Marcolini/cross-species-translational-alignment

Cross-Species Translational Alignment — TG-GATEs + DrugMatrix × Tox21

Goal: build a training substrate for detecting subtle / pre-histopathological toxicity signatures in animal transcriptome data, with mechanism-of-toxicity labels attached. This directory contains the compound-level linkage layer: every compound that has rat in-vivo perturbation data cross-referenced to Tox21 mechanism assays via standardized chemical identifiers.

Background — the hackathon

Built at Building an AI Scientist, an AI-for-Drug-Discovery hackathon by TernaryTx, future.bio, Pluto House & Anthropic (3–5 July 2026, 50Y Soho Square, London — agenda). The brief: use agentic AI to make real progress on a drug-discovery problem over one weekend, judged on innovation, technical execution, scientific relevance, potential impact and presentation.

Our question: does rat in-vivo gene expression add anything to chemical structure when predicting Tox21 mechanism-of-toxicity outcomes? We built the full pipeline, got an encouraging first signal, then stress-tested it — and report where it held and where it didn't (see Results).

• Code: https://github.com/Bl4ckd09/Cross-Species-Translational-Alignment
• Data + results: https://huggingface.co/datasets/Marcolini/cross-species-translational-alignment
• Visual overview: open the interactive report ↗ (renders in your browser) · overview.pdf (printable, 14 pp) · HTML source. A self-contained field report: hypothesis → methods → every result → conclusions.

The deliverable: master_cohort.csv

One row per unique compound (keyed on InChIKey connectivity) across the two rat transcriptome resources, annotated with data-source flags and Tox21 mechanism labels.

column	meaning
compound_name	display name (TG-GATEs name preferred, else DrugMatrix)
inchikey	full standardized InChIKey
connectivity	first 14 chars — the salt/stereo-insensitive join key
in_tggates / tggates_liver / tggates_kidney	present in Open TG-GATEs (rat in vivo)
in_drugmatrix	present in DrugMatrix (rat, multi-tissue)
cross_platform_replicate	in both transcriptome sources — use to separate biological signal from batch effects
in_tox21	has Tox21 assay data
tox21_assays_labelled / tox21_assays_active	of 12 assays, how many are measured / positive
NR-AR … SR-p53	the 12 Tox21 mechanism assay calls (0/1/blank)

Headline numbers

	compounds
Unique compounds (TG-GATEs ∪ DrugMatrix)	672
— in TG-GATEs	160
— in DrugMatrix	624
— cross-platform replicates (both)	112
With Tox21 mechanism labels	613
TG-GATEs liver + Tox21	141

Tox21 coverage on shared compounds is dense (~10 of 12 assays measured per compound), so this is not a missing-data swamp.

Tox21 mechanism assays (the 12 labels)

Nuclear-receptor panel: NR-AR, NR-AR-LBD, NR-AhR, NR-Aromatase, NR-ER, NR-ER-LBD, NR-PPAR-gamma. Stress-response panel: SR-ARE (Nrf2 oxidative stress), SR-ATAD5 (genotoxicity), SR-HSE (heat-shock/proteotoxic), SR-MMP (mitochondrial membrane potential), SR-p53 (DNA damage).

Caveat: Tox21 assays are human cell-based in vitro; the transcriptomes are rat in vivo. Treat Tox21 as an orthogonal mechanism prior, not as ground truth about the rat.

Results — does gene expression add to structure?

We ran the controlled comparison (identical everything except the feature block: structure only / expression only / fusion), leakage-safe — repeated stratified CV, every data-dependent transform (PCA, scalers, ComBat) fit on the training fold only — then stress-tested the headline four ways. The headline held only under a linear head at N=177; pooling a second dataset revealed the real bottleneck is cross-dataset comparability, not sample size.

Stage	Setup	Fusion macro	SR ΔAUC	ComBat r	SR-vs-NR
First run	N=177, DrugMatrix liver, logistic head	0.766 (vs 0.757)	+0.025	—	p=0.074
Baseline 2	N=177, GBM head (chemprop stand-in)	0.723 (−0.010)	−0.016	—	—
Richer structure	N=177, ECFP-counts + physchem + L2 (L1 worse)	0.776 (vs 0.763)	+0.019	—	—
Multitask MLP	N=177, shared-trunk net, 12 heads (torch)	0.680 (vs 0.622)	+0.060 †	—	—
Plan A · single-dose	N=256, +TG-GATEs hours (mismatched), ComBat	0.752	+0.001	0.38	p=0.38
Plan A · repeat-dose	N=256, +TG-GATEs days (time-matched), ComBat	0.756	+0.013	0.44	p=0.27

† The MLP's ΔAUC is not SR-specific (NR +0.057 ≈ SR +0.060) — the net adds expression roughly uniformly; only the linear head produces the clean SR>NR split. See RESULTS.md §5c–5d.

At N=177 with a linear head, fusion adds a small benefit concentrated in the stress-response (SR) assays (SR-p53, SR-MMP, PPAR-γ) and neutral-to-negative on receptor-binding endpoints — "expression sees stress programs that structure can't." That signal reverses under a stronger nonlinear head (Baseline 2, gradient-boosted trees, which overfits at N=177) and washes out when a mismatched second source is pooled (single-dose, exposure = hours: SR +0.025 → +0.001, cross-dataset agreement r=0.38). But the fair, time-matched test — TG-GATEs repeat-dose (exposure = days, bracketing DrugMatrix's ≤7 d) — recovers it partway (SR → +0.013, agreement r=0.44). The recovered signal tracks the agreement: 0.38 → 0.44 ⇒ +0.001 → +0.013.

We also gave structure its best classical shot and finally ran the deferred neural net: a richer ECFP-counts + physchem structure baseline lifts structure only +0.006 (0.757 → 0.763, within the CI; an L1 sparse head is worse — N=177 can't fit 2066 sparse coefficients), and a regularised multitask MLP is worse on every arm (structure 0.622, fusion 0.680) and blurs the SR>NR split. The SR-tilted benefit survives the stronger structure baseline (fusion 0.776, SR +0.019 > NR +0.010), and both checks confirm the per-assay linear head is structure at its best here — the model under which the mechanism stays legible. N is the ceiling, not the encoder or the head (RESULTS.md §5c–5d).

Honest conclusion: the bottleneck is cross-dataset comparability, not sample size. More data didn't help; better-matched data helped partially, in proportion to how comparable it was. Even carefully harmonised, same-platform rat liver signatures for identical molecules top out at r ≈ 0.44 — a measured, in-rat preview of the animal→human translational gap. The clean N=177 result stays the primary evidence; the three-point pooling experiment (177 / hours-256 / days-256) is the supporting story. Full per-result detail and reasoning: RESULTS.md.

Per-stage numbers: results_table.csv (N=177 three-arm), baseline2.csv (logistic vs GBM head), structure_rich.csv (ECFP-counts + physchem + L1), mlp_results.csv (multitask MLP), sr_vs_nr.txt (SR-vs-NR test), pc_sweep.csv (PCA-component sweep), results_combined_singledose.csv + results_combined_repeat.csv (per-assay N=177 → N=256, hours vs days), pipeline logs in data/results/logs/. Full narrative: RESULTS.md.

Files

Data:

• master_cohort.csv — the fusion table (above)
• data/tox21_keyed.csv — 7,586 unique Tox21 compounds, standardized + keyed, with 12 assays
• data/tggates_keyed.csv — 170 TG-GATEs compounds keyed (160 small molecules; rest are biologics/mixtures with no structure)
• data/dm_keyed.csv — 641 DrugMatrix chemicals keyed (635 resolved)
• data/shared_tggates_tox21.csv, data/dm_shared_tox21.csv — per-source intersections with assay calls
• data/open_tggates_main.csv — source compound list (LSDB Archive)
• data/chemicals.Rds — source DrugMatrix treatment table (combspk/Complete-DrugMatrix)

Scripts (run from data/):

• scripts/key_tox21.py — standardize + InChIKey the Tox21 SMILES → tox21_keyed.csv
• scripts/resolve_tggates.py — TG-GATEs names → InChIKeys via PubChem → tggates_keyed.csv
• scripts/resolve_dm.py — DrugMatrix names → InChIKeys via PubChem → dm_keyed.csv
• scripts/intersect.py — print all overlap statistics
• scripts/build_master.py — assemble master_cohort.csv

Modeling & experiment scripts:

• scripts/retrieve_expression.py — parse GSE57815 → DrugMatrix liver expression matrix + manifest
• scripts/build_signatures.py — per-compound logFC signatures (treated − vehicle), configurable collapse
• scripts/structure_embed.py — SMILES → ECFP4 fingerprints (swappable featurize(); ChemBERT drop-in)
• scripts/expr_embed.py — expression → PCA latent (swappable embed(); future rat→human drop-in)
• scripts/run_experiment.py — the controlled structure / expr / fusion comparison → results_table.csv
• scripts/baseline2.py — logistic vs GBM head robustness check → baseline2.csv
• scripts/run_structure_rich.py — stronger structure baseline (ECFP-counts + physchem, L1 vs L2) → structure_rich.csv
• scripts/run_mlp.py — the deferred regularised multitask MLP (needs torch) → mlp_results.csv
• scripts/sweep_and_stats.py — PCA-component sweep + formal SR-vs-NR test → pc_sweep.csv, sr_vs_nr.txt
• scripts/rma_tggates.R — RMA-normalise the TG-GATEs liver CELs (R + affy)
• scripts/build_tggates_signatures.py — TG-GATEs logFC signatures
• scripts/combat_merge.py — ComBat-align TG-GATEs onto DrugMatrix → combined_logfc.parquet
• scripts/run_combined.py — N=256 combined comparison, N=177 vs N=256 → results_combined.csv

Reproducing

Large matrices (raw GEO, RMA output, feature parquets) are not committed — size + third-party redistribution terms. Everything regenerates from the scripts plus public downloads; the full dataset list is in DATA.md.

N=177 (DrugMatrix only) — the first run:

pip install -r requirements.txt
mkdir -p data/_raw
curl -L -o data/_raw/GSE57815_series_matrix.txt.gz \
  https://ftp.ncbi.nlm.nih.gov/geo/series/GSE57nnn/GSE57815/matrix/GSE57815_series_matrix.txt.gz
python scripts/retrieve_expression.py     # -> data/expression/drugmatrix_liver_expr.parquet
python scripts/build_signatures.py         # -> data/signatures/drugmatrix_liver_logfc.parquet + labels.csv
python scripts/run_experiment.py           # -> data/results/results_table.csv
python scripts/baseline2.py                # -> data/results/baseline2.csv       (logistic vs GBM head)
python scripts/sweep_and_stats.py          # -> data/results/pc_sweep.csv + sr_vs_nr.txt

N=256 (add TG-GATEs) — the robustness expansion. Same code, different signature file. Needs the E-MTAB-799 liver CELs + R with affy (see DATA.md):

Rscript scripts/rma_tggates.R              # RMA -> data/expression/tggates_liver_rma.tsv
python scripts/build_tggates_signatures.py # -> data/signatures/tggates_liver_logfc.parquet
python scripts/combat_merge.py             # ComBat -> data/signatures/combined_logfc.parquet + combined_labels.csv
python scripts/run_combined.py             # -> data/results/results_combined.csv (N=177 vs N=256)

Same pipeline for both N. run_experiment.py is parameterised by which signature file its config points at — drugmatrix_liver_logfc.parquet + labels.csv for N=177, combined_logfc.parquet + combined_labels.csv for N=256. The two file sets are separate: running the N=256 pipeline does not overwrite the N=177 inputs or results_table.csv, so you can re-run N=177 any time with just python scripts/run_experiment.py.

Method notes

• Standardization: RDKit Cleanup → LargestFragmentChooser (desalt) → Uncharger (neutralize) → InChI → InChIKey. Applied identically to all three sources so keys are comparable.
• Matching: on both the full InChIKey and the 14-char connectivity block. Connectivity matching recovered ~7% more real compounds (salt/stereo form mismatches, e.g. ketoconazole, rifampicin, etoposide).
• Name resolution: PubChem PUG-REST name → SMILES. Note PubChem renamed its output field to SMILES/ConnectivitySMILES (old CanonicalSMILES/ IsomericSMILES are gone). Calls go through curl because the sandbox TLS proxy uses a cert Python's ssl module does not trust.

Provenance / sources

• Open TG-GATEs — LSDB Archive: https://dbarchive.biosciencedbc.jp/en/open-tggates/
• DrugMatrix — NIEHS/NTP; compound table via https://github.com/combspk/Complete-DrugMatrix
• Tox21 — MoleculeNet (tox21.csv), EPA/NCATS Tox21 program
• PubChem PUG-REST — https://pubchem.ncbi.nlm.nih.gov/

Next steps

Comparability — not data volume or model capacity — is the bottleneck, so the highest-leverage next move is a learned rat→human translation layer (RESULTS.md §11): align rat transcriptional responses into a human-toxicity space through the swappable embed() drop-in, in three tiers — linear CORAL/OT → contrastive/domain-adversarial encoder → pathway-level translation. Falsifiable ask: lift the human-DILI expression arm from chance (0.52) toward structure (0.70), and help SR over NR. That is where the animal→human gap actually lives, and the pipeline is already built to receive the fix.

Modeling is now exhausted as a lever — every attempt to out-model N=177 was tried and confirmed the setup:

1. Structure arm — done, every direction: ChemBERT (STRUCT_KIND=chembert) is weaker than ECFP (Tox21 0.68 vs 0.76; DILI 0.59 vs 0.70, §5b); richer ECFP-counts + physchem lifts it only +0.006 and an L1 head is worse (§5c). ECFP-logistic is the strong, conservative baseline; a fine-tuned transformer is the one open modeling follow-up.
2. Model head — done: GBM overfits (§5); a regularised multitask MLP is worse on every arm and blurs SR>NR (§5d). The per-assay linear head is best at this N.
3. Better harmonisation — go beyond ComBat (the r≈0.44 ceiling), e.g. tissue/time as covariates or a learned rat→rat alignment, before pooling. (Done: repeat-dose day-matching lifted r 0.38 → 0.44.)
4. Harder target — clinical failure labels (DILIrank, DILIst, withdrawal) over Tox21 priors, where in-vivo transcriptomics may carry signal structure lacks. (Done: retrospective validation — structure wins on DILI; biology edges withdrawal, but noisily.)