Aligning the Off-Target Shadow: Cross-Tier Interference Alignment for Drug-Combination Transcriptomics

When a cell is treated with a small-molecule inhibitor, the resulting RNA sequencing profile contains two things at once: the intended modulation of the drug's annotated target pathway, and a cloud of unintended "off-target" transcriptional changes that spread across unrelated programs. When two or more drugs are combined, these clouds overlap, and the measured combination signature becomes an entangled mixture that standard differential-expression analysis cannot untangle. Wireless engineering solved a structurally identical problem years ago: interference alignment coordinates transmitters so that all unwanted signals arrive at a receiver collapsed into a single low-dimensional subspace — an "interference shadow" — which the receiver then projects away, leaving clean dimensions for the wanted signals. This proposal imports that machinery, specifically its cross-tier form and its convex rank-minimization formulation, into pharmacological perturbation transcriptomics.

The core move is a change of casting. Each drug in a combination plays the role of a transmitter with two tiers: an on-target tier (the annotated mechanism of action, akin to the macrocell primary tier) and an off-target tier (uncoordinated secondary effects, akin to dense femtocells). We propose to learn a shared projection that aligns every drug's off-target tier into one common low-rank nuisance subspace while keeping each on-target tier full-dimensional and decodable — exactly the design principle of Interference Alignment in a Poisson Field of MIMO Femtocells, where secondary femtocell interference is deliberately steered into the subspace already occupied by macrocell interference so that one zero-forcing filter removes both. Applied to RNA-seq, this yields the first method that both predicts unseen drug-combination transcriptomic signatures and provides a single filter that strips multi-drug off-target leakage across cell lines.

Core Claim

The off-target transcriptional responses of chemically distinct compounds occupy overlapping, approximately low-rank subspaces, and these subspaces can be jointly compressed by a convex program. We claim that a transcriptomic adaptation of rank-constrained rank minimization — the formulation Interference Alignment as a Rank Constrained Rank Minimization. introduced for multi-antenna channels — will (a) predict combination signatures better than additive composition, retrieval similarity, or graph-based perturbation models, and (b) transfer across cell lines better than unconstrained deep models, because the nuisance subspace it isolates is shared rather than relearned per context. This fills the sharpest gap in the file: every state-of-the-art perturbation model — What Makes a Representation Good for Single-Cell Perturbation Prediction?, DeMixPert: Decomposed Response Modeling with Gaussian Mixtures for OOD Single-Cell Perturbation Prediction, Gene-Embedding Perturbation Operators for Zero-Shot and Transferable Prediction of Transcriptional Responses, and Predicting transcriptional outcomes of novel multigene perturbations with GEARS — is built and evaluated on CRISPR perturbations, where a perturbation is a clean binary gene erasure. None handles a small molecule that hits multiple enzymatic tiers at once with dose-dependent, sometimes biphasic behavior, as documented for IB-MECA in A3Adenosine Receptor Activation Triggers Phosphorylation of Protein Kinase B and Protects Rat Basophilic Leukemia 2H3 Mast Cells from Apoptosis, where 1–10 µM protects against apoptosis and concentrations above 10 µM kill the same cells.

Research Program

The method has three stages. First, tier assignment: for each compound in the Connectivity Map L1000 corpus (over 1.3 million perturbational profiles of small molecules and knockouts across cell lines), we split its measured signature into an on-target tier — the response attributable to the compound's annotated mechanism of action, using the Drug Repurposing Hub annotations that Leak Proof CMap; a framework for training and evaluation of cell line agnostic L1000 similarity methods already uses for mechanism-of-action splitting — and an off-target residual. This mirrors the logic of comparative pharmacological inhibition: Homoharringtonine regulates the alternative splicing of Bcl-x and caspase 9 through a protein phosphatase 1-dependent mechanism separated PP1 from PP2A effects by exploiting the differing specificity of calyculin A versus okadaic acid; we do the same computationally, using annotated specificity rather than matched inhibitor pairs.

Second, alignment. Write $\Delta_d^{\mathrm{on}} \in \mathbb{R}^{g \times n_d}$ for the on-target response matrix of drug $d$ ($g$ genes, $n_d$ replicate-by-concentration profiles), and let $M_{\mathrm{off}}$ be the column-stack of all drugs' off-target response matrices. We learn a shared receive filter $U \in \mathbb{R}^{g \times k}$ by

$$\min_{U}\; \| U^\top M_{\mathrm{off}} \|_* \quad \text{subject to} \quad \lambda_{\min}\!\big( U^\top \Delta_d^{\mathrm{on}} (\Delta_d^{\mathrm{on}})^\top U \big) \ge \epsilon \;\text{ for every } d,$$

where $\|\cdot\|_*$ is the nuclear norm (the sum of singular values, the standard convex stand-in for rank, as in Interference Alignment as a Rank Constrained Rank Minimization.) and $\lambda_{\min}$ is the smallest eigenvalue, which enforces that each on-target tier stays decodable after filtering. The subproblems are solved by the same alternating convex scheme that paper used for precoders and decoders. The output is a "clean subspace" — the orthogonal complement of the learned nuisance shadow — in which on-target signals live.

Third, composition. To predict the combination of drugs $d_1$ and $d_2$, we project both single-agent signatures into the clean subspace and compose the survivors two ways: by adding the projected on-target vectors (the rule implicit in GEARS and DeMixPert) and by composing learned low-rank operators multiplicatively, following Gene-Embedding Perturbation Operators for Zero-Shot and Transferable Prediction of Transcriptional Responses, which found operator multiplication 9.6% more accurate than addition for genetic double perturbations. Which rule wins for pharmacological combinations is itself an open question the proposal settles empirically. One constraint disciplines the design: Geometry of the 3-user MIMO interference channel showed that alignment feasibility is governed by algebraic structure, not by counting parameters, and exact alignment fails when interferers exceed the receive dimension. The transcriptomic analog is unavoidable — with roughly twenty thousand genes and far fewer replicates, exact zero-forcing is impossible — so we will report residual leakage (off-target energy surviving the filter) as a first-class diagnostic rather than claim exact removal.

Baselines

The do-nothing reference is the Mean Baseline that simply predicts the unperturbed average cell state, which PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction showed to be a surprisingly hard bar: the best zero-shot foundation model, scGPT, scored an AUSPC of roughly 0.045, barely beating it. Alongside it sits the naive combiner, $\Delta_{d_1} + \Delta_{d_2}$, which is what most practitioners would actually do today. Both run on identical held-out genes and combinations with no tuning, so any structured method must clear them by a margin that a paired test supports.

The strongest existing competitor is GEARS, the graph neural simulator of Predicting transcriptional outcomes of novel multigene perturbations with GEARS, which reported a 53% improvement over prior models on zero-shot combinatorial prediction and a 30–50% MSE improvement over CPA for the most relevant genes; on the Norman dataset it achieved an AUSPC of roughly 0.008 in the PertEval-scFM benchmark, an order of magnitude better than any foundation model. We will retrain GEARS with compounds (rather than genes) as perturbation nodes on the same L1000 profiles, cell lines, and split structure we use, with a matched tuning budget, since no reported result found on Lacuna for combination-signature prediction on CMap L1000 by any method — the benchmark itself has to be built. For cross-cell-line transfer, the competitor is the TripletLoss metric-learning model of Leak Proof CMap, which achieved split-averaged uniqueness AUROC of 0.916 ± 0.011 across 25 leakage-controlled splits, evaluated under its exact split protocol.

The key ablation replaces the learned cross-tier alignment with a plain PCA projection of the same dimension: Leak Proof CMap found EuclideanPCA the strongest method on distinctness, and Benchmarking Transcriptomics Foundation Models for Perturbation Analysis : one PCA still rules them all found PCA competitive with 30M+ parameter transformers. If this ablation matches the full method on combination prediction, the alignment machinery is not earning its keep, and we will say so. A second ablation removes the rank constraint (training $U$ by plain reconstruction loss) to isolate whether the nuclear norm, specifically, is the active ingredient, echoing the finding of DeMixPert that subtracting the systematic basal component lifted cluster purity from 0.308 to 0.769 on Replogle — decomposition matters, but nobody has tested whether the convex rank-minimizing form of decomposition matters.

Why not just…?

Why not just run DESeq2 and PCA? This is the most serious objection, because the evidence partly supports it: PCA dominates distinctness in Leak Proof CMap and preserves structural integrity better than foundation models in Benchmarking Transcriptomics Foundation Models for Perturbation Analysis. But PCA is unsupervised over the pooled matrix: it has no notion of which drug a directional component belongs to, and therefore no composition rule for predicting an unseen combination — it can describe what happened, not what will happen when two clouds overlap. The decisive check is the PCA ablation above, run at matched dimension: if PCA-projected composition predicts held-out combinations as well as cross-tier-aligned composition, the proposal's claim fails, and the honest conclusion would be that a simple subspace suffices for description while prediction needs something else.

Why not just add the single-drug profiles? Because the pharmacological record is full of combinations whose effect is not the sum of parts: aprepitant plus itraconazole in Aprepitant, an antiemetic agent, interferes with metal ion homeostasis of Candida auris and displays potent synergistic interactions with azole drugs reduced minimum inhibitory concentration eight-fold, drove a roughly 600% rise in intracellular ROS, and reached a synergy index (ΣFICI) as low as 0.08 — none of which is recoverable from either single-agent profile by addition. On the modeling side, DYNAMO found multiplicative operator composition 9.6% better than addition for genetic pairs. Our method should beat the additive baseline specifically where the two drugs share off-target tiers, and we will test that prediction directly: performance gain should correlate with measured overlap of the off-target subspaces, which is a falsifiable signature of the alignment mechanism rather than generic model capacity.

Why not just fine-tune scGPT or Geneformer? Because two independent benchmarks show foundation models actively suppress the signal we need: PertEval-scFM found scGPT barely beating the Mean Baseline, and What Makes a Representation Good for Single-Cell Perturbation Prediction? showed their masked-language pretraining prioritizes housekeeping programs, leaving PerturbedVAE at an RMSE of 0.4474 against 0.56–0.61 for the foundation models. Fine-tuning does not rescue this: Benchmarking Transcriptomics Foundation Models for Perturbation Analysis reports that fine-tuning scGPT on evaluation data degraded zero-shot biological retrieval. The one caveat we take seriously is that these benchmarks cover genetic perturbations; chemical perturbation prediction by foundation models is under-measured, so we will run the best fine-tuned foundation model as a reference line rather than dismiss it from theory alone.

Proposed Experiments

The first experiment builds the benchmark: tier-labeled L1000 profiles with compounds grouped by Drug Repurposing Hub mechanism of action, following the leakage-controlled split design of Leak Proof CMap, holding out cell lines and mechanisms of action. The decisive ground-truth combination test in the corpus is GEO GSE148749, the Candida auris bulk RNA-seq from the aprepitant study, which measured aprepitant alone, itraconazole alone, and the combination. We will fit the alignment on the two single-agent arms plus the L1000 corpus and predict the combination arm, scoring RMSE and Pearson correlation on held-out genes and, more importantly, directional recovery of the reported program (downregulation of the metal transport genes FTR1, ZRT2, and CTR1; upregulation of the ergosterol genes ERG1, ERG2, and ERG10). The additive baseline and GEARS-on-compounds run under identical conditions. The second experiment evaluates transfer on the 25 Leak Proof CMap splits using their compactness, distinctness, and uniqueness metrics, comparing against TripletLoss's reported 0.916 ± 0.011 uniqueness AUROC and EuclideanPCA's distinctness dominance. The third experiment stress-tests dose: following the biphasic behavior documented in the A3 adenosine receptor study, we will model dose as a deformation of the tier subspaces and test whether the aligned nuisance subspace predicts where non-monotonic sign flips occur. If the alignment method does not beat the additive and PCA baselines on the GSE148749 combination task and at least match the linear leaders on the L1000 transfer splits, the hypothesis is wrong, and we will report the residual leakage analysis that explains why.