# In Silico Perturbation

Across several large Perturb-seq datasets, a generative framework called Graph-guided Flow Matching (GFM) combines three biological knowledge graphs, covering Gene Ontology, protein interactions and perturbation similarity, with latent flow matching to predict single-cell transcriptomes after unseen perturbations. It outperforms a linear baseline and the deep-learning models CellFlow, GEARS and scGPT on unseen single-gene perturbations, trains up to 30 times faster, and keeps its advantage as training data shrink; on two-gene perturbations it does better when neither gene was seen in training.

Across several large Perturb-seq datasets, a generative framework called Graph-guided Flow Matching (GFM) combines three biological knowledge graphs, covering Gene Ontology, protein interactions and perturbation similarity, with latent flow matching to predict single-cell transcriptomes after unseen perturbations. It outperforms a linear baseline and the deep-learning models CellFlow, GEARS and scGPT on unseen single-gene perturbations, trains up to 30 times faster, and keeps its advantage as training data shrink; on two-gene perturbations it does better when neither gene was seen in training.

mRNA foundation models usually average token embeddings over the whole transcript before fitting a linear probe, so signals from short regulatory regions are diluted by longer ones. A parameter-free region-aware readout instead averages within the 5′ UTR, coding sequence and 3′ UTR and concatenates the three means; across 20 frozen models it raises Pearson correlation on translation and stability tasks by 11–25% on average, with no model significantly worse on any regression task. On classification tasks AUPRC changes range from −3% to +11%.

mRNA foundation models usually average token embeddings over the whole transcript before fitting a linear probe, so signals from short regulatory regions are diluted by longer ones. A parameter-free region-aware readout instead averages within the 5′ UTR, coding sequence and 3′ UTR and concatenates the three means; across 20 frozen models it raises Pearson correlation on translation and stability tasks by 11–25% on average, with no model significantly worse on any regression task. On classification tasks AUPRC changes range from −3% to +11%.

In single-cell foundation models, predicting the effect of overexpressing genes had been limited to one gene at a time or to enumerating a few hundred combinations. A Kyushu Institute of Technology team recast choosing which genes to overexpress as a discrete adversarial attack on gene tokens and searched combinations with Greedy Coordinate Gradient (GCG), a search built for language models. In Geneformer and scGPT, GCG raised similarity to the target cell state above gradient-pool greedy at equal forward-pass cost, with the margin growing with set size; measured against published reprogramming factors, however, ranking by single-gene effect recovered more of them.

In single-cell foundation models, predicting the effect of overexpressing genes had been limited to one gene at a time or to enumerating a few hundred combinations. A Kyushu Institute of Technology team recast choosing which genes to overexpress as a discrete adversarial attack on gene tokens and searched combinations with Greedy Coordinate Gradient (GCG), a search built for language models. In Geneformer and scGPT, GCG raised similarity to the target cell state above gradient-pool greedy at equal forward-pass cost, with the margin growing with set size; measured against published reprogramming factors, however, ranking by single-gene effect recovered more of them.

Using the DNA language model Nucleotide Transformer v3 on the JUPITER supercomputer, the team substituted each of roughly 3.12 billion positions of the telomere-to-telomere human reference genome with each of the three alternative bases and recorded how model predictions changed elsewhere, yielding a genome-wide directed map of nucleotide dependencies. These dependencies preferentially link evolutionarily conserved positions and repeatedly assemble into seven geometric structures called Relons, some outside current annotation. Combined with CADD, dependency features raised intronic-variant pathogenicity AUPRC from 0.299 to 0.356.

Using the DNA language model Nucleotide Transformer v3 on the JUPITER supercomputer, the team substituted each of roughly 3.12 billion positions of the telomere-to-telomere human reference genome with each of the three alternative bases and recorded how model predictions changed elsewhere, yielding a genome-wide directed map of nucleotide dependencies. These dependencies preferentially link evolutionarily conserved positions and repeatedly assemble into seven geometric structures called Relons, some outside current annotation. Combined with CADD, dependency features raised intronic-variant pathogenicity AUPRC from 0.299 to 0.356.

Single-cell perturbation experiments are costly and low-throughput, yet computational prediction has been hampered by inconsistent evaluation and by models that do not reliably beat simple baselines. DCAF combines data pairing, an evaluation criterion and a prediction model in one framework: entropy-regularized optimal transport recovers a pre-perturbation state for each perturbed cell, log fold changes of non-silenced genes drive scoring, and a JEPA structure encodes control and perturbed cells jointly. Across cross-gene, cross-cell-type and cross-region settings, DCAF-model beats GEARS, scGPT, linear and mean baselines on R² and MSE of log fold change for non-silenced genes.

Single-cell perturbation experiments are costly and low-throughput, yet computational prediction has been hampered by inconsistent evaluation and by models that do not reliably beat simple baselines. DCAF combines data pairing, an evaluation criterion and a prediction model in one framework: entropy-regularized optimal transport recovers a pre-perturbation state for each perturbed cell, log fold changes of non-silenced genes drive scoring, and a JEPA structure encodes control and perturbed cells jointly. Across cross-gene, cross-cell-type and cross-region settings, DCAF-model beats GEARS, scGPT, linear and mean baselines on R² and MSE of log fold change for non-silenced genes.

JuloVelo moves gene kinetics pattern identification ahead of neural network training: each gene is first classified as Circle, Complete induction, or Complete repression from its unspliced–spliced RNA relationship, and training starts from pattern-matched pretrained weights, steering cell-specific kinetics away from biologically implausible parameter regimes. Across eight biological datasets spanning erythrogenesis, neurogenesis, and endocrinogenesis, plus simulated data with known kinetics, it was more stable than constant-rate and cell-specific models, was the only method to resolve all five hippocampal neurogenesis branches, and used promoter chromatin accessibility to process multiome data.

JuloVelo moves gene kinetics pattern identification ahead of neural network training: each gene is first classified as Circle, Complete induction, or Complete repression from its unspliced–spliced RNA relationship, and training starts from pattern-matched pretrained weights, steering cell-specific kinetics away from biologically implausible parameter regimes. Across eight biological datasets spanning erythrogenesis, neurogenesis, and endocrinogenesis, plus simulated data with known kinetics, it was more stable than constant-rate and cell-specific models, was the only method to resolve all five hippocampal neurogenesis branches, and used promoter chromatin accessibility to process multiome data.

Time-series spatial transcriptomics measures each tissue section once, so how cells migrate, proliferate and influence one another can only be inferred indirectly. CytoBridge learns continuous population evolution in a joint space of transcriptomic state and spatial coordinates, splitting velocity into an interaction term estimated from spatial neighborhoods and ligand–receptor priors and an intrinsic-context term, while allowing cell numbers to grow or shrink. Across five datasets on three spatial transcriptomics platforms it generates unmeasured intermediate stages; in a leave-one-stage-out benchmark it had the lowest Sliced-W2 in 12 of 33 combinations, more than any other method or control.

Time-series spatial transcriptomics measures each tissue section once, so how cells migrate, proliferate and influence one another can only be inferred indirectly. CytoBridge learns continuous population evolution in a joint space of transcriptomic state and spatial coordinates, splitting velocity into an interaction term estimated from spatial neighborhoods and ligand–receptor priors and an intrinsic-context term, while allowing cell numbers to grow or shrink. Across five datasets on three spatial transcriptomics platforms it generates unmeasured intermediate stages; in a leave-one-stage-out benchmark it had the lowest Sliced-W2 in 12 of 33 combinations, more than any other method or control.

DynaVelo puts RNA expression and transcription-factor motif accessibility from single-cell multiome data into one neural differential equation, learning a velocity field that can be perturbed rather than only drawn. In wild-type mouse germinal center B cells it recovers dark-zone cycling, the switch to the light zone and recycling; in Arid1a-heterozygous cells it nominated Ctcf, Stat3 and Bcl6 as rescue targets, and all three double-heterozygous mouse strains restored the dark-zone phenotype.

DynaVelo puts RNA expression and transcription-factor motif accessibility from single-cell multiome data into one neural differential equation, learning a velocity field that can be perturbed rather than only drawn. In wild-type mouse germinal center B cells it recovers dark-zone cycling, the switch to the light zone and recycling; in Arid1a-heterozygous cells it nominated Ctcf, Stat3 and Bcl6 as rescue targets, and all three double-heterozygous mouse strains restored the dark-zone phenotype.