# Functional overview of the SDK¶

The software development kit (SDK), formally known as selene_sdk, is an extensible Python package intended to ease development of new programs that leverage sequence-level models through code reuse. The package is composed of six submodules: sequences, samplers, targets, predict, interpret, and utils. It also provides two top-level classes: TrainModel and EvaluateModel. In the following sections, we briefly discuss each submodule and top-level class.

## Sampling¶

We start with the modules for sampling data because both training and evaluting a model in Selene will require a user to specify the kind of sampler they want to use.

### sequences submodule (API)¶

The sequences submodule defines the Sequence type, and includes implementations for several sub-classes. These sub-classes–Genome and Proteome–represent different kinds of biological sequences (e.g. DNA, RNA, amino acid sequences), and implement the Sequence interface’s methods for reading the reference sequence from files (e.g. FASTA), querying subsequences of the reference sequence, and subsequently converting those queried subsequences into a numeric representation. Further, each sequence class specifies its own alphabet (e.g., nucleotides, amino acids) to represent query results as strings.

### targets submodule (API)¶

The targets submodule defines the Target class, which specifies the interface for classes to retrieve labels or “targets” for a given query sequence. At present, we supply a single implementation of this interface: GenomicFeatures. This class takes a tabix-indexed file of intervals for each label we want our model to predict, and uses this file to identify the labels for a given sequence drawn from the reference.

### samplers submodule (API)¶

The samplers submodule provides methods and classes for randomly sampling and partitioning datasets for training and evaluation. The Sampler interface defines the minimal requirements for a class fulfilling these functions. In particular, samplers must be able to partition data (i.e. into training, validation, and testing datasets), sample data from each partition, and, if needed, save the sampled data to a file. Further, a file of names must be provided for the features to be predicted. We provide several implementations adhering to the Sampler interface: the RandomPositionsSampler, IntervalsSampler, and MultiFileSampler.

MultiFileSampler draws samples from structured data files for each partition. There is currently support for loading either .bed or .mat files via the FileSampler classes BedFileSampler and MatFileSampler, respectively (see API docs for file samplers). It is worth noting that the .bed file used by BedFileSampler includes the coordinates of each sequence, and the indices corresponding to each feature for which said sequence is a positive example. We hope that users will request or contribute classes for other file samplers in the future. MultiFileSampler does not support saving the sampled data to a file, so calling the save_dataset_to_file method from this class will have no effect.

RandomPositionsSampler and IntervalsSampler are what we call online samplers. Online samplers generate examples from the reference sequence (e.g. genome, proteome) on-the-fly–either across the whole reference sequence (random positions sampler), or from user-specified regions (intervals sampler)–using a tabix-indexed .bed file. These samplers automatically partition said data according to user-specified parameters (e.g. validate on a subset of chromosomes or on some percentage of the data). Since OnlineSampler’s samples are randomly generated, we allow the user to save the sampled data to file. This file can be subsequently loaded with the BedFileSampler. They rely on classes from the sequences and targets submodules for retrieving each sequence and its targets in the proper matrix format.

## Training a model (API)¶

The TrainModel class may be used for training and testing of sequence-based models, and provides the core functionality of the CLI’s train command. It relies on an OnlineSampler (or a subclass of OnlineSampler) to automatically partition the dataset into subsets for training, validation, and testing. These subsets are then used to automatically train and validate performance for a user-specified number of steps. The testing subset is used to evaluate the model performance after training is completed. The model’s loss, area under the receiver operating characteristic curve (AUC), and area under the precision-recall curve (AUPRC) are logged during training. (In the future, we plan to support other performance metrics. Please request specific ones or use cases in our Github issues. The frequency of logging is provided by the user. At the end of evaluation, TrainModel logs the performance metrics for each feature predicted, and produces plots of the precision recall and receiver operating characteristic curves.

## Evaluating a model (API)¶

The EvaluateModel class is used to test the performance of a trained model. EvaluateModel uses an instance of Sampler class or subclass to draw samples from a test set. After using the provided model to predict labels for said data, EvaluateModel logs the performance measures (as described in “Training a model”) and generates figures and a performance breakdown by feature.

## Using a model to make predictions (API)¶

Selene’s predict submodule includes a number of methods and classes for making predictions with sequence-based models. The AnalyzeSequences class is the main class to use. It leverages a user-specified trained model to make predictions for sequences sequences in a FASTA file, apply in silico mutagenesis to sequences in a FASTA file, or perform variant effect prediction on variants in a VCF file. In each case, the user can specify what AnalyzeSequences should save: raw predictions, difference scores, absolute difference scores, and/or logit scores. Note that the aforementioned “scores” can only be computed for in silico mutagenesis and variant effect prediction.

## Visualizing model predictions (API)¶

The interpret submodule of selene_sdk provides methods for visualizing a sequence-based model’s predictions made with AnalyzeSequences. For example, interpret includes methods for processing variant effect predictions made with AnalyzeSequences and subsequently visualizing them with a heatmap or sequence logo. The functionality included in the interpret submodule is not heavily incorporated into the CLI, but is instead intended for incorporation into user code.

## The utilities submodule (API)¶

Unlike the aforementioned submodules designed around individual concepts, the utils submodule is a catch-all submodule intended to prevent cluttering of the selene_sdk top-level namespace. It provides diverse functionality at varying levels of flexibility. Some members of utils are general-purpose (e.g. configuration file parsing) while others have highly specific use cases (e.g. CLI logger initialization).