When two transcription factors (proteins that bind DNA and switch genes on) act together, the combined effect on their target is often larger than the sum of their separate effects. This super-additive behavior is called synergy. Genome-wide data seem like a natural place to find it: compare regions that carry binding sites (motifs) for both factors with regions that carry only one. That comparison is badly confounded, though. Regions carrying two motifs are not random regions. They tend to be GC-rich, dense with motifs of every kind, close to gene starts, and already active. Any of these properties can inflate the apparent joint effect without any cooperation between the proteins.
We propose treating motif-pair presence as a four-arm observational "treatment": neither motif, motif A only, motif B only, or both. For each double-motif region, we find matched regions in the other three arms with the same background properties. The match uses a generalized propensity score, which is the predicted probability of being in each arm given those properties. Synergy is then estimated as a balanced interaction contrast on this matched population. We validate the estimator on a semi-synthetic benchmark with planted ground truth and on yeast knockout data. We then use it to audit whether the motif-interaction weights learned by an interpretable sequence model reflect cooperation or background composition.
Core Claim
Synergy is an interaction, meaning a difference of differences across four conditions. Binary propensity matching, including its existing genomics port, cannot estimate it directly. matchRanges builds covariate-matched null sets of genomic ranges for a single focal set versus a background pool, balancing GC content, gene density, and accessibility. It does not handle several treatment arms or estimate transcriptional output contrasts. Matching double-motif regions against a single-motif background would compare "both" with "one" and silently absorb main effects into the synergy estimate.
Our estimand takes each region $i$, with background covariates $X_i$, outcome $Y_i$ (accessibility or expression), and arm $(a,b)\in\{0,1\}^2$. It is defined on the population where all four arms have support:
$$\tau_{AB} = \mathbb{E}\big[\,Y(1,1) - Y(1,0) - Y(0,1) + Y(0,0)\,\big],$$
where $Y(a,b)$ is the potential outcome under arm $(a,b)$. We report $\tau_{AB}$ on the raw scale (super-additive) and on the log scale (super-multiplicative). DYNAMO found that multiplicative composition of perturbation effects fits combinatorial screens 9.6% better than additive composition. Calling synergy on the additive scale alone would therefore flag many ordinary multiplicative pairs.
The claim is that, with this estimand and honest covariates, a sizable share of motif pairs that look synergistic will shrink toward zero after balancing. We expect this especially for GC-rich or CpG-associated motifs. The pairs that survive should be enriched for independently supported cooperative pairs.
Research Program
The method has three parts.
Propensity model. A multinomial generalized propensity model $e_k(X) = P(\text{arm}=k \mid X)$ is fit with gradient-boosted trees. Tree models guard against the misspecification that leaves residual imbalance in logistic models.
Matching and estimation. Each double-motif region is matched to its nearest neighbor in each of the other three arms, using the propensity vector with a caliper, a maximum allowed distance beyond which no match is made. Regions without full four-way support are dropped and the size of the retained population is reported, so the reader knows which regions the estimate describes. On the matched quadruplets we fit an augmented, doubly robust estimator. This version stays consistent if either the propensity model or a flexible outcome regression is correct. That property matters because covariate-adjusted regression is the main competitor.
Covariate rules. Covariate choice is where the biology enters, and it carries one strict rule: never condition on something the motifs could cause. When the outcome is chromatin accessibility, accessibility cannot be a covariate. The covariates are GC and CpG content, dinucleotide composition, repeat fraction, region length, distance to the nearest transcription start site, and, critically, the count of all other cisBP motifs in the region. Motif density is the confounder we expect to matter most, because double-motif regions are usually motif-rich regions. For expression outcomes in yeast, basal expression in an unperturbed reference condition is allowed as a pre-treatment covariate. Motif spacing and helical orientation are not covariates. They are secondary stratifiers: within the "both" arm, synergy is estimated by spacing bin, which addresses the spacing-grammar gap left by sequence models.
Auditing a sequence model. tiSFM initializes its convolution filters with 862 cisBP motifs and adds an explicit motif interaction layer meant to capture cooperativity. It is trained on ImmGen accessibility across 81 mouse immune cell types, with no control for sequence composition. We compare its learned pair weights against balanced $\tau_{AB}$ estimates in each lineage. We also retrain it on propensity-weighted data to test whether interaction weights for composition-linked pairs collapse. In this part, balancing is the measuring instrument and the sequence model is the object being tested.
Baselines
Naive reference. The unadjusted four-arm interaction contrast of mean outcomes, with no covariates. It shows how much of the apparent synergy is pure confounding.
Covariate-adjusted interaction regression (strongest classical competitor). This is the regression a reviewer would propose: outcome on both motif indicators, their product, and the same covariates. It is given flexible spline terms and fit with boosted trees. It uses identical covariates, data, and tuning budget. It sets the bar for whether matching adds anything.
Binary matching as currently practiced. matchRanges-style nearest-neighbor matching of double-motif regions against the pooled background, run with the same covariates and caliper. It isolates the effect of the four-arm estimand.
tiSFM interaction weights and in-silico insertion. The learned interaction-layer weights, plus in-silico insertion of motif B into sequences that carry A, scored by the trained model. The published comparison reports only that tiSFM "consistently outperformed" AI-TAC on ImmGen with about 1M versus about 3M parameters. No reported result found on Lacuna for synergy-detection accuracy of any of these methods. Beating these baselines means lower absolute bias and higher precision-recall area on planted synergy, and fewer false synergy calls among null pairs.
Binarized logic. Loregic binarizes expression and scores consistency with the 16 two-input logic gates, treating AND as synergy. On yeast it reports 871 AND-consistent triplets where deleting either factor down-regulated the target, but gives no precision figure. It is run on the same yeast data.
Ablations. We match on GC alone, drop the other-motif count, and replace propensity matching with exact binning. These isolate which covariates carry the correction.
Why not just…?
Why not just fit a regression with an interaction term and covariates? This is a real threat. With enough flexibility, regression and matching can agree. Our estimator is doubly robust, so it contains the regression as one component rather than competing against it. The advantage we expect is narrower: an explicit overlap check, meaning the method declines to estimate where some arm has no comparable regions instead of extrapolating. The semi-synthetic benchmark settles it by varying overlap. If flexible regression matches our bias and precision even under poor overlap, we will report that the contribution lies in the four-arm estimand and the covariate rules, not in matching.
Why not just stratify into GC or accessibility bins? Exact binning on one or two variables is feasible. Joint binning on GC, distance to transcription start, repeat fraction, and motif density empties most bins, which is the dimensionality problem propensity scores exist to solve. The exact-binning ablation, run with an increasing number of covariates, measures how many regions each approach retains and how much bias remains.
Why not just ask the model through in-silico motif insertion? Insertion reports the interaction the model learned. Whether that interaction was learned from cooperation or from composition is exactly the open question, so insertion cannot answer it. In the semi-synthetic setting, where true synergy is zero for most pairs, we test directly whether tiSFM trained on confounded data plants spurious interactions that insertion then "discovers."
Why not just use Boolean gate calls? Binarization discards dose, and Loregic does not adjust for basal expression or shared regulators. The yeast comparison shows whether balanced continuous estimates recover knockout-supported cooperative pairs better.
Proposed Experiments
Semi-synthetic benchmark (primary test). We keep real ImmGen regions, their real covariates, and their real motif calls, so the natural correlation between motifs and GC is preserved. We simulate outcomes as a nonlinear function of the covariates, plus main motif effects, plus planted synergy for a chosen subset of pairs, with synergy set to zero for all other pairs. Outcomes are generated on both additive and multiplicative scales. Metrics are absolute bias of $\tau_{AB}$, confidence-interval coverage, precision-recall area for synergy calls, and false-positive rate among null pairs involving CpG-rich motifs.
Yeast expression analysis. We compute balanced synergy for factor pairs on promoter sequences and test enrichment against the 871 knockout-supported AND-consistent pairs from Loregic, compared with the unadjusted and regression estimates.
Real-data ImmGen audit. We report the fraction of tiSFM's strongest interaction weights that survive balancing in each lineage. Heterodimerizing families such as the bZIPs serve as a qualitative positive-control class, motivated by the obligate bZIP73–bZIP71 pairing in rice cold adaptation.
Decisive outcome. The proposal succeeds if, on the semi-synthetic data, four-arm balancing cuts false synergy calls among null pairs relative to binary matching and to tiSFM weights, while matching or beating flexible regression under poor overlap. On the yeast data, the surviving calls should be more enriched for knockout-supported pairs. If flexible regression ties everywhere, we report that the correction comes from the estimand and the covariate rules, not from matching.
References
- matchRanges : Generating null hypothesis genomic ranges via covariate-matched sampling
- Gene-Embedding Perturbation Operators for Zero-Shot and Transferable Prediction of Transcriptional Responses
- An intrinsically interpretable neural network architecture for sequence-to-function learning
- Loregic: A Method to Characterize the Cooperative Logic of Regulatory Factors
- Early selection of bZIP73 facilitated adaptation of japonica rice to cold climates