Chapter 20 — Can We Turn Toxicogenomics Into a Reasoning Problem? A Critique and a Plan

Your idea: convert the rat/human toxicogenomics matrices (gene × drug, with side info) into a text reasoning dataset, fine-tune an LLM to reason, use it for our tasks (cross-tissue translation, missing-value prediction), convert its output back to the numeric format, and benchmark against our baselines. This chapter takes that idea apart honestly — where it breaks, where it’s strong, the one reframe that makes it real, and a concrete plan if we do it.

Verdict up front: the literal “LLM predicts the numbers and beats Funk-SVD” version loses. The version that works is benchmark-first — build a mechanism-grounded reasoning dataset from DrugMatrix, probe how well current LLMs reason over it (with faithfulness checks), then fine-tune to test whether that reasoning generalizes or just memorizes. Beating the numeric baselines on our tasks is the ambitious follow-up, not the headline. Everything below is about that difference.

20.1 The idea, stated fairly

Let me restate it precisely so we’re critiquing the real thing, not a strawman.

1. Serialize our data. DrugMatrix (rat) and TG-GATEs (rat + human) are matrices of gene × drug/treatment, with side information: drug identity, dose, duration, tissue, platform. Turn each cell (or profile) into text.
2. Build reasoning traces. For each example, write a natural-language chain that connects the inputs to the outcome — ideally grounded in mechanism (pathways, Adverse Outcome Pathways).
3. Fine-tune an LLM on those traces so it learns to reason about toxicogenomics.
4. Run our tasks as reasoning. Cross-tissue translation (“given liver, reason to kidney”), missing-value / matrix completion (“given what we know, reason to the missing entry”), maybe new-drug prediction.
5. Convert back to the numeric format and evaluate against our existing baselines (Funk-SVD, ToxCompl, GenTox, mean-predictor, etc.).

It’s a clean idea, and the instinct behind it is correct. But steps 4 and 5 hide a landmine, and step 2 hides most of the work. Let me do the strong version of the criticism first — the version a skeptical reviewer would write — before I tell you why it’s still worth doing.

20.2 The instinct is right — take it seriously

Four reasons this is not a crank idea.

1. The field is pivoting to mechanism, and this rides that wave. Chapter 19 showed the whole July-2026 stack moving from labels to reasoning (CoTox, ToxReason, AOP mapping). Our four papers predict numbers and stop. An idea that adds reasoning is aligned with where the field is going, not against it.

2. Our data is already half-categorical. This is the most important supporting fact, and people forget it. DrugMatrix values are log₁₀ fold-changes, and ToxCompl already bins them into 5 categories (strong-down / down / no-change / up / strong-up), with 92% sitting in “no change” (04_TOXCOMPL.md §4.7, §4.12). ToxCompl’s headline metric is Mean F1 over those categories, not MSE. So the target we actually care about is closer to a classification problem than a dense regression — and classification of the rare up/down signals is exactly the kind of thing an LLM can do by reasoning. The data is meeting us halfway.

3. It uses interventional data. The single sharpest negative result in the field (Nature Methods 2025, 18_GENOMIC_FM_LANDSCAPE.md §18.4) found that pretraining on observational cell atlases doesn’t help perturbation prediction — only pretraining on interventional (perturbation) data helps. Our data is all interventional: someone dosed a rat/cell and measured the response. A reasoning dataset built from it is the right kind of data, by that paper’s own logic.

4. It’s an NLP paper, and that’s your edge. Dataset construction, reasoning fine-tuning, structured generation, faithfulness evaluation — this is your home turf, and it’s the exact skill the tox modelers lack (Chapter 19 §19.15, Opportunity A).

Hold all four. Now the knife.

20.3 The critique — where the literal version breaks

Break 1 — a dense high-dimensional regression is the wrong job for a language model

A cross-tissue profile is 8,565–31,042 numbers. Matrix completion fills a continuous value. An LLM asked to output a full expression vector has to emit tens of thousands of floating-point numbers as text tokens. This fails on every axis:

• Accuracy. LLMs are notoriously bad at precise numeric regression, especially over long correlated vectors. There is no reason to expect GPT-class token-by-token float generation to approach Funk-SVD’s MAE. It won’t.
• Cost and length. 30,000 numbers × several tokens each blows past context windows and costs a fortune per prediction. You cannot run this over a test set of thousands of cells with a frontier model.
• Determinism. Sampling makes the same input give different vectors. Matrix factorization is deterministic.

If the plan is “LLM outputs the numeric profile and we compare MAE to Funk-SVD,” stop. That comparison is lost before it starts. This is the same category error as feeding fold-change to scGPT’s value encoder (11_SC_FOUNDATION_MODELS.md §3): using a tool for the one thing it is worst at.

Break 2 — the “convert back to numeric” step is a lossy bottleneck, and it’s where the paper dies or lives

Steps 4→5 assume you can round-trip: text reasoning → numbers → compare. But an LLM reasoning in words does not naturally produce calibrated floats. Whatever decoding you bolt on (a regression head, a parser, a lookup) is doing the real predictive work, and now you have to ask: is the reasoning helping, or is the decoder? If a linear head on the LLM’s hidden state does the prediction, you’ve built a worse linear model with an expensive front-end. The round-trip is not a formatting detail — it is the central methodological risk.

Break 3 — the baseline wall is real and it is high

The skeptical literature (Chapter 18 §18.4) is unanimous and it is about our exact task:

• Deep perturbation models do not beat linear baselines (Nature Methods 2025).
• “One PCA still rules them all” for perturbation analysis.
• Highly-variable-gene selection beats scGPT/Geneformer zero-shot.

So the bar isn’t “beat a transformer.” It’s “beat PCA, Funk-SVD, a mean-predictor, and highly-variable-gene selection.” Those are cheap, strong, and deterministic. An LLM-reasoning pipeline that costs 10,000× more per prediction and loses to Funk-SVD on Mean-F1 is not a paper — it’s a cautionary tale (and the folder already has enough of those).

Break 4 — reasoning is not mechanism (the faithfulness trap)

Chapter 18 §18.4 and ToxReason (§19.3) both show it: an LLM will produce fluent, confident, biologically-plausible-sounding reasoning that is wrong, and it will do so while sometimes getting the label right. If we fine-tune on reasoning traces and then report the reasoning as a selling point, a reviewer will (correctly) ask: did you check the reasoning is faithful, or did you just check the final number? If we don’t have a faithfulness evaluation, the “reasoning” is decoration.

Break 5 — the quieter problems

• Contamination. Frontier LLMs may have seen DrugMatrix, TG-GATEs, and the relevant AOPs in pretraining. A zero-shot “win” could be memorization, not reasoning (Chapter 18 §18.4). We need contamination-controlled splits.
• Scale / evaluation cost. DrugMatrix is ~375,000 gene-rows × thousands of drugs. You cannot reason over all of it; you must sample, and be honest that you evaluated a slice.
• No gold reasoning. We have measured expression and known drug–target/AOP facts, but we do not have ground-truth reasoning chains for each (drug, tissue). Constructing them without injecting hallucinated “mechanism” is the hardest single part of the whole idea.

That’s the honest case against. Now the reframe that survives all five.

20.4 The reframe that makes it real

The fix is one sentence: stop asking the LLM to predict numbers, and stop competing where matrix factorization is strong. Instead, make the LLM do the thing it is uniquely good at — structured, semantic, mechanistic reasoning — and compete only in the regimes where the numeric baselines are weak or undefined.

Three moves.

Move 1 — change the output type: from floats to structured, semantic predictions

Don’t predict the vector. Predict one of these, all of which are well-posed for an LLM and all of which map cleanly to metrics we already use:

• Direction / category of the toxicologically important genes. For the target tissue, predict which of the “interesting” genes go up / down / no-change (the 5-category scheme ToxCompl already uses). This is classification, and it’s exactly the rare-signal problem our whole program cares about (04_TOXCOMPL.md §4.7).
• A differentially-expressed gene set. Predict the set of genes that move in the target — evaluated with F1 / Jaccard / rank metrics against the measured set. (Sets are LLM-native; dense vectors are not.)
• Key Events / pathways. Predict which AOP Key Events or pathways the compound triggers in the target tissue — the mechanism, scored against enrichment on the measured profile (15_FRONTIER.md F6) and against AOP-Wiki/CTD.

To compare against Funk-SVD/ToxCompl, you threshold the baselines into the same categorical/set form — which is legitimate precisely because ToxCompl already reports 5-category Mean-F1. Everyone gets scored on the same discretized target. Now the LLM is playing a game it can win, on a metric the field already accepts.

Move 2 — compete only where matrix factorization can’t

Pure matrix completion is excellent at the transductive, data-rich case (fill a hole in a well-sampled matrix). Do not fight it there — you’ll lose, and you should say so. Compete in the three regimes where MF is weak or literally undefined:

This is the same lesson as GenTox’s whole existence (06_GENTOX.md): the interesting frontier is inductive, not transductive. The LLM’s natural home is exactly where lookup tables die.

Move 3 — evaluate reasoning quality, not just the answer

Adopt ToxReason’s design (§19.3): score the mechanism with an LLM-as-judge (validated against a toxicologist on a subset) on logical consistency and biological fidelity, and check the predicted Key Events against AOP-Wiki. This turns “reasoning” from a selling point into a measured deliverable, and it’s the honest-evaluation move Chapter 18 keeps demanding.

The reframed claim of the paper is not “LLM reasoning beats Funk-SVD on MAE.” It is: “On the inductive, cross-species, and interpretability regimes — where matrix factorization is undefined or underdetermined — LLM reasoning over a mechanism-grounded serialization is competitive on the categorical metrics the field already uses, and additionally produces a faithful, checkable mechanism. Here is the honest head-to-head, including the linear baselines, and here is where each approach wins.”

That is a paper. The literal version is not.

The cleanest framing: measure first, then intervene

The reframe above says what to predict and where to compete. The safest way to stage the work is to make the first paper a benchmark-and-probe study, not a beat-the-baselines contest. Two stages:

Stage 1 — build the benchmark and probe the current state (no training). Turn DrugMatrix into mechanism-grounded reasoning items, then evaluate off-the-shelf LLMs zero/few-shot with faithful evaluation — score not just the answer but whether the reasoning chain is biologically valid (ToxReason-style judge + AOP grounding, §20.6). This answers the honest first question: how far can current LLM reasoning get on this kind of data, and is it faithful or just fluent?

Stage 2 — fine-tune, and separate generalization from memorization. Only if Stage 1 shows signal. Fine-tune, then test on splits and probes built to tell real reasoning from recall (§20.6). The payload is not “we got higher accuracy” — it is “fine-tuning did / did not teach generalization, and here is the evidence.”

The three deliverables, all of which publish regardless of who wins: (1) a mechanism-grounded toxicogenomics reasoning benchmark; (2) a faithfulness-aware characterization of what current LLMs can and can’t reason on it; (3) evidence on whether fine-tuning yields generalization or memorization. Beating Funk-SVD/ToxCompl on our tasks is the ambitious follow-up, not the first paper. This framing sidesteps the baseline wall (§20.3, Break 3) entirely: the finding is the point, not the win.

20.5 The plan, part 1 — the dataset is the real contribution

For an NLP audience, the dataset is the paper (this is how ToxReason and half of Chapter 19’s stack got published). Here’s how to build it.

Sources (all public)

• DrugMatrix (rat, NIEHS; the completed version too) — gene × drug × 8 tissues, log-fold-change, with dose/duration.
• Open TG-GATEs — rat + human, in vitro + in vivo, dose–time series. This is our rat↔human bridge and it’s the same data O’Donovan used (19 §19.10).
• Side info — drug identity → structure (SMILES/IUPAC via PubChem), targets (ChEMBL), chemical class.
• Mechanism scaffolds — AOP-Wiki (the reasoning chains), CTD (chemical–gene–pathway), and the KE-to-gene map from §19.4. These turn “gene X moved” into “Key Event Y.”

Serialization (solve the context problem)

Never serialize 30,000 genes. Serialize:

• the moved genes only — top-K by |fold-change| (the 8% that aren’t “no change”), signed. This is scFoundation’s “skip the zeros,” and it’s principled here because a fold-change zero means “didn’t move” (11_SC_FOUNDATION_MODELS.md §8). A profile becomes tens-to-low-hundreds of GENE: +1.85 lines — a few hundred tokens.
• the side info — drug, dose, duration, source tissue, species.
• the retrieved mechanism context — the pathways/Key Events the moved genes map to (via the KE-to-gene resource). This is retrieval-augmented input, exactly like CoTox (§19.2).

Reasoning-trace construction (the hardest part — do it hybrid)

We have no gold reasoning. Three ways to make traces, in increasing faithfulness:

1. Template / deterministic — from measured data + KE-to-gene map, fill a fixed scaffold: “Drug D (dose, target T) → moved genes {G} → Key Events {KE} → AOP → outcome/target-tissue prediction.” Faithful by construction, but rigid and unnatural.
2. LLM-distilled — prompt a frontier model to narrate the chain (CoTox-style). Natural, but risks hallucinated mechanism.
3. Hybrid + verified (recommended) — LLM narrates within the template’s fixed facts, then every mechanistic claim is verified against AOP-Wiki/CTD/ChEMBL; unverifiable claims are dropped or flagged. This is ToxReason’s MIE↔AO matching idea (§19.3) generalized. Faithful and natural.

The verified hybrid traces are both the fine-tuning data and, itself, a citable resource.

Splits (avoid the contamination trap)

• Scaffold / drug-family split — test drugs must be structurally novel vs training (Bemis–Murcko scaffolds), so you measure generalization, not memorization.
• Held-out tissue — for the cross-tissue task, hold out target tissues entirely (mirrors TransPlatformer’s zero-shot design, 05_TRANSPLATFORMER.md §5.5).
• Held-out species — train on rat, test on human (the headline cross-species claim).
• Contamination probe — check whether a base (un-fine-tuned) frontier LLM already “knows” the answers; report the gap.

20.6 The plan, part 2 — tasks, models, baselines, metrics

Pick ONE task for paper #1 (resist doing all of them)

The menu, best first:

• A. New-drug direction prediction (inductive). Given an unseen compound’s structure/targets + a tissue, predict the up/down/no-change categories of the key genes (or the DE gene set). Best choice — it’s where MF is undefined, so any competitive number is a clean win, and it’s GenTox’s exact setting reframed.
• B. Cross-species direction transfer (rat → human). Train on rat, predict human direction/gene-set. Highest-impact, hardest ground truth (sparse human in-vivo).
• C. Cross-tissue direction translation. Given source-tissue moved genes → target-tissue moved genes. Directly comparable to TransTissueFormer, but it’s the regime where MF is strong, so hardest to win — save it for later or use it as the “honest loss” comparison.

Recommendation: lead with A, include B as the ambitious result, mention C as the honest hard case.

Feasibility of Task B (rat → human): does the data actually support it?

Task B is the ambitious one, so it deserves a check rather than a promise. Three points, and a toy that makes them concrete (code/demo_rat_human_direction.py).

The data supports it — via TG-GATEs specifically. The whole idea needs the same drugs measured in both species. Open TG-GATEs has exactly that: an overlapping compound set run in rat (in vivo + in vitro) and human (in vitro, primary human hepatocytes). That paired design is the enabler, and it’s the same data O’Donovan already used for rat→human transfer (19_PAPER_STACK_JUL2026.md §19.10) — so this is a demonstrated setup, not a hypothetical one. DrugMatrix is rat-only (training/cross-tissue side); LINCS L1000 and DILImap add human-side data. One honest caveat: “human” here mostly means human hepatocytes in a dish, not human in-vivo.

Why it’s possible: direction is conserved even when magnitude isn’t. Orthologous genes keep their function (that is why rats are used as models), and drug mechanisms are conserved in sign — an AhR agonist induces Cyp1a1, oxidative stress switches on the Nrf2 program, in both species. And our reframe predicts direction (up/down/none), the conserved axis — not magnitude, the divergent one. That is not a lucky coincidence; it is why the reframe makes cross-species tractable.

The toy, showing the mechanism and its limit. Simulate 300 genes in 6 pathways (5 conserved, 1 “species-divergent” like metabolism/P450s), 160 drugs, with rat and human sharing mechanism but differing in magnitude scale, noise, and the divergent template:

Direction (92%) transfers far better than magnitude (0.63) — the premise holds. And the inductive test, on 50 unseen drugs (macro-F1 over −1/0/+1):

The learned model matches “copy rat” on conserved genes and wins on the divergent ones (0.55 vs 0.40) — by inferring the drug’s mechanism from the conserved part of the rat profile and applying the human-specific template — and it does so on drugs it never saw.

The honest limits (build them into the paper). (1) The real baseline is not “no change,” it is copy-rat-direction; beating that by a real margin is the bar, and that gap is the contribution. (2) The P450 / species-divergent metabolism genes are exactly the toxicologically important ones (18_GENOMIC_FM_LANDSCAPE.md §18.5), and they are where concordance and ortholog mapping are worst — so the result must be stratified (strong on conserved biology, weak on divergent), which is itself a mechanistic finding. (3) In the toy I made divergence a learnable function of mechanism; real divergence is partly irreducible, so real numbers will be lower — measuring how much is learnable vs noise is the empirical question. Net: Task B is feasible for conserved biology, against the copy-rat baseline, stratified by conservation — and that honest framing is what turns it from an overclaim into a credible result.

Models

• Fine-tuned open LLM (Qwen2.5/3, Llama-3.x, 4–14B) with LoRA — cheap, reproducible, runs on one GPU (Chapter 18 §18.8), and lets you evaluate over a real test set without frontier-API costs.
• Prompted frontier LLM (GPT-4o / Claude) — zero/few-shot, as an upper-reference and to justify distillation.
• Optional: a numeric model + LLM narrator hybrid — let Funk-SVD/ToxCompl predict the numbers, and the LLM only explains them (this is Chapter 19’s Opportunity A, and it’s the safest fallback if pure-LLM prediction underperforms).

Baselines (non-negotiable — this is the whole credibility of the paper)

Include all of: mean-predictor, highly-variable-gene selection, Funk-SVD / PCA, ToxCompl (5-category), GenTox (for the inductive task), and a co-expression baseline for any pathway/network claim. The baseline is the experiment (Chapter 18 §18.7). If we skip the linear baselines, no serious venue will believe us.

Metrics

• Prediction: Mean-F1 over the 5 categories (matches ToxCompl); DE-gene-set F1 / Jaccard; directional accuracy on the toxicologically important genes.
• Mechanism: enrichment consistency (do predicted & true profiles imply the same pathways? 15_FRONTIER.md F6); AOP Key-Event accuracy vs AOP-Wiki.
• Faithfulness: ToxReason-style LLM-judge (validated on a human-scored subset).
• Cost: report tokens/compute per prediction honestly, so the interpretability gain is weighed against the price.

Ablations that make it publishable

• reasoning trace on vs off (does the chain help the answer, or is it decoration?);
• mechanism context on vs off (does the retrieved pathway/KE context help — the CoTox result?);
• shuffled mechanism context (content vs any structure — the control we always attach);
• fine-tuned vs prompted (does training on our traces beat a frontier model cold?).

Telling generalization from memorization (the Stage-2 payload)

Accuracy alone can’t tell reasoning from recall, and this is the paper’s core measurement. Four probes:

• Hard splits — scaffold / held-out-drug / held-out-tissue / held-out-species. Does it work on structurally novel inputs, or only near-training ones?
• Contamination probe — DrugMatrix is public and old, so a base LLM may already have seen it. Test whether an un-fine-tuned model already “knows” answers, and report that gap; otherwise a “win” is just recall.
• Corruption tests — shuffle the gene names or scramble the mechanism context. A reasoning model degrades; a memorizing one keeps predicting. This is the cleanest memorization detector.
• Memorization baseline — nearest-training-example retrieval. If the fine-tuned model barely beats “copy the closest training drug,” it memorized.

20.7 What counts as success — and what honest failure looks like

Decide this before running anything (pre-registration is good science and good defense). Because paper #1 is a benchmark-and-probe study, “success” is a finding, not a leaderboard win:

• The benchmark is the floor. A mechanism-grounded toxicogenomics reasoning dataset with a faithful-evaluation protocol is a contribution on its own — this outcome is guaranteed.
• Stage 1 finding: whatever the probe shows — “current LLMs reason partially and faithfully,” or “they answer but their mechanisms are unfaithful” — is a publishable characterization (exactly what ToxReason reported for its own domain).
• Stage 2 finding (the headline): whether fine-tuning yields generalization (holds on novel drugs/tissues, survives the corruption tests) or memorization (collapses on the hard splits) — and either answer is interesting, because nobody has measured it for toxicogenomics.
• Ambitious follow-up: only if Stage 2 shows real generalization do we chase the head-to-head against Funk-SVD/ToxCompl/GenTox on the inductive and cross-species tasks. There, “competitive-not-dominant but interpretable, and winning where MF is undefined” is already a paper; losing everywhere with unfaithful reasoning is still publishable as a Nature-Methods-2025-style negative result.

All outcomes yield something, because the benchmark + the honest characterization is the durable contribution regardless of who wins — the same “both outcomes publish” discipline from Chapter 18 §18.6.

20.8 Risks and mitigations

20.9 Is it novel? Where it sits, and the venue

Prior art to cite honestly (so a reviewer doesn’t think we missed it):

• Serializing expression as text for LLMs already exists — Cell2Sentence (turning a cell into a ranked gene “sentence”) and GenePT (gene/cell embeddings from text) are the closest. We must cite these and say clearly what’s new.
• LLM tox reasoning — CoTox, ToxReason (§19.2–19.3).
• LLM + biology reasoning — BioReason (DNA FM + LLM).

What’s genuinely new here (the honest novelty statement): nobody has taken the toxicogenomics matrices (with dose/tissue/species side info), serialized them into mechanism-grounded reasoning items, characterized how well current LLMs reason over them with faithfulness checks, and measured whether fine-tuning produces generalization or memorization. The combination — a toxicogenomics reasoning benchmark + AOP-grounded serialization + a faithful probe of current LLMs + a generalization-vs-memorization analysis — is the contribution. Cell2Sentence did the serialization; it didn’t do toxicology, AOPs, faithfulness, or the memorization question.

Distinct from ToxReason (§19.3) — the closest neighbor — on the one thing that matters: ToxReason reasons molecule → organ-toxicity label and never touches expression data; this reasons over the measured transcriptomic response (our DrugMatrix/TG-GATEs data) and predicts the response in a new tissue/species/drug. We reuse ToxReason’s faithful-evaluation machinery; we do not repeat its task.

Venue fit: this is an ACL/EMNLP Findings paper (exactly where ToxReason landed), or NeurIPS Datasets & Benchmarks (the dataset framing), or a bioinformatics venue if we lead with the biology. The dataset-first framing is the safest, because the dataset is bulletproof even if the model result is “partial.”

20.10 Verdict

Bottom line. Your idea is realistic if you stage it right: make paper #1 a benchmark-first probe — build a mechanism-grounded DrugMatrix reasoning dataset, measure how well current LLMs reason over it with faithfulness checks, then fine-tune to test whether that reasoning generalizes or just memorizes. When you do predict, predict meaning (direction, gene sets, Key Events), not floats, and compete where lookup tables fail (new drugs, rat→human), not where they shine. The benchmark + the honest characterization is the durable contribution; beating Funk-SVD/ToxCompl on our tasks is the ambitious follow-up, with inductive rat→human as the most ambitious of all. And it’s distinct from ToxReason because it reasons over measured transcriptomic data, not molecule→toxicity labels.

This chapter is analysis and planning, not a result. Every “it will lose” claim traces to a specific cited finding (Nature Methods 2025, “one PCA,” Kedzierska — all in 18_GENOMIC_FM_LANDSCAPE.md §18.4); every “this can work” claim traces to a mechanism the reframe exploits (ToxCompl’s 5-category metric, GenTox’s inductive setting, the interventional-data argument). The prior-art positioning (Cell2Sentence, GenePT) should be verified against the latest versions before writing the paper’s related-work section.