A transcription model is any mathematical or computational framework that describes how, when, and how much a gene gets read into RNA. These models range from simple switches that treat a gene as “on” or “off” to deep learning systems that scan raw DNA sequence and predict expression levels across cell types. The field has expanded rapidly, and different classes of models address fundamentally different questions about gene regulation, from how physical forces position proteins on DNA to how a single cell decides which genes to activate in a given moment.
Why Scientists Need Models of Transcription
Transcription is not a passive copying process. The cell’s machinery must locate a gene, assemble a complex of proteins at the right spot, and then decide whether to actually start making RNA or wait. Signals from outside the cell, the physical packaging of DNA, and even random molecular collisions all influence the outcome. This complexity means that measuring gene activity in an experiment gives you a snapshot but not an explanation. Models fill the gap: they provide a testable framework for understanding why a gene is active in one cell type and silent in another, or why genetically identical cells can show wildly different levels of the same transcript.
Early modeling efforts treated genes as binary switches and used differential equations to simulate how networks of genes influence one another. That approach showed how common circuit motifs, like feedback loops, can produce multiple stable states, oscillations, and other dynamic behaviors that help cells respond to their environment.
1PubMed Central. Modeling transcriptional control in gene networks–methods, recent results, and future directions Those insights still underpin much of systems biology, but the field has since branched into several distinct modeling traditions, each tackling a different layer of the problem.
Thermodynamic Models
One of the most influential approaches treats gene regulation as a problem of physical chemistry. In a thermodynamic model, every possible arrangement of proteins on a stretch of DNA has an associated energy cost. The model calculates the probability that RNA polymerase (the enzyme that makes RNA) is bound and ready to transcribe, based on the energy landscape created by all the transcription factors, co-activators, and other molecules competing for space on the DNA. The predicted expression level of a gene is essentially the fraction of time polymerase ends up in the right position to start working.
2Nucleic Acids Research. Discrimination between thermodynamic models of cis-regulation using transcription factor occupancy data
This framework has been extended in several directions. One line of work incorporates how nucleosomes, the protein spools around which DNA is wrapped, physically block or expose binding sites. By modeling how nucleosome positioning changes in response to varying concentrations of regulatory proteins, researchers can predict how a promoter’s activation dynamics shift under different conditions.
3PubMed Central. Incorporating nucleosomes into thermodynamic models of transcription regulation Another extension accounts for transcription factors that do not bind DNA directly but are recruited through protein-protein interactions, a common arrangement at enhancers where multiple factors cooperate.
4iScience. Thermodynamics-based modeling reveals regulatory effects of indirect transcription factor-DNA binding
Thermodynamic models work best when you know which proteins are involved, where they bind, and roughly how strong those interactions are. Their appeal is interpretability: every parameter maps to a physical quantity you can, in principle, measure. Their weakness is that they assume the system reaches a kind of equilibrium, which is not always a safe assumption in a living cell where energy is constantly being spent.
Stochastic Models and Transcriptional Bursting
If you could watch a single gene in a single cell, you would not see a smooth, steady stream of RNA being produced. Instead, you would see something more like a flickering light: bursts of activity separated by quiet stretches. This burstiness, called transcriptional noise, is not a flaw. It is an intrinsic feature of how genes operate, driven by the random switching of a gene’s promoter between active and inactive states.
5Science. Genetic Determinants and Cellular Constraints in Noisy Gene Expression
The workhorse model for this behavior is often called the telegraph model. It imagines a gene flipping between an “on” state and an “off” state. RNA is produced only in the on state, and it degrades at a constant rate regardless. Even with just a handful of parameters (how fast the gene switches on, how fast it switches off, and how fast RNA is made while it is on), this model captures the characteristic bursting patterns seen in real cells.
6Bioinformatics. Scalable inference and identifiability of kinetic parameters for transcriptional bursting from single cell data By fitting the telegraph model to data from thousands of individual cells, researchers can back out the burst frequency and burst size for specific genes, giving a quantitative handle on how noisy a gene’s expression really is.
This noise has functional consequences. In developing embryos, for example, live imaging has revealed that adjusting burst frequency across different positions helps shape precise spatial patterns of gene expression. But bursting alone cannot fully explain how sharply defined those patterns are. The amount of time each cell spends transcribing also matters, suggesting that at least two distinct molecular processes work together to sculpt gene expression in space.
7Proceedings of the National Academy of Sciences. Multimodal transcriptional control of pattern formation in embryonic development
Polymerase Pausing as a Regulatory Checkpoint
For many genes, the bottleneck in transcription is not getting polymerase to the promoter in the first place. It is what happens immediately after. RNA polymerase II often starts transcribing a short stretch of RNA and then stalls, pausing near the start of the gene. This promoter-proximal pause is a widespread phenomenon in fruit flies and mammals, and it acts as a regulatory checkpoint: the gene is loaded and ready, but a separate signal is needed to release the enzyme into productive elongation.
8Science. Transcription Regulation Through Promoter-Proximal Pausing of RNA Polymerase II
This mechanism has important implications for how quickly cells can respond to signals. Because polymerase is already in place and waiting, the gene can be activated almost instantly once the release signal arrives, rather than having to assemble the transcription machinery from scratch. Cellular signaling pathways exploit this by controlling the proteins that either establish or release the pause, tuning both the timing and the magnitude of transcriptional responses.
9Cell Press. Trends in Biochemical Sciences Models that ignore pausing miss this entire layer of regulation and can misjudge how fast a gene can ramp up output.
Phase-Separated Condensates
A newer and still-debated idea is that transcription is partly controlled by liquid-like droplets, called condensates, that form inside the nucleus. These condensates concentrate transcription factors, co-activators, and parts of the transcription machinery into small compartments, much like oil droplets in water. At super-enhancers, which are large clusters of regulatory DNA that drive expression of cell-identity genes, proteins like BRD4 and the Mediator complex have been observed forming such condensates.
10Science. Coactivator condensation at super-enhancers links phase separation and gene control
11PubMed Central. Super-Enhancers, Phase-Separated Condensates, and 3D Genome Organization in Cancer
Direct observation has shown that when a condensate physically approaches a gene locus (within about one micrometer), both the frequency and size of transcriptional bursts increase. When the condensate drifts away, the gene drops back to a lower, basal level of bursting. Researchers have proposed a “three-way kissing” model in which the condensate transiently contacts both the gene and its regulatory DNA elements to boost output.
12PubMed. Direct observation of a condensate effect on super-enhancer controlled gene bursting This adds a spatial, almost mechanical dimension to transcription modeling that older equilibrium-based and stochastic frameworks do not capture well.
Deep Learning Models That Predict Expression from DNA Sequence
The models described so far are mechanistic: they try to represent the physical or biochemical processes that control transcription. A very different approach skips the mechanism entirely and asks a brute-force question: given a stretch of DNA, can a computer learn to predict how much RNA will be produced? Deep learning models trained on large datasets of genome sequence paired with expression measurements have made striking progress on exactly this question.
13PubMed. Predicting gene expression from DNA sequence using deep learning models
One landmark architecture, called Enformer, uses transformer layers (the same technology behind large language models) to integrate information from DNA sequences up to 100 kilobases away from a gene, capturing the influence of distant enhancers and other regulatory elements that simpler models miss.
14PubMed Central. Effective gene expression prediction from sequence by integrating long-range interactions Separate transformer-based models have been designed specifically to predict which enhancers physically contact which promoters, a key question for understanding how genes are regulated across large genomic distances.
15Briefings in Bioinformatics. Capturing large genomic contexts for accurately predicting enhancer-promoter interactions
These models have also begun to incorporate epigenomic data, like histone modifications and chromatin accessibility, to improve predictions. A deep learning framework that integrates both DNA sequence and chromatin accessibility can predict the locations of specific histone marks more accurately than sequence-only approaches, and models using attention mechanisms can predict transcription levels from histone mark activity at both promoters and distant regulatory regions.
16PubMed Central. DeepHistone: a deep learning approach to predicting histone modifications
17Nucleic Acids Research. Predicting gene expression from histone marks using chromatin deep learning models depends on histone mark function, regulatory distance and cellular states
RNA Velocity and Single-Cell Kinetics
Single-cell RNA sequencing generates snapshots of gene expression in thousands of individual cells, but each cell is measured only once. To infer the direction a cell is heading, in terms of whether a gene is ramping up or winding down, researchers developed a concept called RNA velocity. The idea exploits the fact that newly transcribed RNA still contains unspliced segments (introns), while mature RNA has been processed. By comparing the ratio of unspliced to spliced RNA for each gene in each cell, you can estimate the rate of change in expression, essentially a velocity vector pointing toward the cell’s future state.
18PubMed Central. RNA velocity of single cells
Later refinements introduced a dynamical version of this model that infers gene-specific rates of transcription, splicing, and degradation, allowing each cell to be placed along a differentiation trajectory and highlighting the genes most likely to be driving transitions between cell states.
19PubMed Central. Generalizing RNA velocity to transient cell states through dynamical modeling RNA velocity has become a standard tool in single-cell biology, though it rests on simplifying assumptions (such as uniform splicing rates across genes) that can produce misleading trajectories when those assumptions break down.
20Molecular Cell. RNA Velocity: Molecular Kinetics from Single-Cell RNA-Seq
How Stress Reshapes Transcription on a Genome-Wide Scale
Cells do not transcribe the same genes at the same rates all the time. Under stress, like heat shock or nutrient starvation, the transcriptional landscape is rapidly reorganized. Recent kinetic studies have revealed that cells appear to “prewire” their stress responses: the machinery is configured so that genome-wide changes in transcription can be executed quickly, and the system is simultaneously set up for recovery once the stress lifts.
21PubMed Central. Molecular mechanisms driving transcriptional stress responses
Interestingly, the transient shape of stress responses is sculpted by the interplay of RNA production and RNA degradation. Many genes that are rapidly induced during stress are also simultaneously destabilized, meaning their RNA is being broken down faster even as more is being made. This counterintuitive strategy allows a gene to maintain high steady-state levels under normal conditions while still responding quickly when conditions change.
22PubMed Central. Transient transcriptional responses to stress are generated by opposing effects of mRNA production and degradation Stochastic modeling has added another layer: in yeast, bursts of a stress-responsive transcription factor moving in and out of the nucleus produce a more robust transcriptional response than sustained nuclear residence, and this pulsing behavior arises from noise in the signaling pathways amplified by the small number of regulatory molecules in each cell.
23Molecular Biology of the Cell. Noise and interlocking signaling pathways promote distinct transcription factor dynamics in response to different stresses
Cross-Species Predictions and Their Limits
If the grammar of gene regulation is partly conserved across evolution, a model trained on one species might work in another. Training deep learning models simultaneously on human and mouse genomes does improve gene expression prediction in both species, and mouse-trained regulatory models can be repurposed to analyze human genetic variants associated with disease.
24PubMed Central. Cross-species regulatory sequence activity prediction In plants, a model trained jointly on rice and maize can make reasonable predictions in other grass species it has never seen, as long as they share conserved regulatory features.
25PubMed Central. Cross-species prediction of histone modifications in plants via deep learning
The catch is that cross-species performance is consistently worse than within-species performance. Much of this drop comes from species-specific repetitive sequences that confuse models trained on a different genome. Domain adaptation techniques, which adjust the network architecture to discourage learning of species-specific repeat patterns, can partially correct these errors, but the gap does not close entirely.
26PubMed Central. Domain-adaptive neural networks improve cross-species prediction of transcription factor binding In plants, performance degrades further as evolutionary distance increases: predictions within a plant family are much more reliable than predictions across families like grasses and mustard plants.
Where Current Models Fall Short
The most sobering result in recent benchmarking comes from testing whether deep learning models can predict how gene expression differs between individuals based on their personal DNA sequence. Using paired genome and expression data from over 800 people, researchers found that current models struggle to correctly predict even the direction in which a genetic variant pushes expression, let alone its magnitude. The problem traces to insufficiently learned sequence grammar: the models pick up broad patterns well enough to distinguish tissue types and genomic regions, but they have not yet cracked the subtle motif interactions that cause one person’s version of a gene to produce more RNA than another’s.
27PubMed. Benchmarking of deep neural networks for predicting personal gene expression from DNA sequence highlights shortcomings
This gap matters for clinical applications. Interpreting non-coding genetic variants, the vast majority of variants flagged in genome-wide studies, requires models that can reliably say whether a given mutation increases or decreases the expression of a nearby gene. Deep learning approaches for this task have proliferated, and they show real promise for uncovering regulatory mechanisms and prioritizing candidate variants.
28PubMed Central. Deep learning approaches for non-coding genetic variant effect prediction: current progress and future prospects But the benchmarking results make clear that we are not yet at the point where a model can serve as a reliable personal DNA interpreter for individual patients.
Feedback Loops and Network-Level Dynamics
Zooming out from single genes to gene regulatory networks introduces another class of models altogether. When genes regulate each other in loops, the system can exhibit behaviors that no single gene model would predict. Networks with interlocking positive and negative feedback loops can produce multiple stable states (where the same network settles into different expression patterns depending on initial conditions) or sustained oscillations, as seen in circadian rhythms and cell cycle control.
29PubMed Central. Identification, visualization, statistical analysis and mathematical modeling of high-feedback loops in gene regulatory networks
A single network can even switch between different modes of oscillation. In one mode, the frequency stays constant while the amplitude changes; in another, the amplitude is fixed while the frequency varies. The key parameters that control the switch between modes have been identified through simulation, and the ability of a simple network to produce both behaviors highlights how rich the dynamics of even small gene circuits can be.
30International Journal of Modern Physics C. Two different modes of oscillation in a gene transcription regulatory network with interlinked positive and negative feedback loops
Engineering Transcription with Generative AI
If a model can predict expression from sequence, the logic can be run in reverse: start from a desired expression pattern and design a DNA sequence that produces it. Deep learning models have been used to design synthetic enhancers that drive gene expression only in specific cell types, beginning from random sequences and optimizing them computationally. The process allows researchers to trace which sequence features matter at single-nucleotide resolution, turning the model into both a design tool and a discovery tool.
31Nature. Cell-type-directed design of synthetic enhancers
Toolkits built on generative AI now package multiple design strategies into user-friendly frameworks. These allow researchers to compare different network architectures for designing promoters and enhancers from scratch, lowering the barrier for labs that want to engineer regulatory elements without building their own deep learning pipeline.
32Bioinformatics. GPro: generative AI-empowered toolkit for promoter design The practical applications range from gene therapy, where you might want a synthetic promoter that activates only in liver cells, to industrial biotechnology, where fine-tuning expression levels in engineered microbes can make the difference between a process that works and one that fails economically.