TL;DR
This paper investigates whether genomic signatures in early SARS-CoV-2 samples can reveal their temporal origin. The researchers found that n-mer distributions in early samples could predict their collection month with ~78% accuracy, suggesting a temporal signal in the virus's genome. However, this signal diminishes with later samples, and isolation sources could not be predicted from these genomic patterns.
What Problem Does This Paper Address?
The rapid global spread of SARS-CoV-2 in early 2020 generated a massive amount of genomic data. Understanding the virus's evolution, origin, and spread is crucial for controlling pandemics. While many studies focused on phylogenetic relationships and protein structures, this paper tackles a specific question: can the genomic sequence itself, particularly in its early stages, reveal temporal information about when and where a sample was collected? The authors aim to determine if variations in genomic composition, specifically n-mer distributions, can serve as a "temporal footprint" for SARS-CoV-2, especially during its initial outbreak phase.
Methods
The core of this study revolves around analyzing the genomic sequences of SARS-CoV-2 samples. The primary method employed is the extraction and analysis of n-mer distributions. N-mers are contiguous subsequences of length 'n' within a DNA sequence. By counting the frequencies of different n-mers (specifically 1-mers, 2-mers, and 3-mers, resulting in 84 features), the researchers aimed to capture unique genomic signatures. These frequency counts were normalized by sequence length.
To test the predictive power of these n-mer features, the researchers applied several classification algorithms, including k-NN, AdaBoost, and Random Forest. These models were trained to predict the collection month of the samples. The performance of these models was evaluated using standard metrics like accuracy, precision, recall, F1 score, and AUC, with 5-fold cross-validation.
Additionally, the paper includes hierarchical clustering to group samples based on their n-mer similarity, and phylogenetic analysis to understand the evolutionary relationships between SARS-CoV-2 and other coronaviruses. Protein alignment was also performed to identify conserved protein regions.
Datasets
The study utilized publicly available SARS-CoV-2 genomic sequences from the NCBI Virus portal. The initial dataset comprised 87 sequences, from which 36 were selected for detailed analysis. These 36 sequences were chosen based on the criterion that their originating countries had at least 5 samples available. The selected samples were from China (17), Japan (5), and the USA (14). Metadata collected for these sequences included geographical location, isolation source, and month of collection. For phylogenetic analysis, sequences of other coronaviruses were also obtained from NCBI.
Key Insights and Ideas
- Genomic "Time Capsules": The most striking finding is that early SARS-CoV-2 genomic sequences contain discernible temporal information. The n-mer patterns act like a snapshot, reflecting the virus's state at a particular point in time.
- Early Temporal Signal Fades: The predictive power of n-mer distributions for collection month was significant for early samples but disappeared when later samples were included. This suggests that the virus's genomic landscape evolved rapidly, obscuring the initial temporal signal.
- Geographic vs. Temporal Influence: While samples from Japan (collected later) formed a distinct cluster, the study highlights that the temporal aspect (collection month) was a stronger predictor than geographical location alone for early samples.
- Isolation Source is Not a Genomic Predictor: Despite variations in isolation sources (e.g., oronasopharynx, saliva), n-mer patterns could not reliably predict this information, indicating that these genomic features are not strongly linked to the host tissue from which the virus was isolated.
- Evolutionary Connections: Phylogenetic analysis confirmed SARS-CoV-2's close relationship with human SARS and bat coronaviruses, reinforcing existing knowledge about its origins.
Key Results
- Temporal Prediction Accuracy: The Random Forest model achieved approximately 78% accuracy in predicting the collection month of SARS-CoV-2 samples based on n-mer distributions.
- Classification Performance: Table 1 shows that for temporal prediction, Random Forest achieved an accuracy of 77.8%, with an AUC of 0.820. AdaBoost also performed well with 77.8% accuracy and an AUC of 0.793.
- Isolation Source Prediction: Classification accuracy for predicting isolation sources was significantly lower, with the best model (Random Forest) achieving only 58.3% accuracy and an AUC of 0.800, indicating poor predictive power.
- Phylogenetic Overlap: The phylogenetic tree showed a strong overlap between SARS-CoV-2, human SARS, and bat coronaviruses, with some USA-collected SARS-CoV-2 samples forming a distinct cluster.
Limitations and Future Work
- Limited Temporal Scope: The study was limited to samples collected within the first few months of the outbreak (December 2019 - January 2020). The authors acknowledge that with the inclusion of newer samples (March and April 2020), the distinctive temporal pattern disappeared.
- Data Skew: The authors note that the data is skewed over time due to the exponential growth in sequencing efforts.
- Future Directions: The paper suggests that further investigation is needed for a deeper understanding of the temporal change of SARS-CoV-2 sequences at a later phase. It also highlights the importance of exploring how mutations are dependent on the geographic origin of samples, especially given the virus's known diversity and recurrent occurrence in humans. The potential for future outbreaks due to aberrant mutations in different locations is also raised.
Summary updated