Statistical Genomics Foundry
Paper
domain/comparative-annotation

CESAR 2.0 substantially improves speed and accuracy of comparative gene annotation

1. Citation

Sharma V, Schwede P, Hiller M. 2017. “CESAR 2.0 substantially improves speed and accuracy of comparative gene annotation.” Bioinformatics 33(24):3985–3987. doi:10.1093/bioinformatics/btx527. PMID 28961744. Source code: https://github.com/hillerlab/CESAR2.0 Access date: 2026-07-03.

2. Access note

Read the article HTML at academic.oup.com (article/33/24/3985/4095639). This is a 3-page Applications Note; figures were available as captions in the fetched text. I could read the Abstract, Introduction, Methods description, and Results. I did NOT open the GitHub repository — per contract I worked only from the paper. Consequences: CLI invocation form and exact parameter names are NOT in the paper (paper only points to the repo), so those are unrecoverable from this note (see §7, §10).

3. Thesis (1 sentence)

CESAR 2.0 is a re-engineered Hidden-Markov-Model exon realigner that projects coding-gene annotation from a reference genome to aligned query genomes ~77× faster and with ~31× less memory than the previous version, while also improving splice-site and exon-boundary accuracy.

4. Problem & context

Comparative gene annotation transfers the annotation of a reference genome to other genomes using whole-genome alignments. CESAR (“Coding Exon-Structure Aware Realigner”) realigns exon sequences using an HMM that captures the exon’s reading frame and splice-site annotation, so that projected exons preserve reading frame across insertions/deletions and place splice sites correctly. The prior CESAR version had two limitations the paper targets: (1) it realigns only individual exons (one exon at a time), and (2) it is slow and memory-hungry, and it mis-aligns splice sites that have shifted a larger distance into the intron or exon.

5. Method / approach

  • Model. CESAR 2.0 is a Hidden Markov Model that captures the reading frame and splice-site annotation of the reference exon; it uses the Viterbi algorithm to realign the exon sequence against the query genomic locus.
  • Gene mode (new). 2.0 implements a “gene mode” that aligns ALL exons of a coding gene to the query genomic locus jointly, rather than realigning individual exons one at a time (the only capability of the previous version).
  • Splice-site handling (new). The model assigns higher probability to frame-preserving insertions and deletions immediately downstream of the acceptor splice-site state and immediately upstream of the donor splice-site state; this is what lets it recover splice sites that have been shifted a larger distance.
  • Speed/memory re-engineering. The implementation was re-engineered to reduce runtime and memory (the paper reports the resulting factors; it does not enumerate low-level parameter values / gap penalties / probability weights in the text).
  • Benchmarks used.
    • Speed: measured across 5,977 exons; separately reported for exons >500 bp.
    • Real-scale run: mapped all 19,919 human genes (196,259 exons) to the mouse genome on a single desktop core.
    • Splice-site accuracy: in-silico evolution shifting one splice site into the intron or into the exon; sampled 10 different evolutionary distances at 0.1–1 substitutions per site (Fig. 1C).
    • Exon-boundary accuracy: counted exons where both boundaries were correctly detected.
  • Deliverable. Source code freely available at the GitHub URL above; the authors provide a workflow so users can apply CESAR 2.0.

6. Key claims / findings (atomic)

  • Average speedup of CESAR 2.0 over the previous version is 77× across 5,977 exons.
  • Memory requirement is 31 times less than the previous version.
  • For exons longer than 500 bp, the average speedup is 178×.
  • Mapping all 19,919 human genes (196,259 exons) to mouse took 7 h on a single core on a desktop computer, 126× faster than the previous implementation.
  • The previous version correctly aligned shifted splice sites in 91% of cases.
  • CESAR 2.0 gives ~33% higher accuracy on the dataset where splice sites were shifted over a large average distance of 36 bp.
  • CESAR 2.0 correctly detected both boundaries of 1,396 additional exons, raising precision of correctly annotated exons from 95.9% to 96.3%.
  • Example (Fig. 1D): for a 30 bp acceptor shift in cow in the MAMSTR gene, only CESAR 2.0 (not the previous version) identifies the shifted splice site and thus the correct exon start.
  • The gene mode aligns all exons of a gene jointly to the query locus; the previous version could realign only individual exons.

7. Load-bearing statements — OWN-WORDS MODE (source is All-Rights-Reserved)

License is ”© The Author 2017. Published by Oxford University Press. All rights reserved.” (Standard Journals Publication Model) — NOT open-access / not Creative Commons. Therefore no verbatim prose quotes; the statements below are paraphrases. Functional strings (numeric factors, counts, thresholds, the repo URL) are reproduced verbatim as facts.

  • CESAR is a coding-exon-structure-aware realigner; its HMM encodes the reference exon’s reading frame and splice-site annotation and realigns via Viterbi. (own-words)
  • The new gene mode aligns all exons of a coding gene to the query genomic locus jointly; the prior version realigned only individual exons. (own-words)
  • Headline factors (verbatim as facts): 77× faster, 31× less memory, 178× for exons

    500 bp, 126× faster on the 19,919-gene / 196,259-exon human→mouse run (7 h, single core), previous splice-site accuracy 91%, ~33% higher accuracy at 36 bp average shift, precision 95.9% → 96.3%, 1,396 additional exons with both boundaries correct, 30 bp acceptor shift in the cow MAMSTR example, in-silico evolution 0.1–1 substitutions per site, 10 evolutionary distances.

  • Repo (verbatim): https://github.com/hillerlab/CESAR2.0

8. Stated scope, assumptions, limitations

  • CESAR realigns exon sequences that are already located via a whole-genome alignment — it is a realignment/projection step operating on aligned loci, not a de-novo gene finder.
  • Benchmarks are reference→query projections (human→mouse used for the real-scale timing) and in-silico-evolved sequences for the splice-site-shift accuracy test.
  • The paper does not state an evolutionary-divergence limit (see §9, §12).

9. Failure modes / invalidity patterns

  • Splice-site shifts relative to the reference are the named hard case: when a splice site has moved a larger distance into the intron or exon, the previous version fails to align it correctly (91% accuracy overall; worse at large shifts). CESAR 2.0’s downstream-of-acceptor / upstream-of-donor indel-probability handling is the fix; the residual error is the un-recovered fraction (2.0 improves ~33% at 36 bp average shift but the paper does not claim 100%).
  • Exon boundaries: even 2.0 correctly detects both boundaries at 96.3% precision — i.e. ~3.7% of exons still have at least one mis-detected boundary. [summarizer-inferred: complement of the stated precision; the paper reports precision, not the error rate directly.]
  • The paper does NOT name a divergence threshold beyond which projection becomes invalid, nor an identity cutoff, nor an error message / diagnostic string.

10. What the source does NOT address (confident silences)

  • No CLI invocation syntax, command form, flag names, or parameter/default values in the paper text (deferred to the GitHub repo).
  • No explicit divergence range in Myr and no percent-identity reliability window.
  • No mention of TOGA (the paper predates it / does not reference it).
  • No specific HMM numeric parameters (transition probabilities, gap penalties, the exact “higher probability” values for the frame-preserving indels near splice sites).

11. Open questions / ambiguities

  • Over what evolutionary distance does exon projection stay reliable? The in-silico test spans 0.1–1 substitutions/site but the paper draws no reliability cutoff from it.
  • What are the actual command-line and parameter defaults? Not in the paper — would require the repo / manual.

12. Guidance answers

  • What CESAR 2.0 is / does: Confirmed — HMM-based, codon/reading-frame- and splice-site-aware realigner projecting reference exons onto aligned query genomes, preserving reading frame across indels and placing splice sites. (§3–§5)
  • Gene mode (joint vs one-at-a-time) — quote requested: Under the restrictive license, own-words: 2.0’s gene mode aligns ALL exons of a coding gene jointly to the query locus; the previous version realigned only individual exons. (§5, §6)
  • Speed/accuracy vs CESAR 1.0 — exact numbers: 77× faster, 31× less memory (headline); 178× for exons >500 bp; 126× faster on the human→mouse whole-proteome run (7 h, single core); previous splice-site accuracy 91%, +~33% at 36 bp average shift; exon-boundary precision 95.9%→96.3% (+1,396 exons). “Improved splice-site alignment over larger shifts” concretely = recovering splice sites shifted a larger distance (avg 36 bp) into intron/exon, e.g. the 30 bp cow MAMSTR acceptor shift that only 2.0 recovers. (§6)
  • Divergence calibration (~150 Myr claim): The paper gives NO divergence figure in Myr and NO percent-identity reliability window. The only distance-like quantity is the in-silico range 0.1–1 substitutions/site over 10 sampled distances, from which the paper draws no reliability boundary. So the skill’s “~150 Myr (vertebrate)” number is NOT supported by this paper — treat it as convention / sourced elsewhere, not attributable here.
  • Relationship to TOGA: Not mentioned in this paper. Any CESAR-inside-TOGA linkage is external to this source. [summarizer-inferred that it is external — the paper is silent.]
  • Must-capture verbatim (CLI form / parameters): NOT present in the paper — no CLI syntax or parameter names given; only the repo URL https://github.com/hillerlab/CESAR2.0. Unrecoverable from this source.
  • Must-capture verbatim (numeric divergence/identity thresholds): None stated in the paper (see divergence answer above).