Assembly graph forks stop being a topology guess: sample-specific sequence scoring ranks extensions, and raw reads can check the choice

Choosing a path at a branching point can now be scored from sequence statistics instead of relying only on graph topology or paired-end evidence. Graph construction tools typically stop unitig extension when several outgoing edges appear and leave the ambiguity to downstream assemblers, which then use paired-end information, coverage, or topology heuristics. On HIV-1 and SARS-CoV-2 unitig graphs, the sample-adaptive RASTRO-S improved its match with read-supported choices as context length grew from n=5 to n=10, above random selection at roughly 40% and 43% and a longest-extension greedy baseline at 24.7% and 32.8%; for HIV-1 at n=10 the abstract reports 83.5% while the results table reports 75.8%, and the two disagree.

Scoring uses only the sequence formed by the upstream context and a candidate extension, with no reference genome and no assembled contig. RASTRO-S is an n-th order Markov model trained on the target sample's own unitigs, estimating the conditional probability of each base given the preceding n bases, with unobserved contexts falling back to the unigram distribution. A candidate's score is the average negative log conditional probability, or surprisal, over scorable positions, reported as bits per base; RASTRO-L instead uses the pretrained DNA language model Evo2 to produce autoregressive per-base probabilities, normalized by extension length, and both pick the lowest-scoring candidate.

Branch decisions are evaluated against raw reads rather than final contigs, avoiding the bias of a particular assembly. Established quality metrics such as contiguity, misassembly rate, and consensus accuracy measure a finished assembly or corrected reads, and cannot say whether one branch choice is supported by the reads. Evaluation counts only boundary k-mers spanning the junction between context and extension in the raw reads and takes the best-supported candidate as the read-supported choice; among 20,000 sampled branching points, 15,319 HIV-1 branches (76.6%) and 19,882 SARS-CoV-2 branches (99.4%) yielded a unique read-supported choice.

The reported scope covers two viral samples, HIV-1 and SARS-CoV-2, using Logan unitig graphs, and evaluates local choices at branching points rather than complete genome assembly. Readers interested in wiring sequence scoring into an assembly pipeline can watch the direction the authors propose: using RASTRO scores alongside coverage, paired-end information, and graph topology, and testing at larger scale whether they improve path ranking and branch resolution.

The authors note that model scores and read-supported choices capture different aspects of the sequencing data, so agreement should not be expected to reach 100%; the two criteria may diverge at branches with multiple plausible paths, heterogeneous coverage, polymorphic signals, or sampling effects. RASTRO-S outperforming RASTRO-L is explained as sample-adaptive modeling aligning more closely with sample-specific read evidence in this setting, not as general superiority over larger language models. The authors also state that computational cost limited the number of replicates, the size of evaluated genomes, and the range of comparable methods, and that two datasets are insufficient to establish broad generalizability, which requires evaluation on more diverse genomes and sequencing conditions. In addition, the abstract and the results table give different RASTRO-S match values for HIV-1 at n=10, so readers should treat the results table and figure values as the reported ones and note the disagreement.