Identifying Emerging Biological Threats in Grove

Operators need to rapidly identify the pathogen in a sample with high confidence to guide downstream decisions. In a real-world scenario, a potential pathogen could comprise a small fraction of the sequencing data, making it difficult to distinguish a true detection from background noise or misleading matches. Detection needs to be sensitive enough to catch these signals without generating false alarms that take away limited force health protection resources.

All short fragments of genomic sequence from dengue virus were flagged in our sample, with all but two of them passing confirmatory checks. The agreement between the screening and confirmatory checks provides strong confidence for the presence of dengue virus (Figure 3).

Grove uses a two-stage approach. First, a fast and sensitive screen identifies potential pathogens. Then, more targeted checks determine whether the evidence supports each detection.

The dengue example is a single scenario; benchmarking lets us assess detection performance more broadly across other priority biological threats. We evaluated Grove detection accuracy in a comprehensive benchmark, which measures performance across high-priority biological agents in a complex sample background. Grove correctly identifies the target in all samples (120/120), and correctly avoids false-positive detections in all samples designed to test misidentification (25/25) (Figure 4).

The geographic origin of a sequence can indicate whether we’re seeing the result of local spread or a new introduction from somewhere else. The case of this hypothetical individual in Hawaii with no recent travel raises concern about how dengue virus may have reached Hawaii. Estimating where a virus came from can be difficult because the timing of its introduction may be unknown, viral surveillance leaves sampling blind spots, and closely related viruses can circulate in multiple places. Geographic attribution requires strong supporting evidence and an explicit accounting of uncertainty.

In our hypothetical example, the results indicated the dengue virus sequence originated from lineages in Nicaragua, and suggested a possible introduction from Nicaragua to Hawaii. This is consistent with our ground truth in this scenario, the ancestry proposed for the original isolate by Shin et al. 2013.

Grove applies two complementary methods: sequence similarity comparison and an evolutionary tree analysis. Both approaches rely on stitching together the dengue virus sequence fragments we recover into a genome, and comparing this genome to other known dengue virus isolates. Applying two complementary methods allows us to check whether the geographic output holds up under different ways of analyzing the sequences, strengthening the result when they agree and adding uncertainty when they do not.

Benchmarking lets us assess how well geographic attribution performs across a broader set of sequences. We used a larger panel of human viruses, and tested methods in Grove against a simple baseline that applies BLAST to find the closest reference sequence and assigns its recorded collection country. For dengue, Grove assigned the country to 88% of sequences and abstained on 12%; among calls made, 95% were correct. Across the larger panel, performance varied by virus, with more accurate assignments when closely related sequences tend to come from the same country. Overall, Grove assigned a country to 47% of sequences and, of those, 79% were correct (Figure 7). The BLAST baseline, which always makes an assignment, assigned 90% of dengue sequences correctly, and 59% across all viruses. Together, these results show that Grove’s sequence similarity method achieves a higher precision than the baseline in exchange for leaving more cases unresolved when the evidence is insufficient.

As part of our framework for detecting engineered threats, Grove looks for evidence of chimeric viral genomes: genomes containing segments from two or more viruses joined within the same molecule. While chimerism can arise naturally through recombination, it can also be a sign of deliberate engineering or manipulation of a sequence in the lab. If the dengue virus sequence found in our scenario showed evidence of chimerism, it would warrant escalation and further investigation into the origin.

In this dengue virus case study, Grove did not flag a chimeric signal. This is the expected result because we know that the genome in our sample is a natural dengue virus genome. Nevertheless, for an unknown sample, a negative result would not exclude other forms of manipulation or chimeric junctions without sequencing coverage.

Grove looks for genomic fragments whose two ends match different viral species. That provides evidence that material from two viruses may be joined into a single genome.

Benchmarking lets us measure how well chimera detection performs across a broader range of viruses and engineered changes. We built a benchmark consisting of several RNA virus families, with chimeric examples containing changes involving different parts of the viral genome. Using discordance of the fragment ends, Grove distinguishes samples containing chimeras from unmodified controls better than two baseline approaches (Figure 9).