Target Audience
- Readers interested in the evolution of machine learning technologies for genomic analysis.
- Researchers and engineers interested in biological simulation and predicting gene expression directly from DNA sequences.
- Professionals interested in applying AI to drug discovery, gene therapy, and synthetic biology.
Summary
- MPRA (Massively Parallel Reporter Assay) is an experimental technology that directly measures the transcriptional regulatory activity of large numbers of DNA sequences.
- The 2025 paper "Massively parallel characterization of transcriptional regulatory elements" measured more than 680,000 candidate regulatory elements, creating a large-scale dataset of functional measurements suitable for machine learning.
- While models such as DeepSEA and Enformer advanced the ability to predict biological properties from DNA sequences, MPRA has improved the quality and quantity of experimentally measured training labels.
- Today, two complementary directions are emerging: long-range sequence models such as Borzoi and AlphaGenome, and functional prediction models trained on MPRA data.
- Together, these advances are expanding the field from predicting gene expression toward designing regulatory DNA sequences with desired biological functions.
1. Introduction: The Role of MPRA in Genomic Machine Learning
Machine learning for genomic analysis has evolved from predicting transcription factor binding and chromatin states to predicting gene expression and, more recently, designing DNA sequences with desired functions.
An important contribution to this evolution is the 2025 Nature paper "Massively parallel characterization of transcriptional regulatory elements" by Agarwal, Inoue, and colleagues [1].
Rather than introducing a fundamentally new AI architecture, this study represents an important advance in generating experimentally measured functional training data at a scale suitable for machine learning.
The researchers directly measured more than 680,000 candidate cis-regulatory elements (cCREs) using lentiMPRA in three human cell types. They then used these measurements to develop models that predict regulatory activity and the effects of sequence variants.
To understand the significance of this study, it is helpful to review how genomic machine learning has evolved.
2. The Evolution of Genomic Machine Learning: From Motif Discovery to Deep Learning
2.1 Stage 1: The Era of Motif Discovery
From the 2000s through the early 2010s, a major focus of genomic analysis was identifying short DNA sequence patterns recognized by transcription factors.
Representative approaches included:
- Position Weight Matrices (PWMs)
- Motif analysis
- k-mer analysis
In these methods, researchers typically defined sequence features based on biological knowledge before supplying them to computational models.
In other words, humans needed to determine which sequence patterns might be important before machine learning could use them effectively.
2.2 Stage 2: CNNs Learn Sequence Features Automatically
Around 2015, deep learning models such as DeepBind and DeepSEA, followed by Basset, began transforming this approach [2][3][4].
DeepBind used convolutional neural networks (CNNs) to learn the sequence specificities of DNA- and RNA-binding proteins.
DeepSEA predicted biological properties directly from DNA sequences, including:
- Transcription factor binding
- DNase sensitivity
- Histone modifications
It also enabled the prediction of how single-nucleotide variants might affect these properties.
Basset used DNase-seq data across multiple cell types to predict chromatin accessibility directly from DNA sequences.
The major breakthrough was the ability to learn the following relationship in an end-to-end manner:
DNA sequence → Regulatory phenotype
Instead of requiring researchers to manually define every important sequence feature, neural networks could automatically learn patterns associated with regulatory activity.
2.3 A Fundamental Limitation of Early Deep Learning Models
However, this generation of models had an important limitation.
Most training targets were observational biochemical phenotypes, such as:
- ChIP-seq signals
- DNase-seq signals
- ATAC-seq signals
- Histone modifications
These measurements are extremely powerful, but they primarily provide observational or descriptive information.
For example, researchers may observe that chromatin is accessible at a particular DNA sequence.
However, this observation does not necessarily reveal how strongly the DNA sequence itself can drive transcription.
There is an important distinction between:
"Chromatin is accessible at the location of this sequence."
and:
"This sequence itself has the ability to activate transcription."
This distinction is central to understanding the importance of MPRA.
3. What Is MPRA? Directly Measuring the Function of Regulatory DNA Sequences
MPRA (Massively Parallel Reporter Assay) is an experimental technology that introduces thousands to hundreds of thousands of DNA sequences into cells and quantitatively measures their ability to drive transcription [5].
3.1 How MPRA Works
A typical MPRA experiment follows this general process:
DNA element
↓
Minimal promoter + Reporter
↓
Barcode RNA abundance / Barcode DNA abundance
↓
Regulatory activity
The key distinction is that MPRA does not merely observe where a DNA sequence is located or whether chromatin is accessible.
Instead, it measures the transcriptional activity produced by the sequence under defined experimental conditions.
In lentiMPRA, lentiviral vectors are used to integrate the tested sequences into the genome.
Compared with episomal MPRA, where reporter constructs remain outside the chromosomes, lentiMPRA provides a measurement environment that more closely resembles a genomic context [5].
3.2 The Difference in Machine Learning Training Data
Conventional genomic machine learning and MPRA differ in the types of information they provide.
| Comparison | Conventional observational assays | MPRA |
|---|---|---|
| Main measurement | Chromatin states, transcription factor binding, histone modifications | Transcriptional activity driven by DNA sequences |
| Representative technologies | ATAC-seq, ChIP-seq, DNase-seq | MPRA, lentiMPRA |
| Information obtained | Biochemical features associated with regulation | Functional activity of tested sequences under experimental conditions |
| Machine learning application | Indirect prediction of regulatory function | Direct experimental labels for transcriptional activity |
| Main limitation | Observed biochemical features do not necessarily represent regulatory function itself | The original genomic context is not fully reproduced |
Conventional approaches can be represented conceptually as:
DNA sequence → ATAC / ChIP / Histone marks → Putative enhancer
In contrast, MPRA directly measures:
DNA sequence → Measured transcriptional activity
From a machine learning perspective, this represents a major improvement in the nature of the training labels.
However, MPRA measures regulatory activity under specific experimental conditions.
Its results do not necessarily correspond exactly to the function of the same sequence at its native genomic location.
4. What the 2025 Paper Achieved: Scaling MPRA and Machine Learning
The importance of "Massively parallel characterization of transcriptional regulatory elements" is not that it invented MPRA.
MPRA had already been developed and applied in earlier research.
The study's contribution was to combine large-scale, high-quality lentiMPRA measurements with machine learning to investigate the regulatory rules encoded in DNA sequences [1].
4.1 Direct Measurement of More Than 680,000 Candidate Regulatory Elements
The researchers studied three human cell types:
- HepG2
- K562
- WTC11
They measured more than 680,000 candidate cCREs and reported that approximately 41.7% exhibited functional regulatory activity under the experimental conditions.
Furthermore, approximately 60,000 shared sequences were tested across all three cell types, enabling direct comparisons of cell-type specificity.
Generating functional measurements at this scale has important implications for machine learning.
4.2 Training Deep Learning Models on MPRA Measurements
The researchers used the resulting measurements as training labels for sequence-based models.
Models evaluated in the study included:
- MPRAnn
- MPRALegNet
- EnformerMPRA
- SeiMPRA
MPRALegNet was particularly interesting.
Although it is a relatively small CNN with approximately 1.3 million parameters, it achieved competitive performance compared with EnformerMPRA and SeiMPRA, which utilize features from much larger pretrained models.
The study also found that sequence-based models outperformed models relying exclusively on biochemical features.
4.3 Progress Does Not Depend Only on Larger Models
An important implication of this research is that increasing model size is not the only path toward better genomic AI.
Alongside developing more powerful neural network architectures, producing larger quantities of direct, high-quality experimental training data is equally important.
In machine learning, the choice of training targets matters as much as the design of the model.
This study demonstrated how large-scale measurements of regulatory function can be used to learn relationships between DNA sequences and transcriptional activity.
5. The Relationship with Enformer: Long-Range Genomic Context and Functional Training Data
Enformer, introduced in 2021, represented another major milestone in genomic machine learning [6].
Earlier CNN-based models often focused on relatively local DNA sequences, typically ranging from hundreds to thousands of base pairs.
Enformer introduced Transformer-based modeling to capture the effects of genomic interactions over much longer distances, including tens to approximately 100 kilobases.
This expanded the ability to predict gene expression and chromatin properties while considering interactions between distant enhancers and promoters.
Enformer was also evaluated on variant-effect prediction tasks involving MPRA saturation mutagenesis data.
5.1 Enformer and MPRA Represent Different Types of Progress
Enformer and large-scale MPRA are better understood as complementary advances rather than competing technologies.
| Technology | Main advance | Primary focus |
|---|---|---|
| Enformer | Modeling long-range genomic interactions | Incorporating broader DNA sequence context |
| Large-scale lentiMPRA | Direct measurement of regulatory activity | Generating experimentally grounded functional training data |
Enformer-type models improve the ability to interpret long genomic sequences.
MPRA-based approaches improve the availability of direct functional measurements for shorter regulatory elements.
Modern regulatory genomics is increasingly moving toward combining these two directions.
6. Connections to Current Technologies: Borzoi and AlphaGenome
Genomic machine learning is increasingly developing toward sequence-to-function models that predict multiple biological properties directly from DNA sequences.
Two representative examples are Borzoi and AlphaGenome.
6.1 Borzoi: Predicting RNA-seq from DNA Sequences
Borzoi is a model that predicts cell-type- and tissue-specific RNA-seq coverage from DNA sequences [7].
It can model multiple aspects of RNA regulation, including:
- Transcription
- Splicing
- Polyadenylation
It can also evaluate how DNA variants may affect these processes.
This expands the prediction task beyond:
DNA → Chromatin
toward:
DNA → RNA production / processing
As a result, sequence-based models can address a broader range of mechanisms involved in gene regulation.
6.2 AlphaGenome: Integrating Larger Genomic Contexts and Multiple Outputs
AlphaGenome is a large-scale DNA sequence model published in Nature in 2026 [8].
It processes DNA sequences of approximately one megabase and predicts numerous functional genomic measurements.
These include:
- Gene expression
- Transcription initiation
- Chromatin accessibility
- Histone modifications
- Transcription factor binding
- Chromatin contacts
- Splicing-related features
It also supports predictions of how noncoding variants affect these biological properties.
6.3 A Historical Overview of Genomic Machine Learning
The evolution of the field can be summarized as follows.
| Period or stage | Representative technologies | Major advance |
|---|---|---|
| Early approaches | PWM, motif analysis | Identifying sequence patterns recognized by transcription factors |
| Around 2015 | DeepBind | Learning transcription factor binding specificities from DNA sequences |
| 2015–2016 | DeepSEA, Basset | Predicting chromatin states from DNA sequences |
| 2018–2021 | Basenji, Enformer | Modeling long-range genomic context and gene expression |
| Development of large-scale MPRA | MPRA, lentiMPRA | Direct functional measurement of large numbers of regulatory sequences |
| 2025 | Agarwal, Inoue, and colleagues | Functional measurement and deep learning using more than 680,000 sequences |
| 2025–2026 | Borzoi, AlphaGenome | Predicting multiple regulatory outputs and variant effects using long-range sequence context |
This table is intended as a conceptual overview of technological development rather than a strictly linear chronology of publication dates.
MPRA technologies and long-range sequence models have evolved in parallel.
7. The Fundamental Difference Between MPRA-Based Models and Long-Range Sequence Models
MPRA-based models and models such as Enformer and AlphaGenome do not solve exactly the same problem.
Understanding this distinction is particularly important when considering applications to biological simulation.
7.1 MPRA Measures the Intrinsic Activity of DNA Sequences
MPRA is particularly useful for measuring how strongly a relatively short DNA sequence, often around 200 base pairs, can regulate transcription on its own.
This is often described as intrinsic or autonomous regulatory activity.
For example, MPRA can evaluate how strongly a particular enhancer sequence can activate transcription under standardized experimental conditions.
7.2 Gene Expression in the Genome Depends on Context
In an actual genome, gene expression is not determined solely by the DNA sequence of an individual enhancer.
Multiple factors contribute, including:
- Which enhancer contacts which promoter
- Chromatin state
- Three-dimensional genome structure
- Surrounding DNA sequences spanning tens of kilobases to megabases
- Transcription factor abundance
- Developmental state
Therefore, understanding biological behavior inside a living system requires accounting for both the intrinsic regulatory potential of a sequence and the environment in which it operates.
7.3 MPRA Does Not Fully Preserve Native Genomic Context
In lentiMPRA, tested DNA sequences are integrated into random genomic locations.
As a result, the measured activities effectively average over the influences of different integration sites, providing an estimate of sequence-intrinsic regulatory effects [1].
However, this does not fully reproduce the regulatory function of the sequence at its native genomic location.
The distinction can be summarized as follows.
MPRA-based models
"How much transcriptional regulatory activity does this DNA sequence possess on its own?"
Enformer- and AlphaGenome-type models
"What functional effects are predicted for this DNA sequence within its broader genomic context?"
The appropriate model therefore depends on the biological question being investigated.
8. Why Reducing Genomic Context Can Benefit Machine Learning
At first glance, losing genomic context may seem like a disadvantage of MPRA.
However, this simplification can be beneficial when the goal is to understand regulatory grammar.
8.1 Conventional Genomic Measurements Combine Multiple Influences
Suppose a gene associated with a particular enhancer is strongly expressed.
In conventional genomic measurements, expression may depend on numerous interacting factors.
Sequence
-
TF concentration
-
Chromatin
-
3D contact
-
Promoter
-
Neighboring elements
↓
Expression
Because these factors act together, isolating the contribution of any single factor is difficult.
8.2 MPRA Simplifies the Learning Problem
MPRA allows many experimental variables to be held relatively constant.
Conceptually, this makes the relationship closer to:
Sequence → Expression
In practice, the experimental environment and cellular state still influence the measurements.
Nevertheless, MPRA makes it easier to investigate the contribution of the DNA sequence itself compared with conventional genomic observations.
From a machine learning perspective, this creates a more clearly defined supervised learning problem.
8.3 Decoding Regulatory Grammar
By analyzing models trained on MPRA measurements, researchers can investigate:
- Which transcription factor motifs are important
- Motif orientation
- Motif position
- Interactions between motifs
- Differences between promoters and enhancers
Model attribution methods can help identify which regions of a DNA sequence contribute most strongly to predicted regulatory activity.
The goal of regulatory grammar research is to understand the rules through which combinations and arrangements of DNA sequence features control transcription.
MPRA provides an experimental foundation for studying these rules.
9. Developments in 2026: From Predicting Regulatory Sequences to Designing Them
In 2026, a Nature paper titled "Regulatory grammar in human promoters uncovered by MPRA-based deep learning" further advanced this direction [9].
The researchers used approximately 10 million MPRA promoter fragments to train a CNN model called PARM.
9.1 What PARM Achieved
PARM predicts promoter activity directly from DNA sequences.
The study also investigated:
- Position-dependent effects of transcription factor motifs
- Interactions between motifs
- Identification of regulatory sequence features
- Design of synthetic promoters
A particularly important contribution was the use of the learned model not only to interpret regulatory behavior but also to design new promoter sequences.
9.2 Measure → Learn → Interpret → Design
Earlier combinations of MPRA and machine learning can be described as:
Measure → Learn → Interpret
The process consists of three steps:
- Measure the regulatory activity of DNA sequences.
- Train a model using the experimental measurements.
- Interpret the learned regulatory rules.
PARM extends this process by introducing another stage:
Measure → Learn → Interpret → Design
The learned regulatory rules are used to design DNA sequences with desired properties.
This demonstrates how MPRA can evolve beyond an analytical technology into an experimental foundation for sequence design.
10. Connections to Biological Simulation and Drug Development
These advances have important implications for biological simulation and pharmaceutical research.
10.1 From Genome Annotation to Sequence Design
The objectives of genomic analysis have changed over time.
Initially, the focus was:
Genome → Annotation
The goal was to identify genes and regulatory elements within DNA sequences.
With the development of machine learning, the focus expanded to:
Genome → Prediction
Models began predicting transcription factor binding, chromatin states, gene expression, and other biological properties directly from DNA sequences.
MPRA introduced another important capability:
Genome sequence → Experimentally calibrated function
Models could now learn from large-scale measurements of regulatory activity rather than relying only on indirect observations.
More recently, research has increasingly addressed the inverse problem:
Desired function → Sequence design
10.2 Designing DNA Sequences with Desired Regulatory Activity
Traditional genomic prediction asks questions such as:
"How strongly will this DNA sequence activate gene expression?"
Sequence design introduces a different objective:
"Can we design an enhancer that drives strong expression in liver cells while remaining inactive in other cell types?"
This represents a shift from predicting the function of existing sequences to generating sequences with desired biological properties.
10.3 Potential Applications in Drug Development
These advances may support applications such as:
- Promoter and enhancer design for gene therapy
- Development of cell-type-specific vectors
- Synthetic biology
- Therapeutic regulation of gene expression
- Functional interpretation of noncoding variants
Agarwal, Inoue, and colleagues also highlighted the potential of large-scale functional measurements to support regulatory element design and therapeutic delivery [1].
Furthermore, research using PARM explored regulatory behavior under different cellular conditions, including responses to stimulation and drug treatment [9].
In the future, combining these approaches with patient-derived cells or organoids may help researchers understand disease progression, treatment responses, and potential drug targets.
However, it is important to distinguish these potential applications from capabilities that have already been experimentally validated.
MPRA measurements alone cannot fully reproduce the complexity of gene regulation inside living organisms.
Accurate biological simulation must also consider long-range genomic context, cellular state, and other regulatory factors.
11. Conclusion: MPRA as an Innovation in Training Data for Genomic AI
"Massively parallel characterization of transcriptional regulatory elements" represents an important milestone in the evolution of genomic AI.
Its primary contribution was not a fundamentally new neural network architecture, but rather an advance in the quality and scale of functional training data available for machine learning.
The roles of major technologies can be summarized as follows.
| Technology | Main contribution to genomic AI |
|---|---|
| DeepBind | Learning transcription factor binding specificities from DNA sequences |
| DeepSEA and Basset | Predicting chromatin states from DNA sequences |
| Enformer | Modeling long-range genomic context |
| Large-scale MPRA | Directly measuring regulatory function and providing experimental training labels |
| Agarwal, Inoue, and colleagues | Learning regulatory grammar from hundreds of thousands of experimentally measured sequences |
| Borzoi and AlphaGenome | Predicting multiple regulatory outputs using long-range DNA sequence context |
| PARM | Learning regulatory grammar from MPRA data and designing new promoter sequences |
DeepSEA and Basset demonstrated that neural networks could automatically learn regulatory features from DNA sequences.
Enformer extended this capability to longer-range genomic interactions.
In contrast, the study by Agarwal, Inoue, and colleagues demonstrated how large-scale direct measurements of regulatory function could be used to learn the rules encoded in DNA sequences.
Today, two technological directions are particularly important:
Large-scale sequence models × Large-scale functional genomics experiments
One direction, represented by AlphaGenome, focuses on predicting diverse genomic functions from long DNA sequences.
The other, represented by PARM, uses functional training data generated through MPRA to learn interpretable regulatory rules and design new DNA sequences.
Although these approaches do not address exactly the same problem, they are complementary.
For biological simulation and drug development, both capabilities are essential: predicting biological functions from DNA sequences and designing sequences that produce desired functions.
MPRA provides experimentally measured functional data that can support both directions, making it an increasingly important technology in the development of genomic AI.
References
[1] Agarwal, V., Inoue, F., Schubach, M. et al., "Massively parallel characterization of transcriptional regulatory elements," Nature, 639, 411–420, 2025. https://www.nature.com/articles/s41586-024-08430-9
[2] Alipanahi, B., Delong, A., Weirauch, M. T., and Frey, B. J., "Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning," Nature Biotechnology, 33, 831–838, 2015. https://doi.org/10.1038/nbt.3300
[3] Zhou, J., and Troyanskaya, O. G., "Predicting effects of noncoding variants with deep learning–based sequence model," Nature Methods, 12, 931–934, 2015. https://doi.org/10.1038/nmeth.3547
[4] Kelley, D. R., Snoek, J., and Rinn, J. L., "Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks," Genome Research, 26, 990–999, 2016. https://doi.org/10.1101/gr.200535.115
[5] Gordon, M. G., Inoue, F., Martin, B. et al., "lentiMPRA and MPRAflow for high-throughput functional characterization of gene regulatory elements," Nature Protocols, 15, 2387–2412, 2020. https://www.nature.com/articles/s41596-020-0333-5
[6] Avsec, Ž., Agarwal, V., Visentin, D. et al., "Effective gene expression prediction from sequence by integrating long-range interactions," Nature Methods, 18, 1196–1203, 2021. https://doi.org/10.1038/s41592-021-01252-x
[7] Linder, J., Srivastava, D., Yuan, H. et al., "Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation," Nature Genetics, 57, 949–961, 2025. https://www.nature.com/articles/s41588-024-02053-6
[8] Avsec, Ž., Latysheva, N., Cheng, J. et al., "Advancing regulatory variant effect prediction with AlphaGenome," Nature, 649, 1206–1218, 2026. https://www.nature.com/articles/s41586-025-10014-0
[9] Barbadilla-Martínez, L., Klaassen, N., Franceschini-Santos, V. H. et al., "Regulatory grammar in human promoters uncovered by MPRA-based deep learning," Nature, 651, 1107–1116, 2026. https://www.nature.com/articles/s41586-025-10093-z
[10] "Massively parallel characterization of transcriptional regulatory elements," PubMed Central, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC11903340/
[11] "Massively parallel reporter assays for investigating gene regulation," Nature Reviews Genetics, 2026. https://www.nature.com/articles/s41576-026-00944-4