ViralVectors: compact and scalable alignment-free virome feature generation

Abstract

The amount of sequencing data for SARS-CoV-2 is several orders of magnitude larger than any virus. This will continue to grow geometrically for SARS-CoV-2, and other viruses, as many countries heavily finance genomic surveillance efforts. Hence, we need methods for processing large amounts of sequence data to allow for effective yet timely decision-making. Such data will come from heterogeneous sources: aligned, unaligned, or even unassembled raw nucleotide or amino acid sequencing reads pertaining to the whole genome or regions (e.g., spike) of interest. In this work, we propose ViralVectors, a compact feature vector generation from virome sequencing data that allows effective downstream analysis. Such generation is based on minimizers, a type of lightweight “signature” of a sequence, used traditionally in assembly and read mapping — to our knowledge, the first use minimizers in this way. We validate our approach on different types of sequencing data: (a) 2.5M SARS-CoV-2 spike sequences (to show scalability); (b) 3K Coronaviridae spike sequences (to show robustness to more genomic variability); and (c) 4K raw WGS reads sets taken from nasal-swab PCR tests (to show the ability to process unassembled reads). Our results show that ViralVectors outperforms current benchmarks in most classification and clustering tasks.

Graphical Abstract

Graphical Abstract showing the all steps of proposed approach. We start by collecting the sequence-based data. Then Data cleaning and preprocessing is applied. After that, we generate the feature embeddings using minimizer based approach. Then Classification and clustering algorithms are applied on the resultant data and predictions are made on the test set.

Appendix A: Baseline Models

We use different baseline and state-of-the-art (SOTA) methods to compare the results with ViralVectors. The baseline model that we are using is One Hot Embedding (OHE) [21] while the SOTA methods are Spike2Vec [3], PWM2Vec [2], and Pango Tool [29].

Table 7 Contingency tables of variants vs clusters after applying k-means on the OHE based feature embedding on GISAID data

Table 8 Contingency tables of variants vs clusters after applying k-means on the Spike2Vec based feature embedding on GISAID data

Table 9 Contingency tables of variants vs clusters after applying k-means on the ViralVectors based feature embedding on GISAID data

Table 10 Contingency tables of variants vs clusters after applying k-means on the OHE based feature embedding on ViPR data

Table 11 Contingency tables of variants vs clusters after applying k-means on the Spike2Vec based feature embedding on ViPR data

Table 12 Contingency tables of variants vs clusters after applying k-means on the ViralVectors based feature embedding on ViPR data

Table 13 Contingency tables of variants vs clusters after applying k-means on the OHE based feature embedding on NCBI short read data

1.1 A.1 One hot embedding (OHE) [21]

Since most machine learning methods does not work with the biological sequence based feature vectors, it is important to convert the them into a numerical representation. A traditional method to convert sequential information into numerical representation is called one-hot embedding [4, 21]. Given a finite set of symbols in a sequence (e.g., spike sequence), we call this set as an alphabet, denoted by $\Sigma$. In the GISAID amino acid sequences, for example, we have 21 unique characters “ACDEFGHIKLMNPQRSTVWXY” (i.e., amino acids). To design a fixed-length feature vector representation, we generate a length 21 binary vector for each amino acid, which contains value 1 for the position of that specific character and zero everywhere else. At the end, we concatenate all these vectors to get a final feature vector representation for a given sequence. In GISAID amino acid sequences, since the length of each spike amino acid sequence is 1273, the length of each OHE based vector is $1273 \times 21 = 26,733$ (more detail on the dataset can be found in Section 4.1). For the ViPR data, since the length of each spike amino acid sequence (after alignment) is 3498 (and the length of unique characters is 24), therefore, the length of OHE vector is $3498 \times 24 = 83,952$. In the case of NCBI raw short reads sequencing data, the OHE does not apply, since we have variable-length unmapped reads rather than a single fixed length sequence. After generating the feature vectors, we can give these vectors as an input to machine learning algorithms for classification and clustering purposes.

Remark 1

Note that one problem with OHE is that it required all sequences in a data to be of fixed-length [1, 4].

1.2 A.2 Spike2Vec [3]

Since OHE does not work with the variable length sequences, a popular alignment-free method is using k-mers to preserve the order of amino acids and then generating a fixed-length feature vector that contains the frequency of each k-mer in a virome sequence. In this setting, the first step is to compute the substrings (called mers) of length k, where k is the user defined parameter. The k-mers are generated using sliding window approach with the increment of 1 (see Fig. 1). The total number of possible k-mers that can be generated from a virome sequence is “N - k + 1", where N is the length of sequence.

Table 14 Contingency tables of variants vs clusters after applying k-means on the Spike2Vec based feature embedding on NCBI short read data

Table 15 Contingency tables of variants vs clusters after applying k-means on the ViralVectors based feature embedding on NCBI short read data

1.2.1 A.2.1 Fixed-length representation

Since each virome sequence can have different number of k-mers, it is important to generate fixed-length numerical representation so that classification and clustering algorithms could be applied. For this purpose, we design a feature vector of length length $|\Sigma |^{k}$ (where $\Sigma$ is the alphabet and k is user defined parameter for k-mers) that contains the frequency/count of each k-mer within a sequence. In this paper, we are taking $k=3$ for all experiments unless specifically mentioned otherwise (decided using standard validation set approach [9]). In the GISAID dataset, since the total number of alphabets are 21, the length of Spike2Vec based feature vector is $21^{3} = 9261$. For the ViPR dataset, the length of Spike2Vec based vector is $25^{3} = 15625$, for NCBI short reads data, the length of Spike2Vec based vector is $24^{3} = 13842$.

1.3 A.3 PWM2Vec [2]

When using a Spike2Vec method, the frequency vectors obtained is comparatively low dimension but still is high dimensional. Moreover, while generating the frequency vectors, matching the k-mers to the appropriate location/bin in the vector (bin matching) can be computationally expensive. To solve these issues, PWM2Vec [2] can be used. It is a recently proposed method for producing a fixed-length numerical feature vector based using the well known position-weight matrix notion [37]. PWM2Vec creates a PWM from the sequence’s k-mers, and the final feature vector contains the score of each k-mer in the PWM. This enables the method to use k-mers ability to collect localization information while also capturing the significance of each amino acid’s position in the sequence (information that is lost in computing k-mer frequency vector). By combining these pieces of data in this way, a compact and broad feature embedding can be created that can be used for a variety of downstream machine learning tasks.

1.4 A.4 Pango tool [29]

For clustering purpose, we also use the state-of-the-art clustering benchmark called pango tool [29]. Since pango tool takes multiple aligned sequence as input, we needed to align each read set to the reference genome, call (genomic) variants, and introduce these variants into the reference sequence to generate a consensus sequence which represents this particular sample — the pipeline is available as a Snakefile [27] in our shared code repository above. The SARS-CoV-2 reference genome sequence (INSDC accession $GCA\_009858895.3$, sequence MN9089047) used in this study is obtained from Ensemble COVID-19 browser database, Ensemble COVID-19 [19, 20]. It is a complete genome of 29903 bps. The genome the reference assembly of the viral RNA genome Isolates of the first cases Wuhan-HU-1, China [44] and has been reportedly used as the standard reference widely [35].