BGPT: Paper Review: Transcriptomic Mutational Profiling of Gastric Adenocarcinoma in Northern Brazil

Fuel Your Discoveries

Quick Explanation Copied

Paper being reviewed: Transcriptomic Mutational Profiling of Gastric Adenocarcinoma in Northern Brazil (preprint DOI: 10.1101/2025.09.15.676424).

Long Explanation

Transcriptomic mutational profiling in northern Brazil gastric adenocarcinoma — rigorous critique

Evidence-based review of methods, results, and key blind spots (no treatment recommendations).

What the paper claims (ground truth: what is explicitly stated)

Cohort: 102 gastric adenocarcinoma tumor tissue samples from HUJBB and Ophir Loyola Hospital (Belém, Pará, Brazil), collected July 2, 2022–July 6, 2023; informed consent and ethics approval (CAAE 47580121.9.0000.5634) are reported.
Pipeline: alignment with STAR (two-pass), variant calling with VarDict, annotation with Ensembl VEP, deleterious prediction with SIFT and PolyPhen-2, conversion to MAF and landscape plots with Maftools, and signature inference with MutationalPatterns using COSMIC v3.
Reported mutation patterns: >90,000 variants; missense mutations most frequent; SNPs predominate; most common substitution reported as C>T; median mutational burden reported as 427.5 variants/sample.
Top mutated genes (reported frequencies): FTH1 (47%), MPDU1 (41%), TXNIP (40%), followed by ARID1A, RHOA, CTNNB1, APC, CDH1, KRAS, PIK3CA, TP53 (percent values reported in text).
Signature inference: SBS3 and SBS6 plus SBS5 reported as most prevalent overall, with interpretation linking them to HRD/replication damage (“SBS3”), MMR-related instability (“SBS6”), and endogenous aging/clock-like processes (“SBS5”).
Co-occurrence: significant co-occurrences mentioned, especially RBM12-ARF1 and RPL10A-CANX (asterisks used for significance in described co-occurrence matrix).

Visuals (from explicitly stated numbers)

Note: The final chart uses only qualitative ordering from the provided text (it is not a numeric claim about proportions).

Pipeline scrutiny: where RNA-derived “mutations” can mislead

Major technical risk #1: RNA-seq variant calling ≠ DNA tumor/normal calling

The paper describes using STAR (two-pass) and VarDict for variant calling from aligned RNA-seq reads.
Because RNA editing, splicing artifacts, allele-specific expression, and mapping bias can produce apparent variants, the authors mention mitigation filters for RNA-seq biases (coverage/strand bias and “RNA editing hotspots”), but the exact impact of these filters on sensitivity vs specificity is not fully quantified in the provided text.
RNA-seq–based variant calling has been studied; one example argues that RNA variant detection can reveal allele-specific differential expression of pathogenic cancer variants, but that does not remove the need for careful controls (e.g., matched normals) to distinguish somatic calls from RNA artifacts.

Major technical risk #2: filtering choices can erase real somatic variants

The paper reports removing variants with DP<10 and excluding allele-frequency windows (45–55% and 95–100%) to mitigate heterozygous and germline homozygous variants, and retaining only PASS calls.
These thresholds may increase specificity but can systematically bias against certain zygosity states or tumor purity patterns; the excerpt provided does not include a validation experiment quantifying how many known somatic variants would be recovered under the filter scheme.

Signature inference: plausible narrative, but confidence depends on called-mutation quality

The paper uses MutationalPatterns and COSMIC v3 to infer signatures.
COSMIC signatures are widely used resources, and foundational work describes the repertoire and interpretation of mutational signatures in human cancer, including major SBS categories.
However, RNA-seq–derived mutation calls can have systematic biases (e.g., mapping/mismatch patterns) that can distort the observed trinucleotide spectrum; therefore the biological attribution (SBS3/SBS6 meaning HRD/MMR) should be treated as hypothesis-generating unless DNA-based validation (matched normal DNA and/or orthogonal confirmation) is provided.

Biological interpretation: where it is strong vs where it’s a leap

Stronger parts (with provided mechanistic anchors)

The paper links FTH1 to ferroptosis/iron homeostasis, and TXNIP to redox control, citing literature that supports mechanistic relevance of these genes in cancer biology.
The paper argues that co-occurrence patterns are statistically significant in a co-occurrence matrix and uses those to suggest coupling between transcript regulation and protein/vesicle trafficking or proteostasis stress.

Potential leap(s): functional impact vs statistical association

The pipeline excludes synonymous mutations and applies SIFT/PolyPhen to classify deleterious variants, but the excerpt does not show how often predicted deleterious variants occur in known driver contexts, nor does it show protein-level effects.
Several mechanistic interpretations rely on separate literature about gene function, but because the study does not include matched-normal DNA validation, there is an added uncertainty layer for whether the observed RNA-seq–derived variants represent true somatic events.

Co-occurrence heterogeneity: what the authors’ data support

The authors report that most genes do not mutate together and interpret limited co-occurrence as sample heterogeneity.
However, co-occurrence significance depends on multiple testing correction, sample filtering, and variant calling artifacts; the excerpt mentions Mann–Whitney U test marking significance with asterisks but does not provide full statistical tables, correction method, or the effect sizes.

Reproducibility & missing information (from what is provided)

Data availability: the excerpt states that data availability is “not provided” for this preprint (no accession numbers in the provided text).
Variant calling exactness: key thresholds are given (DP<10; allele-frequency windows removed; PASS; exclude synonymous), but the excerpt does not show full details needed to reproduce the VEP/SIFT/PolyPhen decision thresholds or the MutationalPatterns signature extraction parameters (e.g., number of signatures fitted, fitting method, or exposure to COSMIC v3 reference).
External validation: the excerpt does not report validation with an independent dataset or DNA-based confirmation.

Author review links (bespoke)

Jump to targeted author-focused critique pages on BGPT.

Feedback:

Updated: March 22, 2026