A Transformer-Based Delta Expression Encoder for Psilocybin Transcriptional Response: Architecture, Representations, and Biological Validation
q-bio.GN · cs.LG
Quick verdict
Single-author Transformer study on public psilocybin transcriptomics; unusually honest about null results, but 69.4% headline accuracy lacks a held-out test.
Summary
The paper presents a 4-layer Transformer (delta expression encoder, 862k parameters) trained to classify per-gene differential expression status (up/down/neutral) from pseudobulk single-nucleus RNA-seq of mouse medial frontal cortex after psilocybin or ketamine (623 examples; 18 cell types; 5 drug timepoints plus baseline), using the public Liao et al. 2025 dataset. Unmeasured timepoints are sentinel-masked, and only the animal's measured timepoint contributes to the loss. The reported headline is 69.4% weighted accuracy versus ~32.8% imbalance baselines. The author reports a novel asymmetry - psilocybin-induced downregulation is more stereotyped across individuals than upregulation (Mann-Whitney U=18615, p<0.0001) - per-cell-type accuracy inversely tracking DEG burden (rho=-0.885), drug-specific attention modules (serotonin/alkaloid vs NMDA/spine genes), and a significant negative HTR2A-expression-versus-response correlation (rho=-0.7088) that contradicts simple HTR2A-gating. Biphasic temporal structure is not recovered, attributed to the between-subjects design. Strengths include unusual transparency about null and negative results, and public data, code, and checkpoints. Key limitations: the headline accuracy appears to be training-checkpoint performance without a described held-out split; perfect drug/cell-type probes are near-circular because conditioning embeddings inject those labels; the pooled variance test ignores non-independence; and several analyses use different checkpoints.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A 4-layer Transformer trained on 623 pseudobulk examples classifies DEG status at 69.4% weighted accuracy, ~2.1x above imbalance-matched baselines.
- Psilocybin-induced downregulation is more stereotyped across individuals than upregulation (Mann-Whitney U=18615.0, p<0.0001 pooled; UP:DOWN variance ratio 1.07x-1.61x across eight excitatory subtypes), claimed as novel.
- Per-cell-type accuracy inversely tracks psilocybin DEG burden (Spearman rho=-0.885, p<0.001), and baseline HTR2A expression negatively correlates with drug-response separability (rho=-0.7088, p=0.0021, n=16), contradicting a simple HTR2A-gating account.
- Attention recovers drug-specific co-regulation modules without pathway supervision (Grin2b ketamine-dominant; Fos/Bdnf and serotonin-receptor genes psilocybin-dominant).
- The model fails to recover psilocybin's biphasic temporal structure, which the author attributes mechanistically to the between-subjects training design rather than model capacity.
Strengths
Unusual transparency about negative and null results
The HTR2A-gating test is reported as significantly negative (rho=-0.7088, p=0.0021, n=16) and discussed rather than buried; the attention-DEG overlap is reported as 0/30 with hypergeometric p=1.0 (below chance expectation 1.3/0.7); the temporal-structure failure is quantified against Liao et al. Figs 5b/6c (flat r=0.897-0.916, peak-to-trough 0.019).
Complete reproducibility stack
Processed AnnData (Zenodo 10.5281/zenodo.19666128), FASTQ (SRA PRJNA1204073), code (GitHub), checkpoints (Zenodo 10.5281/zenodo.21095034), full hyperparameters, and a tiny 862,084-parameter model make retraining cheap and plausible.
Principled handling of between-subjects data
Multi-route biological validation with self-correction
Error-geometry analysis ties model behavior to biology
Weaknesses
No described held-out evaluation for the headline metric
Near-circular linear probing
Statistical treatment has dependence and multiplicity gaps
Attention-based module claims rest on near-uniform attention
Cell-type accuracy gradient partly confounded with class priors
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This single-author paper trains a 4-layer Transformer ('delta expression encoder', 862,084 parameters) to classify per-gene DE status (UP/DOWN/NEUTRAL, with a structurally excluded BASELINE class) from 510-gene x 6-timepoint pseudobulk matrices built from the public Liao et al. 2025 snRNA-seq dataset (mouse medial frontal cortex; psilocybin and ketamine; 623 examples = 588 individual-animal + 35 population-mean; 18 cell types; five drug timepoints plus a population baseline). Because the experiment is between-subjects, each animal contributes real expression at exactly one drug timepoint; unmeasured timepoints carry a -1.0 sentinel and only the measured timepoint contributes loss. The headline result is 69.4% weighted accuracy (final checkpoint) versus 32.7-32.8% imbalance baselines. Further claims: (i) a novel DOWN-more-stereotyped-than-UP asymmetry (Mann-Whitney U=18615.0, p<0.0001; UP:DOWN variance ratio 1.07x-1.61x across eight excitatory subtypes, 6/8 individually significant); (ii) a per-cell-type accuracy gradient (28.3% L2/3 IT to 99.6% endothelial) tracking DEG burden (Spearman rho=-0.885, p<0.001); (iii) a negative HTR2A-vs-separability correlation (rho=-0.7088, p=0.0021, n=16) contradicting simple HTR2A-gating; (iv) attention recovering drug-specific modules (Grin2b ketamine-dominant at 0.57x; Grin2a/Fos/Bdnf psilocybin-dominant at 1.67x/1.52x); (v) failure to recover biphasic temporal structure, attributed to the training design.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper's model actually do?
It is a small Transformer (~862,000 parameters) that reads pseudobulk single-nucleus RNA-seq profiles of mouse frontal cortex (510 genes x 6 timepoints) and predicts, for each gene, whether psilocybin or ketamine upregulated, downregulated, or left it unchanged, reaching 69.4% weighted accuracy against ~32.8% imbalance baselines.
Did the study confirm that higher HTR2A expression means a stronger psilocybin response?
No. Across 16 cell types, baseline HTR2A expression correlated negatively with drug-response separability in the model's embeddings (Spearman rho=-0.7088, p=0.0021) and showed no significant positive link to DEG counts - the opposite of a simple HTR2A-dose prediction. The author notes this does not contradict Shao et al.'s causal finding that HTR2A is necessary, since necessity does not imply monotonic dose dependence.
Is the code and data available?
Yes: processed AnnData is on Zenodo (DOI 10.5281/zenodo.19666128), raw FASTQ at SRA PRJNA1204073, model code at github.com/saijaya/delta-expression-encoder, and trained checkpoints at Zenodo DOI 10.5281/zenodo.21095034.
Why doesn't the model capture psilocybin's 1h/72h biphasic response?
Because the underlying experiment is between-subjects - each mouse is sacrificed at only one post-drug timepoint - so almost no training example contains a real temporal trajectory; only 35 population-mean examples do. The author quantifies the failure: per-gene psilocybin/ketamine score correlations stay flat at r=0.897-0.916 across all timepoints instead of dipping during the quiet 4-24h interval.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Only figure captions are present in the extracted text; the rendered figure panels could not be visually inspected, so claims about plots (e.g., Figures 3A, 5, 7) rest on captions and body text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.