- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper predicts pseudobulk responses (delta vs DMSO, 2,000 HVGs) to drugs on Tahoe-100M, for scaffold-held-out molecules and five fully held-out cell lines. Fig. 1C motivates it via Tanimoto similarity, shared targets and shared MoA correlate with response similarity at r <= 0.041. MechMoE runs a Transformer over gene tokens, a MolFormer drug token and a basal token. M learnable "mechanism queries" cross-attend over the output and a Top-K router mixes them into a vector c. A per-gene "local" decoder on [gene token, c] is summed with a "global" decoder that sees only the drug and basal tokens (Eq. 7-8). Training uses InfoNCE, regression and diversity losses. They test against scDCA, STATE, SAMS-VAE and BioLORD, with delta Pearson 0.397 vs 0.381 on cell lines and 0.580 vs 0.570 with PDS 0.901 vs 0.858 on molecules.
Critical strengths
- The evaluation setting is the right one. Five test cell lines are fully excluded from training and molecules are split by scaffold (L354-358). That's the regime that matters for screening.
- There's a real component ablation, and it does show the routed module helps. Without it the model falls to 0.558 Pearson and 0.851 PDS, below BioLORD. The authors also say where they lose (AUPRC, DE@100, L401-405).
Critical weaknesses
- The headline gains are small, unreplicated and measured against the wrong trivial baseline. Everything is on one dataset, with no seeds, error bars or per-cell-line results, and the delta Pearson margins over the best baseline are 0.016 (Tab. 1) and 0.010 (Tab. 2). To be fair, no baseline is close on both splits, and every learned model beats Mean on delta Pearson. But "Mean" is the global training mean, and in Tab. 1 it still ties MechMoE on MSE (0.001) and MAE (0.009), while the "substantially higher" direction accuracy (L389-390) is 0.954 vs 0.950 (Mean) and 0.949 (scDCA). The test is a per-drug mean over training cell lines (Tab. 1) and a per-cell-line mean over training drugs (Tab. 2). Such averages have matched deep models on genetic screens (Ahlmann-Eltze et al., 2025). Query cross-attention, Top-K routing and InfoNCE are standard, so these numbers have to carry the paper.
- The mechanism claims go beyond what's tested. The central hypothesis is that learned mechanisms explain responses better than predefined categories (L202-204), but no model using targets, MoA or pathways is compared and Fig. 1C is only a pairwise correlation. The module does help, but the biggest drop in Tab. 3 (0.072 Pearson, 0.115 PDS) comes from removing the global decoder, which never sees the mechanisms (Eq. 7, despite the Fig. 2 caption?). The 0.96 -0.98 Jaccard across drugs (L478-479) has no ref point; fixed queries plus gene-identity embeddings (Eq. 1, 4) may give it almost for free.
- Minor: several reference entries have wrong author lists (scGPT, scFoundation, Squair et al.).
Overall a sensible setup, but I don't yet see the model beating stratified averages or the learned slots beating predefined categories.
Questions for authors
- (experiment) How do those two stratified means (per dose, if dose is an input) score on all metrics? If MechMoE clearly beats both by more than the seed spread in Q2, I'd raise my score.
- (experiment) What's the std of delta Pearson and PDS over 3-5 seeds (MechMoE, Tab. 3 variants, best baseline)?
- (experiment) How does the global decoder alone perform, and the full model without InfoNCE? You train with a retrieval loss, and PDS (distance and candidate set unspecified) reads like a retrieval metric.
- (clarification) How were scDCA, SAMS-VAE and BioLORD adapted and tuned?
- (clarification) Is Mean's MSE 0.013 / MAE 0.001 in Tab. 2 swapped? Do Tab. 1-3 use the "corrected model input vocabulary" (L463)?