From DIA-NN Output to Paper Draft: A Complete AI-Assisted Proteomics Workflow (2026)

The Honest Version of "AI Wrote My Paper"

There's a fantasy circulating in 2026: feed your DIA-NN output to an LLM, get a publication-ready manuscript back. It doesn't work that way — and the people selling that fantasy haven't actually tried to publish the result.

But there's a real, useful middle ground. AI genuinely accelerates large parts of the DIA-NN → manuscript pipeline: writing analysis code, interpreting enrichment results, drafting Methods, structuring a Discussion. It also silently fails at specific, predictable points — and if you don't know where, you'll publish something wrong.

This guide walks the full pipeline — DIA-NN quantification, downstream statistics, figure interpretation, Discussion, Conclusion, and manuscript draft — marking at each stage what AI does well, where it fails, and how to verify. It's grounded in a real cross-species ECM proteomics project where AI assistance caught a lot of grunt work but also produced two mistakes I only found by looking at the figures myself.

Related: Reproducing Park et al. 2026 — Three Iterations of a Cross-Species ECM Proteomics Pipeline covers the analysis project this workflow is drawn from.

Stage 0 — What DIA-NN Gives You

DIA-NN (Data-Independent Acquisition by Neural Networks) outputs several files. The two that matter:

• report.tsv — the precursor-level master table (every detected precursor, per run)
• report.pg_matrix.tsv — protein-group quantities (MaxLFQ), the table you usually start downstream from

Key columns you'll work with:

Column	Meaning
Protein.Group	Protein group ID (UniProt accessions)
Genes	Gene symbol(s)
PG.MaxLFQ	MaxLFQ normalized protein quantity
Q.Value / PG.Q.Value	precursor / protein-group FDR
Global.PG.Q.Value	experiment-wide protein-group q-value

AI is genuinely useful here: paste the column header and a few rows, ask "which columns do I use for label-free protein quantification and what q-value threshold is standard?" It will correctly tell you to filter Global.PG.Q.Value < 0.01 and use PG.MaxLFQ.

Where it can mislead: AI sometimes suggests filtering on the precursor Q.Value when you want the protein-group Global.PG.Q.Value. Subtle, but it changes your protein count. Verify against the DIA-NN documentation, not the LLM's memory.

Stage 1 — Quantification and QC

Loading and filtering

A standard starting point in Python:

import pandas as pd

df = pd.read_csv('report.pg_matrix.tsv', sep='\t')
# MaxLFQ columns are the sample intensity columns; metadata columns are the first few
meta_cols = ['Protein.Group', 'Protein.Ids', 'Protein.Names', 'Genes', 'First.Protein.Description']
sample_cols = [c for c in df.columns if c not in meta_cols]

# Log2 transform (MaxLFQ intensities are linear)
import numpy as np
mat = df[sample_cols].replace(0, np.nan)
logmat = np.log2(mat)

AI is good at: generating this boilerplate, suggesting log2 transformation, explaining why you replace 0 with NaN rather than keeping zeros.

The valid-value filter (where pseudocounts kill you)

This is the single most common downstream mistake, and AI does not reliably warn you about it.

If a protein is detected in all of group A and none of group B, the fold change is undefined (division by zero). The naive fix — add a tiny pseudocount — produces exploded log2FC values that dominate your volcano plot with artifacts.

The correct approach is a valid-value filter: require detection in at least k of n samples per group, and handle one-sided proteins separately as "qualitative only."

groupA = [c for c in sample_cols if 'EEM' in c]
groupB = [c for c in sample_cols if 'MAT' in c]

valid_A = logmat[groupA].notna().sum(axis=1) >= 3   # e.g. 3 of 5
valid_B = logmat[groupB].notna().sum(axis=1) >= 3   # e.g. 3 of 3
quantifiable = df[valid_A & valid_B]
qualitative_only = df[~(valid_A & valid_B)]

⚠️ Real failure I hit: an LLM-generated script used np.log2((meanA + 1e-6) / (meanB + 1e-6)). The volcano plot then had a dozen proteins at log2FC ±10 — pure artifacts from the pseudocount. The LLM didn't flag it; I caught it by looking at the volcano plot and asking why the tails were so extreme. This is the part of the job AI can't do for you: skepticism toward your own output.

Stage 2 — Differential Expression (DEP)

Standard label-free DEP workflow:

from scipy import stats
from statsmodels.stats.multitest import multipletests

results = []
for idx, row in quantifiable.iterrows():
    a = logmat.loc[idx, groupA].dropna()
    b = logmat.loc[idx, groupB].dropna()
    # Welch's t-test (does NOT assume equal variance)
    t, p = stats.ttest_ind(a, b, equal_var=False)
    log2fc = a.mean() - b.mean()
    results.append({'Genes': row['Genes'], 'log2FC': log2fc, 'p': p})

res = pd.DataFrame(results)
res['FDR'] = multipletests(res['p'], method='fdr_bh')[1]
res['DEP'] = (res['FDR'] < 0.05) & (res['log2FC'].abs() >= 1)  # FC >= 2

AI does well: generating this, explaining Welch's vs Student's t-test, explaining Benjamini-Hochberg FDR vs Bonferroni.

Where to verify:

• AI sometimes defaults to equal_var=True (Student's). For proteomics with unequal group sizes/variances, Welch's (equal_var=False) is the safer default — confirm it picked the right one.
• For small n (n=3), a moderated test like limma (R) is statistically stronger than a raw t-test because it borrows variance information across proteins. AI will write the t-test version unless you ask for limma. For n≤5, ask for limma.

Related: Statistical Test Selection Guide — t-test, limma, ANOVA covers when each test is appropriate.

Stage 3 — Figures (and Reading Them Critically)

Volcano, PCA, heatmap

AI generates publication-quality matplotlib/seaborn code quickly. A volcano plot, PCA, and clustered heatmap are essentially free now — describe the columns and ask.

The non-negotiable human step: look at every figure and ask whether it makes biological sense.

• PCA: do replicates cluster? If a sample sits alone, is it a batch effect or a real outlier? (In my cross-species data, one Matrigel sample separated on PC2 — turned out to be known batch variation, not an error, but I had to investigate.)
• Volcano: are the extreme points real biology or pseudocount artifacts (Stage 1)?
• Heatmap: does the clustering separate your groups, or is it driven by a confound?

AI cannot do this judgment. It will happily generate a beautiful figure of meaningless data.

GO / pathway enrichment interpretation

This is where AI is genuinely strong. Paste your significant gene list, ask for likely enriched biological processes, and it gives a well-organized starting interpretation. Then run the actual enrichment (clusterProfiler, g:Profiler, Enrichr) and compare — AI's guess is a hypothesis generator, the tool output is the evidence.

Related: GO Annotation and Pathway Enrichment Practical Guide.

Stage 4 — Result Interpretation

Here the division of labor becomes clear.

AI excels at:

• Summarizing what a list of 50 DEPs has in common ("these are predominantly ECM structural proteins and basement membrane components")
• Connecting your findings to known literature ("LAMB1/NID1/LAMC1/HSPG2 are the canonical basement membrane signature")
• Drafting the descriptive part of Results ("Of 285 quantified proteins, 131 were differentially abundant…")

AI fails at:

• Knowing whether your specific finding is novel or already published (it confidently cites papers that may not exist — verify every citation)
• Distinguishing a real biological signal from a technical artifact specific to your platform (e.g., Matrigel nuclear contaminants like EWSR1/RUVBL2 that look like "findings" but are known contamination)
• Judging effect sizes in clinical/biological context

Verification rule: every factual claim AI makes about prior literature, you check against PubMed yourself. LLM citation hallucination is still real in 2026.

Stage 5 — Discussion and Conclusion

AI is a strong drafting partner here, not an author.

A productive workflow:

1. Give AI your verified Results bullets + the key prior literature (that you found)
2. Ask it to draft a Discussion with structure: principal findings → comparison with prior work → mechanistic interpretation → limitations → future directions
3. Rewrite it in your voice, correcting overstatements

The single most important edit: AI systematically overstates. It will write "these findings demonstrate that…" when your data suggest something. Proteomics reviewers punish overclaiming. Downgrade confidence language everywhere:

• "demonstrates" → "suggests" / "is consistent with"
• "proves" → "supports"
• "the first to show" → delete unless you've verified it's true

Limitations section: AI writes generic limitations ("sample size was limited"). Replace with the specific ones only you know — the Matrigel batch variation, the intestine-vs-esophagus tissue mismatch, the 20% ortholog mapping rate. Specific limitations build reviewer trust; generic ones signal a thin paper.

Stage 6 — The Manuscript Draft

What AI can draft well

• Methods: highly templated. Given your DIA-NN version, FASTA, FDR thresholds, and stats, AI produces a solid Methods draft. Still verify every parameter against what you actually ran.
• Abstract: after Results and Discussion are final, AI writes a tight structured abstract quickly.
• Figure legends: fast and usually accurate if you describe the figure.
• Reference formatting: convert between citation styles.

What you must own

• The actual claims and their calibration — this is your scientific responsibility, not delegable
• Every citation's existence and relevance — check each in PubMed
• Statistics reported — numbers must match your actual output exactly
• Novelty framing — only you can verify what's genuinely new

2026 journal disclosure requirements

Most major publishers (Nature, Science, Cell Press, Elsevier, ICMJE) now require disclosure of AI use in manuscript preparation. As of 2026, the standard rules:

• AI tools cannot be listed as authors (they can't take responsibility)
• Disclose AI assistance in Methods or Acknowledgements (e.g., "Large language models were used to assist with code generation and manuscript editing; all outputs were verified by the authors.")
• You remain fully responsible for accuracy, including anything AI drafted
• Some journals prohibit AI-generated images/figures without disclosure

Check your target journal's specific policy before submission — they vary and they change.

A Realistic Time Breakdown

From a real cross-species project, roughly how AI shifted the time distribution:

Stage	Without AI	With AI	AI's role
Analysis code	2-3 days	0.5-1 day	Strong — boilerplate, debugging
Figure generation	1 day	2-3 hours	Strong — but you interpret
Literature context	2-3 days	1-2 days	Mixed — hypotheses yes, citations verify
Discussion draft	2-3 days	1 day	Strong drafting, heavy human editing
Methods/Abstract	1 day	2-3 hours	Strong
Verification/judgment	(folded in)	+1-2 days	None — this is all you

Net: AI roughly halves the time to a first complete draft. But it adds a verification burden — you spend new time checking AI output that you wouldn't have spent writing it yourself. The savings are real but smaller than the hype.

Honest Prompt Patterns That Worked

A few prompts that were genuinely productive (paraphrased):

• "Here are my DIA-NN pg_matrix columns and 5 example rows. Write Python to load, log2-transform, and apply a valid-value filter requiring ≥3 of 5 detection per group. Do not use pseudocounts." (Explicitly forbidding the pseudocount avoids the artifact.)
• "Here are 50 significant up-regulated genes. List the most likely enriched GO Biological Processes as hypotheses I should test with clusterProfiler — do not claim these are confirmed."
• "Draft a Discussion limitations paragraph. Here are the specific limitations: [list]. Do not add generic limitations I haven't given you."
• "Rewrite this Discussion paragraph to downgrade overclaiming — replace 'demonstrates' with 'suggests' where the data is correlational."

The pattern: constrain the AI, forbid known failure modes, and never ask it for facts you haven't verified.

FAQ

Q: Can AI replace a bioinformatician for proteomics analysis? No. It replaces the typing, not the judgment. The hard parts — choosing the right filter, spotting artifacts, calibrating claims — are exactly where AI fails. It's a power tool, not an autopilot.

Q: Which LLM is best for this in 2026? For code: any frontier model handles pandas/scipy well. For literature: be cautious with all of them — citation hallucination persists. For Methods/Discussion drafting: frontier models are strong. The model matters less than your verification discipline.

Q: Is it ethical to use AI to write a paper? Using AI to draft and edit, with full human verification and disclosure, is accepted by major journals in 2026. Submitting unverified AI output as your own work is not — and it's how fake citations end up in published papers.

Q: How do I avoid AI-hallucinated citations? Never let AI generate a citation list. Find your references yourself (PubMed, Google Scholar), give them to the AI, and have it format/insert them. Verify every DOI.

Q: Can AI do the statistics for me? It can write the code and explain the methods. It cannot decide whether your experimental design supports the test, or whether n=3 is enough. For small-n proteomics, prefer limma over raw t-tests — and confirm the AI used the test you intended.

Q: Will reviewers reject a paper that used AI? Not if you disclose appropriately and the science is sound. They will reject overclaiming, hallucinated citations, and artifacts — all of which are more likely if you trust AI uncritically. Used carefully, AI doesn't increase rejection risk.

Closing — The Division of Labor

The realistic 2026 picture of AI in the DIA-NN → manuscript pipeline:

1. Code generation: AI strong, verify logic (especially filters)
2. Statistics: AI writes it, you choose the right test (limma for small n)
3. Figures: AI generates, you interpret critically
4. Enrichment interpretation: AI hypothesizes, the tool confirms
5. Literature: AI suggests, you verify every citation
6. Discussion/Methods/Abstract: AI drafts, you calibrate claims and own the science
7. Judgment, skepticism, novelty: 100% human

AI roughly halves time-to-draft and shifts your effort from typing to verifying. The researchers who benefit most are the ones who already know what good analysis looks like — because they can catch the failures. AI amplifies expertise; it doesn't substitute for it.

If you're starting a DIA-NN project, use AI aggressively for the mechanical parts and ruthlessly verify the judgment parts. That's the workflow that actually gets to a defensible paper.

---

Related posts:

• Reproducing Park et al. 2026 — Cross-Species ECM Proteomics, Three Iterations
• LC-MS/MS Proteomics: Complete Workflow Guide 2026
• Statistical Test Selection Guide — t-test, limma, ANOVA
• GO Annotation and Pathway Enrichment Practical Guide

References:

• Demichev, V. et al. (2020). DIA-NN: neural networks and interference correction enable deep proteome coverage in high throughput. Nature Methods, 17, 41-44.
• Ritchie, M. E. et al. (2015). limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Research, 43, e47.
• ICMJE (2026). Recommendations on the use of AI-assisted technologies in manuscript preparation.
• Nature Portfolio (2026). Editorial policies on AI tools and authorship.