Pair-Conditioned Gene Weighting over Precomputed Meta-Paths for Drug Response

Predicting whether a cancer cell line will respond to a drug is, biologically, a question about a small set of genes: the ones the drug perturbs, through its protein targets and their interaction partners, that are also active in that particular cell. Graph models of drug response try to capture this by linking drugs, genes and cell lines in one network. They face a forced choice. They can learn which genes matter through attention, a learned weight on each neighbor. That is expressive, but it is costly and prone to overfitting, so in practice it runs only on a pre-filtered "landmark" gene panel. Alternatively, they can use fixed, precomputed graph averages, which scale to any graph size but cannot adapt to the task. We propose a third option. First, compute offline and without learning how strongly each gene is reached from a drug along typed biological routes, called meta-paths. Examples are drug → target protein → interacting gene, and gene → expressed-in → cell line. Then learn weights only over a short, precomputed list of candidate genes for each drug–cell pair, and make those weights depend on both the drug and the cell line.

The central claim is narrow and testable. The information that attention actually adds in pharmacogenomics is pair-specific: which intermediate genes matter depends jointly on the drug and the cell line. Pair-specific weighting is also cheap if it is applied after propagation rather than inside it. If the claim holds, the model should match or beat heavier heterogeneous graph models when run on all measured genes. It should also beat fixed-aggregation baselines specifically on held-out drugs and held-out tissue lineages. GPU memory per training batch should not depend on how many genes the graph contains. If the claim fails, and a fixed pair-level path score does just as well, then the literature's disagreement about whether attention helps drug response is resolved in favor of simple models. That is also a useful result.

What exists and where the gap is

Drug response work splits into models that treat cell lines as omics vectors and models that build a biological network. On the vector side, DeepCDR runs a graph convolution over each drug's molecular graph and fuses it with mutation, expression and methylation sub-networks; it reports Pearson 0.923 on GDSC/CCLE IC50 regression. DeepDDS likewise confines graph operations to molecules and reduces cells to a landmark-gene expression vector. On the network side, drGAT builds a drug–cell–gene heterogeneous graph and uses masked graph-transformer attention, reporting AUROC up to 0.945 under random cross-validation. Its attention scores recover known drug–target links (36.9% of predicted links). DRExplainer uses directed convolutions on a bipartite drug–cell graph with similarity edges. ASGCL sparsifies a drug–cell graph and trains with contrastive losses. drGAT supplies our graph schema and is the closest competitor. It computes attention inside full message passing over gene neighborhoods, so it depends on gene filtering to fit in memory. It also does not track composite meta-paths.

The meta-path literature supplies the vocabulary but also a warning. HAN weights meta-path neighbors and meta-paths with two levels of attention, but it throws away the intermediate nodes on each path. MAGNN shows that encoding the intermediate nodes helps: IMDb Macro-F1 rises to 61.44% from 58.51% for HAN. MAGNN does this by enumerating path instances, whose count grows as the product of node degrees along the path. That is intractable on a dense protein-interaction graph. PHGT compresses meta-path instances into tokens for a transformer, but it still needs a coupled message-passing encoder and has not been tested on biology. Our design keeps the one intermediate node that is biologically interpretable, the gene where drug perturbation meets cell state. It collapses the rest of the multi-hop route into a precomputed diffusion score, so no path instance is ever enumerated.

The scalability literature constrains the design. The decoupled methods GRAND+, GBP and LD2 precompute powers of a fixed normalized adjacency matrix and then train an MLP, which makes memory independent of graph size. Because the operator is fixed, they cannot hold learned neighbor weights. Fixed Aggregation Features Can Rival GNNs goes further: fixed statistical aggregates matched or beat GCN, GAT and graph transformers on 12 of 14 tasks. Together with ASGCL's finding that a single-layer GCN beat GAT on GDSC/CCLE, this is the strongest evidence against our proposal, and we treat it as the null hypothesis. The relevant reading of that evidence is that attention computed per node is fragile on sparse biological graphs. Attention computed per query pair over a short, preselected candidate list is a different object. The experiments below separate the two.

Core Claim

For a drug $d$ and cell line $c$, let $s_d(g)$ be a fixed diffusion score of $d$'s curated targets onto gene $g$, computed from a few steps of propagation over a protein-interaction graph. Let $x_c(g)$ be the expression of $g$ in $c$. Let $e_g$ be a static gene embedding (pathway membership, precomputed network features). For each pair we keep a candidate set $G_K(d,c)$: the $K$ genes with the largest fixed score $s_d(g)\,|x_c(g)|$. Each candidate gene $g$ in this set receives the learned weight

$$\alpha_g(d,c)=\frac{\exp\big(q(d,c)^\top W z_g\big)}{\sum_{g'\in G_K(d,c)}\exp\big(q(d,c)^\top W z_{g'}\big)},\qquad z_g=[\,s_d(g),\,x_c(g),\,e_g\,],$$

where $q(d,c)$ is a query vector computed from the drug's molecular embedding and the cell's omics embedding, and $W$ is a learned projection. The prediction head reads $\sum_g \alpha_g z_g$ together with the drug and cell embeddings. The same construction is repeated for a small set of meta-paths: drug–target–interacting gene–cell; drug–target–pathway–gene–cell; and drug–similar drug–cell, which uses training responses only. A second, meta-path-level weight, also conditioned on the pair, combines them.

All graph propagation happens once, offline. The trainable computation per pair is $O(Kh)$ for hidden width $h$, so batch memory depends on $K$ and not on the number of genes. Unlike HAN's node-level attention, the weight on a gene changes with the cell line even for the same drug. That is the property we claim carries the biological signal.

Research Program

We will build the heterogeneous graph following drGAT's schema. Drugs are linked to curated target genes, genes are linked through a protein-interaction network, and cell lines are linked to genes through CCLE expression. The graph uses every measured gene rather than a landmark panel. Offline, we compute $s_d(g)$ for propagation depths of one to three hops and select candidate sets for every drug–cell pair. Response edges enter only the drug–similar drug–cell meta-path, and they are recomputed per fold from training pairs to prevent label leakage. We then train the pair-conditioned model.

The model is evaluated in three settings: random pairs, leave-drug-out, and leave-lineage-out (all cell lines of a held-out tissue removed). The last two are the regimes ASGCL flags as weak because the graph is sparse there. We will vary $K$, propagation depth and gene-set size to map the accuracy–memory frontier. We will also examine whether the highest-weighted genes recover known targets and resistance genes, measured the same way drGAT reports its link-match rate.

Baselines

The naive reference is an MLP on the concatenation of a drug's Morgan fingerprint and the cell's expression vector, with no graph. Beside it we report per-drug mean response, because on random splits drug identity alone explains much of the IC50 variance. Any graph model that does not clearly beat both is not using the graph.

The strongest published competitor on this exact graph is drGAT, which reports AUROC up to 0.945 and R² 0.690 under random cross-validation on GDSC. We will rerun it on our splits under two conditions. The first uses its landmark gene filtering. The second uses the full gene set for as long as memory allows, and we record where it runs out of memory.

We will also rerun DeepCDR, which reports Pearson 0.923 and AUC 0.841 in binary classification, and ASGCL, which reports AUC 96.23% on GDSC and 89.25% on CCLE. These published numbers use different data slices and splits, so they establish only that our reimplementations are sane. They are not comparison targets. We found no reported result on Lacuna for leave-drug-out or leave-lineage-out evaluation of any of these methods, so those comparisons rest entirely on our reruns.

HAN on the same meta-paths represents node-level attention that discards intermediate nodes. A fixed-aggregation model in the style of FAFs, applied to the same precomputed meta-path features, represents the "no learned weighting" hypothesis.

Two ablations isolate the key ingredient. The first replaces $\alpha_g$ with the fixed score $s_d(g)|x_c(g)|$, normalized over the candidate set. This model has identical inputs and identical scalability; only the learned, pair-dependent weighting is removed. The second keeps learned weights but conditions them on the gene alone, not on the $(d,c)$ query, which reproduces per-node attention inside our pipeline.

Fairness is enforced as follows. All models share the same folds, the same gene set where memory allows, the same features and the same hyperparameter-search budget (an equal number of trials per model, selected on validation folds). We report five seeds and peak GPU memory and wall-clock time on the same hardware. Beating the field means two things. On the random split, the proposal must be statistically indistinguishable from or better than drGAT's rerun while using no more memory than it does. On leave-drug-out and leave-lineage-out, it must improve AUPR over both the fixed-score ablation and FAF, with the gap exceeding seed variance.

Why not just…?

Why not just use fixed aggregation features and an MLP? This is the most serious threat. FAFs rivalled learned models on 12 of 14 benchmarks, and ASGCL's GCN-over-GAT result points the same way. Our fixed-score ablation is exactly this alternative on identical inputs. If it ties the full model on the held-out-drug and held-out-lineage splits, we will report that learned weighting adds nothing in this domain and recommend the fixed model. The FAF paper's own failures on tasks needing long-range, input-specific reasoning (Minesweeper, Roman-Empire) are the reason to expect pair-specific relevance to matter. We treat that as a hypothesis, not evidence.

Why not just use a single-layer GCN, as ASGCL found best? ASGCL's comparison was GCN against per-node GAT on a homogeneous projection. It says nothing about weighting that is conditioned on the query pair and restricted to a short candidate list. We include ASGCL itself, and HAN-style per-node attention in our own pipeline, so the comparison separates "attention overfits" from "per-node attention overfits".

Why not just run drGAT on landmark genes and accept the filter? Filtering is cheap and may lose nothing. We will test that by comparing our model on landmark genes against our model on all genes. If the full-gene model is no better, the scalability motivation weakens to a pure cost argument, which we will state plainly.

Why not just enumerate path instances with MAGNN? Path-instance counts grow multiplicatively with degree on a protein-interaction graph. We will run MAGNN on a subsampled gene set and record memory and time. Our claim is that the endpoint gene is the only intermediate node worth keeping, and a comparison on the subsample can check whether anything is lost by collapsing the rest.

Decisive Checks

Three results decide the proposal. The first is the gap on leave-lineage-out and leave-drug-out AUPR between the pair-conditioned model and its fixed-score twin. The second is the gap between pair-conditioned and gene-only weighting. The third is a memory curve showing flat per-batch GPU usage as the gene set grows from the landmark panel to all measured genes, set against drGAT's curve. As a secondary check, the top-weighted genes should recover curated targets at a rate at least comparable to drGAT's reported 36.9%. They should also change with the cell line for a fixed drug; if they do not, pair-conditioning is not doing the work claimed for it.