TxPert

TxPert predicts how a cell's transcriptome changes in response to a genetic perturbation the model has never seen. The combinatorial space is hopeless to enumerate experimentally — roughly 20,000 genes, far more pairs, multiplied by every cell type and tissue context — so the value of a perturbation model lies entirely in whether it generalizes out of distribution. TxPert is evaluated on three such settings: unseen single perturbations within a trained cell line, unseen double perturbations, and perturbations in cell lines held out of training altogether.

Its central design decision is to route generalization through prior biological knowledge. Rather than learning gene relationships from expression data alone, TxPert encodes each perturbation using graph neural networks operating over curated gene-gene networks, so a gene never perturbed in training still receives an informative embedding from its neighbors. A separate basal-state encoder embeds batch-matched control profiles to capture cell type, protocol, and batch context; the two embeddings are combined and decoded into a predicted log-transformed expression profile.

The model was developed by Frederik Wenkel, Alisandra Denton and colleagues at Valence Labs, the AI research arm of Recursion, and posted to arXiv in May 2025. The paper is candid about the state of its field: independent benchmarking has repeatedly failed to reproduce headline claims from transcriptomic perturbation models, and simple mean or additive baselines often beat published methods. TxPert therefore devotes substantial space to evaluation standards, including strong non-learned baselines and an analysis showing that Pearson correlation on delta profiles is the metric that tracks retrieval quality.

#Key Features

• Multi-graph perturbation encoding: The Exphormer-MG variant learns over a unified supra-adjacency representation of several knowledge graphs at once, letting information flow between STRING, Gene Ontology, and screen-derived networks in one message-passing stack.
• Latent transfer architecture: Perturbation embeddings are applied to a basal-state embedding of matched controls rather than to raw counts, keeping batch and cell-context effects separate from the perturbation signal.
• Three out-of-distribution regimes: One framework covers unseen singles, unseen doubles built from seen constituents, and zero-shot transfer to a new cell line.
• Honest baselines: Every comparison includes a non-learned "general baseline" combining mean and additive effects, which the authors show outperforms several published models.
• Graph ablation as evidence: Progressively rewiring STRING edges degrades accuracy monotonically, showing the model exploits real biological topology rather than graph statistics.

#Technical Details

Two module families are tuned per task. For unseen single perturbations, the best configuration is Exphormer-MG, a graph transformer integrating four sources: STRING, Gene Ontology, and two proprietary Recursion relationship datasets — PxMap, derived from large-scale phenomics microscopy screens, and TxMap, derived from single-cell transcriptomic screens. For double perturbations a multilayer graph attention network over GO, PxMap, and TxMap performs best. Benchmarks use genome-scale Perturb-seq data across four cell lines — K562, RPE1, HepG2, and Jurkat, each with more than 2,000 perturbations — from Replogle et al. (2022) and Nadig et al. (2025), plus the Adamson et al. (2016) combinatorial screen.

TxPert outperforms GEARS, scLAMBDA, and the general baseline across the three regimes, approaching estimated experimental reproducibility on some metrics for unseen singles; GEARS itself falls below the non-learned baseline in that setting. Incrementally adding graphs to Exphormer improves performance monotonically, peaking with all four. Ablation shows the Exphormer configuration favors STRING while multi-hop message passing benefits from the denser screen-derived graphs. Accuracy correlates with perturbation effect size (Spearman ρ = 0.52 against the number of differentially expressed genes) and with curated graph degree (ρ = 0.33), but not with how well studied the target gene is. One failure mode is explicit: the architecture does not learn the characteristic downregulation of the perturbed gene itself.

#Applications

The intended use is triage in functional genomics and target discovery — nominating which perturbations, combinations, or cell contexts are worth running on the bench, and interpreting screens that are inevitably sparse relative to the space they sample. The paper's worked example predicts knockdown of TSR2, a ribosome maturation factor implicated in Diamond-Blackfan anemia, recovering both local effects on graph neighbors and transcriptome-wide functional shifts consistent with its role.

#Impact

TxPert's contribution is as much methodological as architectural: by insisting on strong baselines and a metric validated against retrieval, it raises the bar for a subfield whose claims have not survived independent replication. The demonstration that curated interaction graphs carry real, ablatable signal is a counterpoint to expression-only foundation models. Access is constrained. The code carries Recursion's Non-Commercial End User License Agreement, which forbids commercial use, and the released checkpoints — GAT, Exphormer, and Exphormer-MG — are a deliberate subset of what the paper reports: the public Exphormer-MG uses only STRING and GO, because PxMap and TxMap are proprietary. The authors state directly that the paper's best models rely on those withheld graphs and score higher than anything a public user can reproduce.