DeMixPert

In a Perturb-seq screen, much of what separates a knocked-out population from control cells has nothing to do with the gene that was knocked out. A large basal-state-dependent shift appears under almost any intervention in a given cell type, and because it is large it dominates mean-squared error and differential-expression scores alike — so a model can rank well on the standard metrics while recovering almost nothing that distinguishes one target from another. DeMixPert, from the School of Computer Science and Engineering and the School of Future Technology at South China University of Technology, models that shared component separately rather than predicting it.

Each response is factored into three additive pieces: a systematic response that depends only on the basal state encoded from control-cell expression, a perturbation-specific response carrying the effect unique to the target, and a population-level variation term for the spread of individual cells. The first two form a deterministic response center; the third is treated as a distribution to be sampled from. The perturbation-specific piece is read out of frozen pretrained scGPT gene representations, which is what makes prediction possible for a target the model has never observed — the embedding supplies the biological relationships that a training example would otherwise have to.

That framing separates DeMixPert from perturbation predictors that treat the full response as one monolithic target, including graph-based GEARS, set-attention STATE and flow-matching scDFM. It is described in an August 2026 arXiv preprint, has not been peer-reviewed, and covers genetic (CRISPR) perturbation only — not chemical or drug perturbation.

#Key Features

• Three-way response decomposition: Systematic, perturbation-specific and population-level effects each get a dedicated module, so a dominant shared shift cannot mask the weaker target-specific signal.
• Unseen-target generalization from frozen embeddings: The perturbation-specific component comes from pretrained gene representations rather than observations of that target, so a trained model generates responses for held-out conditions without refitting.
• Gaussian prototype invertible network: Reusable Gaussian prototypes are reweighted per condition and pushed through a bijective coupling network, yielding a condition-specific distribution over cell-level variation instead of a single average shift.
• Interpretable prototype weights: Divergence between two conditions' prototype weights tracks the distance between their observed cell populations, showing which perturbations induce similar heterogeneity.

#Technical Details

Three encoders produce the basal-state representation, the systematic response and the perturbation-specific response; their sum is the deterministic response center, and the decoder reconstructs post-perturbation expression from that center concatenated with the basal state. Mixture weights come from a softmax MLP over the basal state and perturbation embedding, applied to K diagonal Gaussian prototypes whose means and variances are periodically re-estimated by expectation-maximization; the reference setting uses K = 8. Samples from the conditional mixture pass through stacked coupling layers, and the training-only inverse projects observed variations back into mixture space to supervise both the weights and the transformation. Training minimizes four losses jointly: response alignment, the mixture negative log-likelihood, gene-expression reconstruction, and an energy-distance term aligning predicted and observed populations.

Evaluation covers four public CRISPR datasets — Adamson (2016) and Papalexi (2021) for single-gene perturbations, Norman (2019) and Replogle (2020) for combinatorial ones — each reduced to 2,048 highly variable genes plus all perturbation-target genes, under condition-disjoint splits. Against GEARS, scGPT, GenePert, scDFM and STATE, DeMixPert reaches the highest perturbation discrimination score on all four (0.904, 0.947, 0.981 and 0.938 respectively) and the lowest energy distance on all four; its common-DEG count of 31.650 on Norman exceeds STATE's 21.006. Removing the decomposition drops centroid accuracy to 0.240 on Adamson and 0.733 on Replogle, and removing the embedding-derived perturbation-specific component cuts common DEGs from 10.733 to 8.773 and from 12.378 to 9.733.

#Applications

The intended use is prioritizing which genetic perturbations to run: given control cells and a candidate target list, the model returns a full predicted distribution of perturbed transcriptomes, including targets and target pairs with no screen data behind them. That distribution — rather than a single mean profile — supports questions about response heterogeneity, such as whether a knockout drives all cells the same way or splits them into subpopulations, and lets groups running large CRISPR screens narrow combinatorial condition spaces that are infeasible to measure exhaustively.

#Impact

DeMixPert's contribution is a design argument rather than a scaling result: that the accuracy gap on unseen perturbations comes from entangling three effects with different statistical character, and that separating them helps discrimination metrics. Its scope is bounded — the model is trained per dataset over a dataset-specific gene space, with no cross-dataset transfer experiment reported, so the demonstrated generalization runs across conditions rather than across biological systems. The authors state that source code will be released upon publication; as of the v1 preprint no code, trained weights or model card have been published, leaving the reported results unreproducible for now.