Every test sequence searched against the train set with MMseqs2, and scored on its best hit. Similarity is min(query coverage, target coverage) × percent identity: how much of the shorter sequence the alignment spans, times how much of that matches. The histogram is the distribution of those best hits.

The flag is the percentage of test sequences at or above the threshold: Pass at 0%, Warning below 2%, Fail at 2% or more. A test sequence that near-duplicates a training one is one the model has already seen, so the accuracy it earns there is not evidence. What to do about it →

The panel under the histogram lists those alignments, up to the first 100; every one of them, listed or not, is exported to mmseqs/mmseqs2_search_result.tsv beside this report. Its columns are:

Query (Q)

Identifier of the test sequence: seq_<i>_test, where i is its 0-based position in the input test file.

Target (T)

Identifier of the train sequence: seq_<i>_train, where i is its 0-based position in the input train file.

Q Cov.

Fraction of the query sequence covered by the alignment (0–1).

T Cov.

Fraction of the target sequence covered by the alignment (0–1).

% Identity

Percent identical aligned positions in the aligned region.

% Similarity

Similarity score used for leakage detection, calculated as min(Q Cov., T Cov.) × % Identity.

E-value

MMseqs2 E-value (lower is more "significant").

Alignment

Click “Show” to expand the alignment visualisation for a row.