To understand how a complex biological system works, biologists traditionally break it. By knocking out a gene via CRISPR or silencing it with RNA interference, researchers observe what goes wrong to infer the gene's original function. However, performing these physical experiments across tens of thousands of genes—let alone combinations of them—is prohibitively expensive and slow.
In silico genetic perturbation turns this physical paradigm into a computational one. The core intuition is straightforward: if a machine learning model can be trained to deeply understand the baseline state of a cell, it must implicitly learn the underlying rules governing how genes interact. Once the model captures this "cellular manifold," researchers can deliberately corrupt or alter specific components of the model's internal representations and observe how the rest of the system responds. This shifts biology from a descriptive science to an executable, predictive framework, fulfilling the foundational vision outlined by Evelyn Fox Keller in Beyond the Gene of bridging static sequences and dynamic biological outcomes through executable models.
Input-Level Masking and Topology Propagation
The most direct way to simulate a genetic knockout is at the input layer of a neural network. As Victor Curean notes in PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction, setting the values of targeted features to zero within an input vector provides a clean representational strategy for simulating gene knockouts without requiring any architectural changes to the underlying model.
When applied to Graph Neural Networks (GNNs) that explicitly model cellular networks, this zeroing strategy becomes a topological tool. In GRAPE: Heterogeneous Graph Representation Learning for Genetic Perturbation with Coding and Non-Coding Biotype, Changxi Chi et al. implement genetic perturbation simulations by masking the feature representations of specific target nodes. The network then propagates the effects of this localized knockout through its learned topology, allowing researchers to observe system-wide shifts in the expression states of other nodes.
A similar logic applies to sequence-level models. For instance, Yoseph Barash et al. in AVISPA: a web tool for the prediction and analysis of alternative splicing utilize a virtual knockout procedure that computationally removes sequence motifs to measure the resulting change in model output, thereby identifying regulatory drivers. Similarly, Kavi Gupta et al. in Improved modeling of RNA-binding protein motifs in an interpretable neural model of RNA splicing demonstrate that simulating the removal of specific binding sites within modular neural architectures can accurately identify whether a protein acts as a molecular activator or repressor, showing high concordance with experimental knockdown data.
Latent Space Geometry and Vector Arithmetic
While input masking is intuitive, cells are highly non-linear systems where genes rarely act in isolation. Modern single-cell foundation models and variational autoencoders (VAEs) overcome this by embedding cellular states into a continuous, high-dimensional latent space. In this latent space, complex biological behaviors are mapped onto geometric coordinates.
Once a model learns this geometry, simulating a perturbation becomes an exercise in vector manipulation. As Seungbyn Baek et al. explain in Single-cell foundation models: bringing artificial intelligence into cell biology, researchers can simulate biological experiments by manually modifying specific gene values within the model's high-dimensional latent representation, predicting how a cell's state shifts in response to knockouts or drug treatments without physical intervention.
This geometric shifting allows for the simulation of entire cellular trajectories. Ioana Bica et al. in Unsupervised generative and graph representation learning for modelling cell differentiation demonstrate that perturbing coordinates within a latent space and passing the results through a decoder allows for the simulation of complex state transitions, such as cellular reprogramming. By measuring the distance and trajectory between the original and perturbed vectors, the authors quantify the biological complexity required to shift a cell from one identity to another.
Remarkably, this latent manipulation does not always require massive foundation models. Suwen Wei et al. in Investigation of cell development and tissue structure network based on natural Language processing of scRNA-seq data apply natural language processing techniques to treat genes as words and cells as sentences. By representing genes as vectors in a shared latent space, the authors simulate genetic perturbations through simple vector arithmetic. Altering specific gene vectors and recalculating cell positions allows them to accurately predict how up-regulating or down-regulating specific transcription factors shifts a cell's developmental path.
Cascading Effects and Generative Dynamics
How do these localized latent adjustments translate into system-wide predictions? Recent architectures leverage generative frameworks and network propagation to model how a single perturbation cascades across an entire cellular network or tissue microenvironment.
In How to Build the Virtual Cell with Artificial Intelligence: Priorities and Opportunities, Charlotte Bunne et al. highlight how generative frameworks like diffusion models and flow matching are used to simulate the effects of perturbations by predicting global shifts in a cell's latent state. This generative approach can even simulate multicellular environments. For instance, Amir Akbarnejad et al. introduce MintFlow in Mapping and reprogramming human tissue microenvironments with MintFlow, which conditions a generative model on neighborhood composition. This allows researchers to perform computational experiments where specific cell types are virtually added or removed from a tissue, predicting the resulting shifts in gene expression for the remaining cells.
Alternatively, network propagation algorithms can simulate these cascades by treating the cell as an interconnected graph of physical interactions. Minwoo Pak et al. demonstrate in Improved drug response prediction by drug target data integration via network-based profiling that computationally simulating drug-induced gene expression changes through network propagation serves as a reliable proxy for actual post-treatment transcriptomic data, capturing downstream effects without requiring expensive physical samples.
Why the Virtual Laboratory Works
The ultimate power of in silico perturbation modeling lies in its ability to act as a high-fidelity, differentiable virtual laboratory. By encoding massive amounts of data, these models capture the underlying constraints of biology well enough to generalize to unseen scenarios.
For example, Ao Li et al. developed GeneCompass in GeneCompass: deciphering universal gene regulatory mechanisms with a knowledge-informed cross-species foundation model, demonstrating that foundation models trained on massive single-cell datasets can accurately simulate the effects of gene deletions. Their in silico perturbations closely match experimental results in complex biological processes like organ development. Similarly, Rui Peng et al. show in KINDLE: Knowledge-Guided Distillation for Prior-Free Gene Regulatory Network Inference that high-fidelity regulatory models can simulate gene knockouts to accurately predict phenotypic outcomes, such as the arrest or promotion of cellular differentiation.
By allowing researchers to simulate thousands of "what-if" scenarios simultaneously, as emphasized by Prashant S. Emani et al. in Single-cell genomics and regulatory networks for 388 human brains and Claudia Cabrera et al. in AI approaches for the discovery and validation of drug targets, virtual perturbations provide a scalable mechanism to screen and prioritize therapeutic targets in a digital environment before moving to costly physical validation.
Key Foundational Work
Four-dimensional realistic modeling of pancreatic organogenesis (2008) — contributed 1 concept elements to this direction.
Beyond the Gene (2007) — contributed 1 concept elements to this direction.