Glioma is 3.6% of Tahoe-100M: predicting held-out glioma lines rests almost entirely on the drug response shared by all lines

Research question

• Can the drug response of a held-out glioma cell line in Tahoe-100M be predicted from the other lines better than by simply averaging them?
• Is prediction for glioma lines mostly interpolation among the three glioma lines, and are glioma lines harder than other lines?

Current understanding

Our judgment: the hypothesis we set out to test ('glioma prediction in Tahoe-100M is mostly interpolation among three glioma lines and does not transfer') is only half right, so we changed the title. Tahoe-100M does contain only three conventional adherent glioma lines (A-172, H4, SW 1088; 3.57% of all cell rows; CCF-STTG1 is listed in metadata but has no cells). But in our recomputation, held-out glioma lines are not predicted from each other: the mean response of the other two glioma lines was clearly worse than the mean of the 47 non-glioma lines (official DE table: MSE +0.18, Pearson −0.21). What predicts a glioma line is the response shared by all lines: the per-drug training mean reached Pearson 0.57 (DE) / 0.65 (cell-level) on held-out glioma lines; a ridge model on baseline expression beat it consistently but only slightly (MSE −0.8% DE, −4% cell-level; CIs exclude zero); nearest-neighbour lines did not help reliably. Glioma lines are not systematically harder than other lines (A-172 ranks 24th–33rd of 48 evaluable lines by difficulty, H4 13th–14th); SW 1088 is among the hardest, which tracks its low cell count. Certainty: moderate for our computations on these data; low for anything beyond these three cell lines. Not tested: glioma stem-like cells, patient-derived models, organoids, in vivo.

Scope and how to read this page. Commentary on one public dataset with our own recomputation; it ranks no model and evaluates no deep-learning model. Public facts in one line: Tahoe-100M (Tahoe Therapeutics, formerly Vevo) profiles ~100M cells from 50 cancer lines under 379 drugs at three doses plus DMSO, with all lines pooled in each well (preprint VC0041; Cell 2026 VC0042). VC0041 · preprint VC0042

B1 — Glioma coverage is small and conventional. From obs_metadata (100,648,790 rows): A-172 2,352,722, H4 1,075,515, SW 1088 163,938 rows (3.57% together; 3.67% of cells passing the full filter); CCF-STTG1 has none. Baseline expression (a proxy derived from the DE table) shows high CD44, VIM and EGFR in all three (ranks 5–19 of 50) but essentially no SOX2, OLIG2, GFAP or PTPRZ1 (CP10k ≤ 0.04); NES is high only in A-172. These are mesenchymal-leaning adherent lines, not stem-like or OPC/NPC-like models. H4 is a neuroglioma line. Certainty: high for counts; moderate for marker levels (proxy baseline). VC0042

B2 — The three glioma lines resemble each other at baseline but not in drug response. Baseline: for each glioma line the other two are among its 1st–6th nearest of 49 lines. Response (correlation of response vectors over effective conditions, official DE table): A-172 and H4 correlate 0.52, but SW 1088's responses rank 46th of 49 for A-172 and H4. Consequently, predicting a glioma line from the other two glioma lines was worse than from the 47-line mean on both data sources (Table 2). The 'interpolation among three glioma lines' part of the hypothesis is not supported. Certainty: moderate.

B3 — The shared response dominates, and simple methods add little. On held-out glioma lines (trained on the 47 non-glioma lines) the per-drug training mean is the benchmark to beat. Ridge on baseline PCs beat it in 87% (DE) / 98% (cell-level) of conditions, but by −0.0029 MSE on a mean of 0.375 (DE) and −0.0004 on 0.0090 (cell-level); its shuffled-feature null equals the mean. Nearest-neighbour by baseline was worse on DE and slightly better at cell level. Seeing the held-out line on half the drugs (in-context kNN) gave −0.0011 MSE on effective conditions at cell level. The half-cell duplicate (Pearson 0.71 all; 0.82 effective) shows how much line-specific signal exists: on effective conditions ridge recovers about a quarter of the gap between the mean and the duplicate (−0.0007 of −0.0030). Certainty: moderate.

B4 — Glioma is not special in difficulty. In leave-one-line-out over 48 evaluable lines, ranked by training-mean Pearson (1 = hardest): A-172 33 (DE) / 24 (cell-level), H4 13 / 14, SW 1088 1 / 3. The hardest lines (NCI-H661, SW 1271, SW 1088, AN3 CA, CHP-212) are those with the fewest cells, so difficulty here mostly reflects measurement noise, not tissue of origin. The 'does not transfer' part of the hypothesis holds only in the weak sense that nothing transfers much better than the mean, for glioma and non-glioma lines alike. Certainty: moderate.

B5 — Why the mean looks strong here. In Tahoe's design every line in a condition shares one well and one plate's DMSO control, so well-level technical effects are shared between the training mean and the truth. The same drug and dose on another plate agreed with itself worse than the mean did (cross-plate replicate Pearson 0.25 vs 0.57 for the mean, DE table, 406 glioma pairs). A glioma model run in a separate experiment would not share these effects, so all numbers here are likely optimistic for real transfer. Certainty: moderate that the shared-well effect exists; its size is not separated from biology here. VC0042

What would change our view. (1) A model that beats the per-drug training mean on held-out glioma lines or organoids by a calibrated margin (e.g. closing a substantial share of the gap to a technical duplicate, with drug-level bootstrap CIs), especially when the held-out line is run in a separate experiment. (2) New data adding glioma stem-like cells or patient-derived models, e.g. the announced Tahoe/Arc/Biohub dataset (VC0158; announcement only, no glioma content confirmed). (3) Evidence that responses of glioma models cluster by tissue once noise is controlled, which our data did not show. VC0158

Evidence map

Primary studies & datasets

• VC0041 · preprint Tahoe-100M : A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling — Zhang J et al., bioRxiv, 2025
• VC0042 Tahoe-100M: Mapping drug-induced molecular phenotypes at single-cell resolution. — Zhang J et al., Cell, 2026

Perturbation & validation evidence

• VC0158 Tahoe Therapeutics, Arc Institute, and Biohub Partner to Generate the Largest Perturbation Dataset for Virtual Cell Models — Tahoe Therapeutics; Arc Institute; Biohub, Tahoe news (company announcement), 2026

3 library items cited on this page; 1 are preprints (not peer reviewed).

Entities on this page

• EGFR Genes
• Simple linear baseline Methods & concepts
• Mesenchymal-like state Cell states
• Neural-progenitor-like state Cell states
• OLIG2 Genes
• Oligodendrocyte-progenitor-like state Cell states
• Tahoe-100M Datasets

Methods and results (our own analysis)

This is an exploratory re-analysis of public data run by GlioState, not a peer-reviewed result. Outputs: tahoe_glioma/README.md · tahoe_glioma/results_de.json · tahoe_glioma/results_de_effmask.json · tahoe_glioma/results_raw.json · tahoe_glioma/cell_counts.json · tahoe_glioma/raw_summary.json · tahoe_glioma/report_de.txt · tahoe_glioma/report_raw.txt (the analysis script itself is not published on this site)

Data, code and caveats

Data: Hugging Face tahoebio/Tahoe-100M, revision 349e2a55 (2026-10-14), CC0. All 1,026 shards of the official pseudobulk DE table (~89 GB, streamed) and the 1,694 even-numbered raw expression shards (~178 GB, streamed, ≈50% of cells). Task: predict the response vector (2,000 genes) of a held-out line to one drug × dose from other lines. LOGO = train on the 47 non-glioma lines; LOO = every line held out in turn. 95% CIs = bootstrap over drugs (2,000 resamples). 'Effective' conditions = mean squared delta above the 95th percentile of a DMSO-well-vs-DMSO-well null for that line. Code and exact steps: analysis/tahoe_glioma/README.md. Runtime about 1.5 h on 8 CPUs.

Table 1. Held-out glioma lines (A-172, H4, SW 1088 pooled), trained on 47 non-glioma lines, all conditions

Reading Table 1. MSE scales differ between sources (log2FC vs log1p). Ridge beats the mean in both, consistently but by little; on the DE table restricted to effective conditions (n = 751) ridge gives −0.0061 [−0.0075, −0.0049] on a mean MSE of 0.225.

Table 2. Is it interpolation among glioma lines? Mean of the other two glioma lines minus the 47-line mean

Table 3. Difficulty rank among 48 evaluable lines (leave-one-line-out; 1 = lowest training-mean Pearson)

Sources for numbers: analysis/tahoe_glioma/results_de.json, results_de_effmask.json, results_raw.json, cell_counts.json, raw_summary.json.

What remains uncertain

Our analysis uses the official DE table (method not documented in the dataset README) and a cell-level pseudobulk from 50% of the raw shards; the DE 'baseline' is a baseMean-derived proxy; only two DMSO wells per plate make the effective-condition threshold coarse; the cell-level duplicate shares the control with the truth; ridge and nearest-neighbour are simple re-implementations with one gene panel (2,000 genes chosen on non-glioma lines). Two lines (NCI-H596, NCI-H2122) had too few cells to evaluate.

What this does not test: glioma stem-like cells, patient-derived cultures or organoids, in vivo responses, cell viability or drug sensitivity (only expression deltas), deep-learning virtual-cell models, and transfer to experiments run outside Tahoe's pooled-well design. Nothing here supports a clinical claim.

Related questions

• When papers report that a mean or linear baseline matches deep models, are they comparing the same object, with the same metric, on the same kind of perturbations?
• Which mechanisms could make a perturbation-agnostic predictor look good (shared response, noise, target gene, metric), and which papers actually test which?
• Which large perturbation datasets and profiling platforms are in the library, and what do their records state about scale and cellular system?
• Which statements are dataset-producer descriptions rather than independent evaluations?