About SJI

What Is the Signal Jaccard Index?

The Signal Jaccard Index (SJI) is a protein-similarity metric that compares proteins by their signal proteins: the set of homologs separated from more distant homologs by data-derived gaps in a two-dimensional similarity plot. Unlike conventional pairwise sequence-similarity approaches, SJI does not compare two proteins only by their direct sequence similarity. Instead, it compares the homolog neighborhoods associated with each protein [1].

For any given protein, referred to here as the seed, its top similar proteins are retrieved across many proteomes and placed on a two-dimensional plot. The x-axis represents the length ratio between each homolog and the seed protein, and the y-axis represents full-length sequence similarity (Figure 1C; Figure 2A–D). In most cases, the points separate into two groups with a visible gap between them. The upper group is close to the seed in both similarity and length and is defined as the signal. The lower group forms a denser cloud of more distant matches and is defined as the noise. This gap is not determined by a fixed similarity cutoff. Instead, it is inferred from the local density structure of each plot using unsupervised spectral clustering, so each protein has its own protein-specific boundary [1].

The signal set is what SJI compares. Given two proteins A and B, each with its own signal set, SJI is defined as the size of the intersection of their signal sets divided by the size of their union (Figure 1D). If A and B share most of their signal homologs across proteomes, their SJI is high. Building the metric from signal sets, rather than from a single pairwise similarity score, makes SJI sensitive to the broader genomic and evolutionary context in which each protein occurs.

What SJI Actually Measures

SJI is not simply another orthology score. It uses whole-proteome composition to identify a signal set for each protein and then measures how much two proteins' signal sets overlap. Two proteins with comparable functions are expected to have similar neighborhoods of homologs across proteomes, and SJI directly measures that shared neighborhood.

This design gives SJI several advantages. First, it incorporates information from many homologs rather than relying on a single pairwise alignment. Second, it handles inparalogs more naturally than strict one-to-one best-hit methods. Third, the resulting SJI network exposes difficult proteins as peripheral nodes. In the ProtDC framework, the network core represents proteins for which orthology prediction is relatively stable across databases, whereas the periphery contains proteins for which database assignments are more inconsistent [1].

In summary, SJI was developed from the observation that ortholog-database inconsistencies are not merely computational artifacts. They reflect, at least in part, the structure of protein evolutionary space. By identifying protein-specific gaps between signal and noise homologs, SJI provides a data-driven framework for evaluating protein similarity, ortholog consistency, and functional annotation transfer.

Why This Matters for Similarity Cutoffs

The gaps shown in Figure 2 also expose a limitation of conventional similarity-search workflows. BLAST and related tools [4] provide alignment scores, E-values, and ranked hits, but the interpretation of these values often depends on thresholds chosen by researchers. Whether a similarity score is considered high enough for ortholog prediction, functional annotation transfer, or comparative analysis is frequently determined by empirical convention or practical convenience rather than by a protein-specific biological boundary.

This is problematic because different protein families evolve under different constraints. A cutoff that is stringent for one family may be permissive for another. The valley in each two-dimensional plot provides a per-protein cutoff inferred from the data rather than imposed in advance. In this sense, SJI provides a way to replace arbitrary global similarity thresholds with locally inferred boundaries that are specific to each protein's homolog distribution.

A Note on Scale and Interpretation

Some criticism of this framework was based on experimental studies showing that a single amino acid substitution can dramatically change protein function. Such cases are real and biologically important, particularly in microevolution, protein engineering, and population genetics. However, they do not directly address the problem of functional annotation transfer across species.

Functional annotation transfer is a macroevolutionary question. It asks whether a protein in one organism can be treated as functionally equivalent to a homolog in another organism, often across substantial evolutionary distances. At this scale, the relevant issue is not whether one mutation can alter function under a specific experimental condition. The relevant issue is whether homologs across many proteomes form distinguishable clusters in similarity space. The recurring valleys observed in the two-dimensional plots suggest that such clusters are objectively defined.