model. A score of a query sequence above that threshold will consider query sequence as Novel, otherwise familiar. Refer to this script for more details on how this is done. Note:

• Each model folder contains a trained knn graph under model_name/post_model/knn_model.sav.
• The learned NLD threshold is logged per model at model_name/analysis/novelty_model_coef.yaml
• To use the model directly, adapt this function transforna/utils/utils.py/get_closest_neighbors
• All Inference output contains the output of the novelty prediction.