Intended Audience

  • Researchers and readers interested in the evolution of machine learning for genomics and the role of massively parallel reporter assays (MPRAs).
  • Those interested in predicting and designing gene regulatory functions from DNA sequences, as well as understanding the scientific limitations of these approaches.
  • Scientists and professionals exploring applications of genomic AI in biological simulation, gene therapy, drug discovery, and synthetic biology.

Key Takeaways

  • MPRA is an experimental technology for measuring the transcriptional regulatory activity of large numbers of DNA sequences under controlled assay conditions. A landmark Nature study published in 2025 characterized more than 680,000 candidate regulatory sequences, generating a large-scale functional dataset for machine learning.
  • While DeepSEA and Basset advanced the prediction of biochemical features from DNA sequences, and Enformer, Borzoi, and AlphaGenome extended predictions to long-range genomic contexts and gene expression, MPRA has advanced the experimental measurement of sequence-dependent regulatory activity.
  • The activity measured by MPRA is not necessarily equivalent to a sequence's regulatory function at its native genomic location. MPRA results can depend on sequence length, chromatin context, promoter choice, integration site, and cellular state. They do not directly establish effects on endogenous target genes or disease phenotypes.
  • By 2026, studies such as PARM had further advanced the prediction, interpretation, and design of regulatory sequences using MPRA data. However, this research direction dates back to early MPRA studies in 2012. Recent progress represents an expansion in scale, predictive power, and applicability rather than the introduction of an entirely new concept.
  • A central challenge for future research is not simply to measure more sequences, but to understand the biological conditions under which regulatory potential measured by MPRA translates into actual regulatory function in vivo.

1. Introduction: What Has MPRA Changed in the Evolution of Genomic AI?

Machine learning in genomics has evolved from predicting transcription factor binding and chromatin states from DNA sequences to predicting gene expression and, increasingly, designing DNA sequences with desired biological functions.

An important milestone in this progression is the 2025 Nature paper, "Massively parallel characterization of transcriptional regulatory elements," by Agarwal, Inoue, and colleagues [1].

In this study, the researchers used lentiviral massively parallel reporter assays (lentiMPRA) to measure more than 680,000 candidate cis-regulatory elements (cCREs) across three human cell lines. They then trained deep learning models to predict transcriptional regulatory activity and the effects of sequence variants.

The significance of this work does not lie primarily in the invention of a new AI architecture.

Rather, it lies in the large-scale experimental generation of functional training data that machine learning models can use to learn relationships between DNA sequences and regulatory activity.

In genomic AI, the architecture of a model is important, but so is the biological meaning of the target it is trained to predict.

For example, a model that predicts whether chromatin is accessible and a model that predicts whether a DNA sequence can activate transcription are learning different aspects of gene regulation.

MPRA provides large-scale experimental data for the latter problem.

However, this raises an important question:

Is the regulatory activity measured by MPRA equivalent to the function that the same DNA sequence performs in its native genomic environment?

MPRA typically removes DNA sequences from their native genomic context and measures their regulatory effects under controlled experimental conditions.

This makes it easier to isolate and compare sequence-dependent effects. At the same time, it removes or alters information about native chromatin structure, interactions with target genes, and cellular context.

In other words, the strengths and limitations of MPRA are closely connected.

This article examines MPRA within the historical development of genomic machine learning, reviews the contributions of the large-scale 2025 study, and explores the assumptions underlying MPRA experiments and the biological questions that remain unresolved.

It also discusses the relationship between MPRA and models such as Enformer, Borzoi, AlphaGenome, and PARM, before considering the challenges that must be addressed to connect these technologies with in vivo simulation and drug discovery.

2. The Evolution of Genomic Machine Learning: From Motif Discovery to Deep Learning

2.1 Stage 1: Identifying Sequence Motifs

From the 2000s through the early 2010s, an important part of understanding transcriptional regulation involved identifying short DNA sequence patterns recognized by transcription factors.

Representative approaches included:

  • Position Weight Matrices (PWMs)
  • Motif analysis
  • k-mer analysis

These approaches generally relied on researchers defining biologically meaningful sequence features and using them in computational analyses.

For example, a transcription factor's preferred binding sequences could be represented as a motif, and researchers could investigate where that motif occurred across the genome.

Such methods made it possible to systematically analyze sequence patterns associated with transcriptional regulation.

However, important challenges remained.

How do multiple transcription factors cooperate? How important are motif positions and spacing? How should complex combinations of regulatory sequences be represented?

These questions motivated the development of more flexible computational approaches.

2.2 Stage 2: CNNs Learn Features Directly from DNA Sequences

Around 2015, deep learning models such as DeepBind and DeepSEA emerged, followed by Basset [2][3][4].

These studies demonstrated that convolutional neural networks (CNNs) could automatically learn sequence features associated with transcriptional regulation.

DeepBind uses CNNs to learn the sequence specificities of DNA- and RNA-binding proteins [2].

DeepSEA predicts biochemical features from DNA sequences, including [3]:

  • Transcription factor binding
  • DNase sensitivity
  • Histone modifications

DeepSEA also made it possible to evaluate how single-nucleotide variants might affect these predicted features.

Basset uses DNase-seq data from multiple cell types to predict chromatin accessibility directly from DNA sequences [4].

A major advance of this period was the ability to learn relationships between DNA sequences and regulatory phenotypes in an end-to-end manner.

Conceptually, the relationship can be expressed as:

DNA sequence → Regulatory phenotype

Rather than requiring researchers to manually define every relevant sequence feature, deep learning models could discover patterns useful for prediction.

2.3 A Fundamental Limitation of Early Genomic Deep Learning

Despite these advances, an important limitation remained.

Many of the training targets used by these models represented observational biochemical phenotypes, such as:

  • ChIP-seq
  • DNase-seq
  • ATAC-seq
  • Histone modifications

These are powerful technologies for investigating transcription factor binding, chromatin accessibility, and other aspects of gene regulation.

However, they primarily provide observational or descriptive measurements. They do not necessarily establish the causal effect of a particular DNA sequence on transcription.

For example, observing accessible chromatin at a genomic location tells us that the region may be involved in regulatory activity.

It does not necessarily tell us how strongly the DNA sequence itself can increase transcription.

In other words, there is an important distinction between:

"This genomic region appears to be associated with active transcription."

and:

"This DNA sequence has the capacity to promote transcription under specified conditions."

This distinction helps explain the importance of MPRA.

However, MPRA does not replace conventional observational genomic data.

Information about chromatin state and transcription factor binding remains essential for understanding regulation in native genomic environments.

MPRA and conventional genomic assays should therefore be viewed as complementary technologies that measure different aspects of gene regulation.

3. The History of MPRA: What Problem Was It Developed to Solve?

3.1 From Conventional Reporter Assays to MPRA

A massively parallel reporter assay (MPRA) is an experimental method for measuring how large numbers of DNA sequences affect transcription.

The basic principle of using reporter genes to investigate regulatory sequences predates MPRA.

In a conventional reporter assay, a candidate regulatory sequence is combined with a reporter gene and introduced into cells. Researchers then measure reporter expression.

For example, a candidate enhancer can be tested for its ability to increase reporter gene expression.

However, conventional reporter assays typically require individual experiments for each sequence or small group of sequences.

Systematically testing tens or hundreds of thousands of sequences was therefore difficult.

MPRA addressed this limitation by enabling massively parallel measurements of regulatory activity.

3.2 2012: MPRA-Based Modeling and Sequence Design Had Already Begun

One of the important early studies was published by Melnikov and colleagues in Nature Biotechnology in 2012 [12].

The researchers investigated more than 27,000 enhancer variants, systematically examining relationships between DNA sequences and inducible enhancer activity.

They also modeled sequence–activity relationships and explored the optimization of enhancer sequences.

This historical point is important.

The idea of using MPRA to measure regulatory activity, learn relationships between sequence and function, and optimize DNA sequences did not originate with the deep learning studies published in 2025 or 2026.

The basic research cycle was already present in early MPRA studies:

Measure → Learn → Design

More recent research has extended this direction through larger datasets, more powerful models, and increasingly sophisticated analyses.

3.3 2017: Differences Between Episomal and Chromosomal Contexts Became Clear

Another important milestone was the comparative study published by Inoue and colleagues in 2017 [13].

Many early MPRA implementations used episomal constructs, in which reporter DNA remained outside the chromosomes.

However, episomal DNA does not necessarily experience the same chromatin environment as DNA located within the native genome.

Inoue and colleagues compared enhancer activities measured using episomal and chromosomally integrated reporter constructs.

They demonstrated substantial differences between the results.

This finding established that MPRA activity depends not only on the DNA sequence but also on the experimental context in which the sequence is measured.

One important approach for addressing this limitation is lentiMPRA, which uses lentiviral delivery to integrate reporter constructs into the genome [5][13].

Nevertheless, lentiMPRA does not restore a candidate sequence to its original genomic position.

It incorporates aspects of the chromosomal environment, but it does not fully reproduce the regulatory function of the sequence at its native locus.

3.4 2020: MPRA Measurements Depend on Assay Design

In 2020, Klein and colleagues systematically compared nine MPRA approaches using 2,440 candidate enhancer and control sequences [14].

Their study demonstrated that assay design can influence reproducibility, dynamic range, and relationships between sequence features and measured activity.

They also compared enhancer fragments of different lengths: 192 bp, 354 bp, and 678 bp.

Substantial differences in activity were observed depending on sequence length.

In contrast, enhancer activity was generally orientation-independent for the tested sequences and assay configurations.

These findings are important for interpreting MPRA results.

Measured regulatory activity is not necessarily a single intrinsic physical property of a DNA sequence that is completely independent of experimental conditions.

The result can depend on which portion of the sequence is tested, which promoter is used, and how the assay is constructed.

3.5 MPRA in Historical Perspective

The historical development can be summarized as follows.

Period Representative Research or Technology Major Contribution
Conventional reporter assays Individual regulatory sequence experiments Measuring the effects of candidate sequences on reporter expression
2000s Genome-wide assays such as ChIP-seq Identifying large numbers of candidate regulatory regions
2012 Melnikov et al. Parallel measurement, modeling, and optimization of enhancer sequences
2015–2016 DeepBind, DeepSEA, Basset Learning regulatory sequence features using deep learning
2017 Inoue et al. Demonstrating differences between episomal and chromosomal MPRA measurements
2020 Klein et al. Systematically evaluating assay design and sequence-length dependencies
2021 Enformer Modeling long-range genomic sequence context
2025 Agarwal, Inoue et al. Functional characterization of more than 680,000 sequences and deep learning
2025–2026 Borzoi, AlphaGenome Predicting diverse genomic functions from longer sequence contexts
2026 PARM Advancing promoter activity prediction, interpretation, and synthetic promoter design

This history shows that MPRA is more than a technology for increasing experimental throughput.

It has also advanced research into how regulatory capacity can be defined, measured, and modeled from DNA sequences.

At the same time, the history of MPRA is also a history of investigating what its measurements actually represent.

4. What Does MPRA Actually Measure?

4.1 The Basic Mechanism of MPRA

In a typical enhancer MPRA, candidate DNA sequences are combined with reporter constructs and introduced into cells [5].

The experimental process can be summarized as follows:

DNA element

↓

Minimal promoter + Reporter

↓

Barcode RNA abundance / Barcode DNA abundance

↓

Regulatory activity

By associating individual DNA sequences with identifying barcodes, researchers can measure the activities of many sequences simultaneously.

Regulatory activity is commonly estimated by normalizing RNA abundance to the corresponding DNA abundance.

Conceptually, this can be written as:

[ A(s)=\log_2\left(\frac{\mathrm{RNA}(s)}{\mathrm{DNA}(s)}\right) ]

Here, (s) represents the tested DNA sequence, and (A(s)) represents its measured activity.

Actual analysis pipelines may also include barcode aggregation, normalization procedures, and comparisons with control sequences.

The important distinction is that MPRA measures a functional expression output rather than simply identifying the presence of a sequence or a particular chromatin state.

More precisely, MPRA measures how strongly a DNA sequence affects reporter expression under the conditions established by the assay.

However, RNA abundance is not determined exclusively by transcription rate.

It may also be affected by RNA stability, RNA processing, barcode effects, and other properties of the experimental system.

Therefore, the RNA/DNA ratio should not automatically be interpreted as a pure measure of transcription rate under all conditions.

4.2 Episomal MPRA Versus lentiMPRA

There are several MPRA implementations.

An important distinction is whether the reporter construct remains outside the chromosomes or becomes integrated into chromosomal DNA.

Feature Episomal MPRA lentiMPRA
DNA location Primarily extrachromosomal Integrated into chromosomes
Main approach Measuring regulatory activity using constructs such as plasmids Using lentiviral delivery to integrate reporter constructs
Chromatin environment Different from native chromosomal context Influenced by the chromosomal environment
Genomic location Does not reproduce the native location Usually integrates at locations different from the native locus
Main limitation Dependence on episomal assay context Dependence on integration site, copy number, and local chromosomal environment

lentiMPRA is important because it enables regulatory activity to be measured in a chromosomal setting rather than exclusively in an episomal environment [5][13].

However, there is a fundamental difference between being integrated into a chromosome and reproducing the native genomic locus.

This distinction matters when connecting MPRA measurements to endogenous gene regulation.

4.3 Observational Genomic Data Versus MPRA Training Data

The differences can be summarized as follows.

Feature Observational Genomic Assays MPRA
Main measurements Chromatin accessibility, transcription factor binding, histone modifications Regulatory activity in a reporter system
Representative methods ATAC-seq, ChIP-seq, DNase-seq MPRA, lentiMPRA
Biological information Biochemical features within native genomic environments Functional activity of sequences under experimental conditions
Role in machine learning Learning features associated with regulatory function Direct experimental supervision for reporter activity
Main limitation Observed features do not necessarily establish causal function Native genomic context is not fully reproduced

A conventional observational analysis can be represented conceptually as:

DNA sequence → ATAC / ChIP / Histone marks → Candidate enhancer

MPRA, by contrast, directly measures the relationship:

DNA sequence → Measured reporter activity

This difference is important for machine learning.

However, the term "direct functional training data" must be interpreted carefully.

MPRA provides experimentally measured labels for reporter activity.

It does not automatically provide direct ground-truth labels for endogenous target-gene expression or disease phenotypes.

5. What the Large-Scale 2025 MPRA Study Achieved

5.1 Measuring More Than 680,000 Candidate Regulatory Elements

The 2025 Nature study by Agarwal, Inoue, and colleagues combined large-scale lentiMPRA experiments with deep learning [1].

The study used three human cell lines:

  • HepG2
  • K562
  • WTC11

The researchers measured more than 680,000 candidate cCREs and reported that approximately 41.7% displayed functional regulatory activity under the tested experimental conditions.

They also examined approximately 60,000 shared sequences across all three cell lines to investigate cell-type specificity.

The ability to generate functional measurements at this scale has major implications for genomic machine learning.

However, the 41.7% figure does not mean that 41.7% of all candidate regulatory sequences in the human genome are functional in vivo.

It reflects the particular candidate sequences, cell lines, reporter conditions, and activity criteria used in the study.

5.2 Training Deep Learning Models on MPRA Data

The researchers used the measured data to train models that predict regulatory activity from DNA sequences.

The models evaluated included:

  • MPRAnn
  • MPRALegNet
  • EnformerMPRA
  • SeiMPRA

MPRALegNet is particularly interesting.

Despite being a relatively small CNN with approximately 1.3 million parameters, MPRALegNet achieved competitive performance compared with approaches that leverage representations from much larger pretrained models, including EnformerMPRA and SeiMPRA [1].

The study also found that models using DNA sequence information achieved better predictive performance than models relying exclusively on biochemical features.

These results suggest that simply increasing model size is not the only way to advance genomic AI.

The quality, scale, and biological meaning of experimental training data are also critical determinants of model performance.

5.3 Learning Cell-Type-Specific Regulatory Activity

The same DNA sequence can show different regulatory activities in different cell types.

This can reflect differences in transcription factor expression, regulatory networks, and cellular state.

By testing shared sequences across multiple cell lines, the 2025 study directly compared cell-type-specific regulatory activity [1].

These results demonstrate that regulatory function cannot always be understood from DNA sequence alone.

Cellular context also matters.

Nevertheless, measurements in three cell lines do not establish that the resulting models generalize to all biological settings.

Additional validation is needed in untested primary cells, differentiating cells, patient-derived disease models, and cells exposed to drugs or other stimuli.

5.4 Predicting Variant Effects

The study also evaluated the ability of its models to predict how DNA sequence variants affect MPRA activity.

For MPRALegNet, comparisons with existing MPRA saturation mutagenesis data produced the following Pearson correlations [1].

Region Pearson Correlation
SORT1 0.49
PKLR 0.65
LDLR 0.66
F9 0.51

These results indicate that the model captures a meaningful portion of the variation in experimentally measured reporter activity caused by sequence variants.

However, substantial unexplained variation remains.

More importantly, these evaluations primarily concern variant effects measured in MPRA systems.

They do not establish equivalent accuracy for predicting changes in endogenous gene expression, disease risk, or clinical phenotypes.

5.5 How Well Does MPRA Activity Correspond to Endogenous Enhancer Function?

The 2025 study also investigated the relationship between MPRA activity and functional evidence from CRISPR interference (CRISPRi) experiments [1].

Sequences with higher MPRA activity were more likely to be supported as functional enhancers by CRISPRi.

The highest-activity group showed approximately twice the proportion of CRISPRi-supported enhancers compared with the lowest-activity group.

However, adding MPRA activity to an existing Activity-by-Contact (ABC) model produced only modest improvements in enhancer–target gene prediction performance in two of the three CRISPRi studies evaluated.

This is an important result.

MPRA provides information about the regulatory capacity of DNA sequences.

However, this information does not necessarily add substantial predictive value beyond existing genomic-context features when identifying endogenous enhancer–gene relationships.

Understanding native gene regulation requires more than measuring the isolated activity of a regulatory sequence.

Chromatin state, promoter relationships, and other genomic features remain important.

The study therefore demonstrates both the value of large-scale functional training data and the challenge of translating MPRA measurements into predictions of endogenous regulatory function.

6. The Underlying Assumptions of MPRA

Understanding MPRA requires careful consideration of how its measurements are interpreted biologically.

It is useful to distinguish assumptions needed for the experimental assay itself from assumptions required when using MPRA results to predict endogenous function.

The assumptions discussed here are not necessarily false.

Some are supported by experimental evidence, some hold under particular conditions, and others remain subjects of ongoing investigation.

6.1 Assumption 1: Short DNA Sequences Have Regulatory Capacities That Can Be Measured Independently

A central principle of MPRA is that a DNA sequence retains at least part of its regulatory capacity when removed from its native genomic location.

This idea is related to the concepts of intrinsic regulatory activity and autonomous regulatory activity.

For example, a candidate enhancer fragment of approximately 200 bp can be combined with an artificial promoter to measure how strongly it activates reporter expression.

This makes it possible to compare the regulatory capacities of large numbers of short sequences.

However, the function of a native enhancer is not necessarily determined entirely by a 200 bp fragment.

Neighboring transcription factor binding sites, surrounding sequences, chromatin structure, and interactions with target promoters may all contribute.

Klein and colleagues demonstrated that measuring enhancer fragments of different lengths could produce substantially different activity estimates [14].

Therefore, the apparent intrinsic activity of a DNA sequence may depend on how the sequence boundaries are defined.

The unresolved question is how far regulatory capacity can be treated as an intrinsic property of a DNA sequence that is independent of its surrounding environment.

Typical enhancer MPRAs combine candidate sequences with an artificial minimal promoter.

Consequently, the experiment directly measures the activity of the sequence in combination with that particular promoter.

However, in native genomes, different enhancer–promoter combinations may produce different regulatory outcomes.

A sequence that strongly activates an artificial promoter will not necessarily activate every endogenous promoter in the same way.

Furthermore, the ability of a sequence to act as an enhancer is not equivalent to evidence that it actually regulates a particular endogenous gene.

Therefore, detecting activity in MPRA does not directly establish that the sequence regulates a specific target gene in its native genomic context.

6.3 Assumption 3: Averaging Across Integration Sites Can Estimate Sequence-Dependent Effects

In lentiMPRA, reporter constructs are integrated into chromosomes.

By measuring multiple independent integration events, researchers can reduce the influence of individual integration sites and estimate the activity associated with each tested sequence.

This is a useful strategy, but it has several limitations.

First, integration-site effects do not necessarily disappear completely.

Second, different sequences may respond differently to their chromosomal environments.

For example, sequence A might show strong activity only in accessible chromatin, while sequence B might remain active across a wider range of genomic environments.

Even if their average activities are similar, their biological properties may differ.

Third, averaging across integration sites may conceal information about context dependence.

Therefore, the average activity measured by lentiMPRA should be distinguished from the activity of the sequence at its native genomic locus.

An important unresolved question is whether we can predict not only average regulatory activity but also the genomic environments in which a sequence becomes active.

6.4 Assumption 4: The RNA/DNA Ratio Is an Appropriate Measure of Transcriptional Regulatory Activity

MPRA commonly uses the ratio between RNA and DNA abundance to estimate regulatory activity.

However, measured RNA abundance can be influenced by factors other than transcription rate.

Potential influences include:

  • RNA stability
  • RNA processing and splicing
  • Barcode-related effects
  • Reporter copy number
  • Systematic differences between assay formats
  • Sequencing depth and normalization procedures

In assays such as STARR-seq, where the candidate regulatory sequence is included in the transcript, distinguishing transcriptional effects from post-transcriptional effects can be especially important.

Therefore, MPRA activity is a useful indicator of regulatory output, but it should not automatically be equated with transcription rate alone.

6.5 Assumption 5: Regulatory Rules Learned Under One Cellular Condition Can Generalize to Others

When MPRA is used to generate training data for machine learning, the underlying assumption is that relationships between DNA sequence and measured activity contain patterns that models can learn.

However, regulatory activity may vary with cellular state.

Relevant factors include:

  • Transcription factor abundance
  • Cell cycle
  • Developmental state
  • Cellular stimulation
  • Drug treatment
  • Culture conditions

Consequently, an MPRA-based model is not necessarily learning a universal relationship of the form:

DNA sequence → Regulatory activity

A more accurate conceptual representation is:

DNA sequence + Cell state + Assay conditions → Measured regulatory activity

A model trained under a limited set of conditions may learn only a particular portion of this broader relationship.

This matters when evaluating predictive performance.

High accuracy within a given cell line and assay system does not necessarily imply high accuracy in unfamiliar cell types or biological conditions.

6.6 Assumption 6: Regulatory Grammar Can Be Learned from Short DNA Sequences

MPRA enables systematic investigation of transcription factor motif positions, orientations, spacing, and combinations.

Such analyses contribute to the study of regulatory grammar.

However, native gene regulation also involves interactions among multiple enhancers, functional redundancy, long-range genomic contacts, and dynamic cellular processes.

Therefore, regulatory rules learned from short DNA fragments may not fully explain transcriptional regulation at native genomic loci.

Furthermore, identifying an important motif through model attribution does not itself prove a causal molecular mechanism.

Additional experiments may be needed to determine whether the predicted motif is actually bound by the relevant transcription factor and contributes to the proposed regulatory function.

7. Limitations of MPRA: What Remains Unresolved?

The limitations of MPRA are not simply questions of experimental accuracy.

A more fundamental issue is that the biological quantity MPRA measures differs from the endogenous functions we ultimately want to understand.

7.1 Native Genomic Context Is Not Fully Reproduced

In MPRA, candidate sequences are removed from their native genomic locations and measured using engineered reporter constructs.

Consequently, the assay may not reproduce:

  • Native genomic location
  • Chromatin state
  • Three-dimensional genome structure
  • Surrounding DNA sequences spanning tens of kilobases to megabases
  • Enhancer–promoter contacts
  • Interactions with neighboring regulatory elements
  • Transcription factor abundance
  • Developmental state

lentiMPRA incorporates aspects of chromosomal context, but it does not completely reproduce the native genomic locus [13].

This limitation becomes especially important when extrapolating reporter activity to endogenous gene expression.

7.2 Regulatory Capacity Is Not the Same as Functional Necessity

A sequence that shows activity in MPRA has demonstrated the capacity to alter reporter expression under specified experimental conditions.

However, this does not establish that the sequence is necessary for regulating a particular endogenous gene.

For example, if multiple enhancers regulate the same gene, removing one enhancer may produce only a small effect because of functional redundancy.

Conversely, a sequence that appears weak in an isolated reporter assay may be important within its native genomic context.

MPRA and CRISPR-based approaches therefore address different biological questions.

Technology Primary Question
MPRA Can this DNA sequence alter reporter expression under specific assay conditions?
CRISPRi Does repressing this region at its native genomic locus alter gene expression?
CRISPR-mediated deletion What changes when this region is removed from its native genomic locus?
Base editing What are the effects of specific nucleotide changes in the native genomic context?
Phenotypic assays Do changes in gene regulation alter cellular or organismal properties?

MPRA is useful for measuring regulatory capacity.

CRISPR-based perturbation and complementary experiments are important for establishing causal roles in native genomic environments.

7.3 MPRA Does Not Directly Identify Target Genes

Even when MPRA shows that a sequence has strong regulatory activity, the experiment does not directly identify the endogenous gene regulated by that sequence.

In native genomes, enhancers can act on promoters located far away along the DNA sequence.

The nearest gene is not necessarily the target.

Therefore, MPRA activity alone cannot completely determine enhancer–target gene relationships.

Additional information about chromatin contacts, gene expression, and perturbation experiments is needed.

7.4 Dependence on Sequence Length and Assay Design

MPRAs that use short DNA fragments may fail to capture effects involving surrounding sequences or long-range interactions.

The comparative study by Klein and colleagues showed that sequence length can substantially influence measured activity [14].

Assay design can also produce systematic differences.

When combining MPRA datasets for machine learning, researchers should therefore consider differences in sequence length, promoter selection, cell type, and experimental configuration.

7.5 Candidate Sequence Selection Bias

Although MPRA can measure very large numbers of sequences, the candidate sequences are often selected in advance.

For example, researchers may prioritize regions based on:

  • ATAC-seq
  • DNase-seq
  • Histone modifications
  • Evolutionary conservation

As a result, the training dataset may not represent all genomic sequences equally.

Such selection bias can affect how well a model generalizes to previously untested regions.

7.6 Trade-Offs Between Experimental Scale and Measurement Precision

Larger MPRA libraries allow more sequences to be measured simultaneously.

However, increasing library size can make it more difficult to obtain sufficient measurement depth and independent observations for every sequence.

Measurement uncertainty can also depend on:

  • Number of barcodes
  • Efficiency of cellular delivery
  • Sequence-dependent expression levels
  • Sequencing depth
  • Experimental replication

Therefore, the usefulness of MPRA data for machine learning depends not only on the number of sequences measured but also on reproducibility and measurement uncertainty.

7.7 Summary of Major Limitations

Challenge Main Issue Required Validation
Missing native genomic context Native location and chromatin structure are not reproduced Functional validation at endogenous loci
Target-gene identification Reporter activity does not reveal which genes are regulated Integration with CRISPRi, expression, and 3D genome data
Cellular-state dependence Activity may change across conditions Evaluation across cell types and stimulation conditions
Promoter dependence Activity can depend on promoter choice Testing multiple enhancer–promoter combinations
Sequence-length limitations Surrounding sequences and long-range effects may be missed Comparisons across different fragment lengths
Assay-specific effects Episomal MPRA and lentiMPRA can produce different results Assay standardization and comparative studies
RNA measurement confounding RNA abundance can reflect post-transcriptional processes Assessing measurement design and RNA processing effects
Sequence selection bias Measured sequences may not represent the entire genome Generalization tests on previously unmeasured sequence classes
Measurement noise Results depend on barcodes, depth, and experimental design Quantifying reproducibility and uncertainty
Disease and organism-level generalization Reporter activity is distant from disease phenotypes Validation in patient-derived cells, tissues, and organisms

It is important to distinguish limitations that have already been experimentally demonstrated from questions that remain unresolved.

For example, dependencies on assay format and sequence length have been established through systematic comparisons [13][14].

However, predicting endogenous regulatory function by integrating these different factors remains an active research challenge.

8. The Relationship Between MPRA and Enformer: Functional Training Data Versus Long-Range Context

8.1 What Did Enformer Change?

Enformer, published in 2021, represented an important advance in genomic machine learning [6].

Many earlier CNN-based approaches focused on relatively short DNA sequence windows, often ranging from hundreds to a few thousand base pairs.

Enformer introduced a Transformer-based architecture designed to incorporate much longer-range sequence information.

It processes approximately 200 kb of input DNA and can model contributions from sequences separated by tens of kilobases or more.

This expanded the ability to predict gene expression and chromatin-related features while accounting for long-range regulatory sequence context.

Enformer has also been evaluated on variant effects measured using MPRA saturation mutagenesis experiments.

8.2 Enformer and MPRA Address Different Problems

Enformer and large-scale MPRA are not competing technologies.

They address complementary research challenges.

Technology Main Advance Primary Emphasis
Enformer Modeling long-range genomic sequence context Incorporating information from larger DNA regions
Large-scale lentiMPRA Measuring regulatory activity for large numbers of sequences Producing experimentally measured functional training data

Enformer-like models predict biochemical and gene expression-related measurements from broader genomic sequence contexts.

MPRA measures regulatory capacity under controlled experimental conditions.

MPRA-based models are particularly useful for learning local sequence-dependent features.

Enformer-like models are particularly useful for predictions involving longer genomic regions.

However, processing longer DNA sequences does not automatically mean that a model fully reproduces native chromatin states or the dynamic molecular mechanisms of gene regulation.

An important challenge is determining how local regulatory potential relates to function within a broader genomic environment.

9. Borzoi and AlphaGenome: Expanding the Scope of Genomic Prediction

9.1 Borzoi: Predicting RNA-seq Coverage from DNA Sequence

Borzoi predicts cell-type- and tissue-specific RNA-seq coverage directly from DNA sequences [7].

Its predictions incorporate information associated with multiple stages of RNA regulation, including:

  • Transcription
  • Splicing
  • Polyadenylation

It can also evaluate the predicted effects of DNA sequence variants on these processes.

This represents an expansion from models focused on:

DNA → Chromatin

toward models that also address:

DNA → RNA production / processing

However, predicting RNA-seq coverage does not mean that every causal mechanism governing RNA regulation has been reproduced.

9.2 AlphaGenome: Integrating Long-Range Context and Multiple Genomic Functions

AlphaGenome is a large-scale DNA sequence model published in Nature in 2026 [8].

The model processes approximately 1 Mb of DNA sequence and predicts multiple functional genomic measurements.

Its prediction targets include:

  • Gene expression
  • Transcription initiation
  • Chromatin accessibility
  • Histone modifications
  • Transcription factor binding
  • Chromatin contacts
  • Splicing-related features

AlphaGenome can also be used to evaluate the predicted effects of noncoding variants on these functions.

This represents progress toward integrated prediction of multiple genomic measurements that were previously often modeled separately.

However, AlphaGenome is not a complete simulator of all molecular states and cellular responses.

Its predictions primarily concern genomic functional measurements represented in its training objectives.

Understanding actual cellular states and therapeutic responses requires further experimental data and models operating at additional biological levels.

9.3 Overview of the Technological Landscape

The major approaches can be summarized as follows.

Technology or Model Primary Role
PWM and motif analysis Identifying sequence patterns recognized by transcription factors
DeepBind Learning sequence specificities of DNA- and RNA-binding proteins
DeepSEA Predicting transcription factor binding, DNase sensitivity, and histone modifications
Basset Predicting chromatin accessibility
Basenji Predicting genomic features from DNA sequences
Enformer Incorporating long-range sequence context into genomic predictions
MPRA and lentiMPRA Experimentally measuring sequence-dependent regulatory activity
MPRAnn and MPRALegNet Predicting MPRA activity from DNA sequences
EnformerMPRA and SeiMPRA Applying representations from existing sequence models to MPRA prediction
Borzoi Predicting RNA-seq coverage and RNA-related regulatory features
AlphaGenome Predicting multiple genomic functions from long DNA sequence contexts
PARM Predicting promoter activity, interpreting regulatory grammar, and designing synthetic promoters

These developments should not be understood as a simple sequence in which each new technology replaces its predecessors.

Different research directions have progressed in parallel, emphasizing different prediction targets, sequence lengths, experimental datasets, and degrees of interpretability.

10. Why Removing Genomic Context Can Be Beneficial for Machine Learning

10.1 Native Gene Regulation Involves Multiple Interacting Factors

The loss of genomic context in MPRA is an important limitation.

However, when the goal is to learn regulatory grammar, simplifying the experimental environment can also be advantageous.

Consider an enhancer associated with a highly expressed gene in its native genomic context.

The observed expression level may depend on multiple factors:

Sequence

  • TF concentration

  • Chromatin

  • 3D contact

  • Promoter

  • Neighboring elements

↓

Gene expression

In this situation, it can be difficult to determine how much of the observed expression is attributable to the DNA sequence itself.

10.2 MPRA Simplifies the Learning Problem

MPRA allows researchers to standardize many experimental factors, including reporter design, promoter configuration, and measurement conditions.

This makes it easier to analyze differences attributable to the tested DNA sequences.

Conceptually, the problem becomes closer to:

Sequence → Reporter expression

Cellular state and assay-specific effects still matter.

Nevertheless, controlling many sources of variation creates a more clearly defined supervised learning problem.

This experimental simplification is one of the key reasons MPRA is valuable for machine learning.

10.3 Decoding Regulatory Grammar

Models trained on MPRA data can be analyzed to investigate questions such as:

  • Which transcription factor motifs are important?
  • Does motif orientation matter?
  • How does motif position affect activity?
  • How do motifs interact?
  • Does spacing between motifs influence regulatory activity?
  • What distinguishes promoter and enhancer sequence features?

Model attribution methods can estimate which sequence regions contribute to predicted regulatory activity.

This is part of the study of regulatory grammar.

However, a motif identified as important by a model is not necessarily proven to have the corresponding causal biological function.

Additional experiments may be necessary to validate predicted transcription factor binding and regulatory mechanisms.

Furthermore, regulatory grammar learned from MPRA should first be interpreted as rules that apply under the conditions of the assay.

Whether the same rules operate at native genomic loci is a separate question.

11. Advances in 2026: From Predicting Regulatory Activity to Designing DNA Sequences

11.1 Learning Promoter Activity with PARM

In 2026, the Nature paper "Regulatory grammar in human promoters uncovered by MPRA-based deep learning" was published [9].

The researchers used approximately 10 million MPRA promoter fragments to train a CNN-based model called PARM (Promoter Activity Regulatory Model).

PARM predicts autonomous promoter activity from DNA sequence.

The study also investigated:

  • Position-dependent effects of transcription factor motifs
  • Interactions between motifs
  • Identification of sequences important for transcriptional regulation
  • Design of synthetic promoters

An important aspect of this research is that it examines not only whether a transcription factor motif is present, but also how its position and interactions with other motifs affect regulatory activity.

11.2 Measure → Learn → Interpret → Design

The research process represented by studies such as PARM can be summarized as:

Measure → Learn → Interpret → Design

This involves four stages:

  1. Experimentally measure the activity of DNA sequences.
  2. Train machine learning models using the measured data.
  3. Interpret the models to infer sequence rules associated with transcriptional regulation.
  4. Use the learned rules to design new DNA sequences.

However, the basic idea of using MPRA to optimize regulatory sequences was already demonstrated by Melnikov and colleagues in 2012 [12].

Therefore, sequence design was not introduced for the first time in 2026.

The recent advances lie in using much larger datasets and deep learning to investigate more complex regulatory grammar and support more sophisticated sequence design.

11.3 Will Designed Sequences Function in Native Genomic Environments?

The PARM study demonstrated the design of synthetic promoters with strong activity [9].

It also investigated how promoter regulation changes in response to cellular stimulation and drug treatment.

These are important advances toward the computational design of regulatory DNA.

However, a sequence that performs well in a reporter assay is not necessarily guaranteed to function as intended in native genomic or organismal contexts.

For example, designing an enhancer that is strongly active in hepatocytes but inactive in other cell types requires more than testing a limited set of cultured cells.

Its tissue specificity, safety, and long-term stability must also be evaluated under relevant biological conditions.

As DNA sequence design becomes more sophisticated, validating whether the predicted function transfers from engineered experimental systems to natural biological environments becomes increasingly important.

12. What Is Missing on the Path Toward In Vivo Simulation?

12.1 MPRA Activity and In Vivo Regulatory Function Belong to Different Biological Levels

When considering in vivo simulation, it is useful to distinguish several levels between DNA sequence and organismal phenotype.

Conceptually, these levels can be represented as:

Level 1: Regulatory capacity encoded in DNA sequences

↓

Level 2: Activity within the native genomic environment

↓

Level 3: Changes in endogenous target-gene expression

↓

Level 4: Responses of intracellular molecular networks

↓

Level 5: Cellular, tissue, and organismal phenotypes

MPRA is particularly useful at Level 1.

Models such as Enformer and AlphaGenome extend prediction toward measurements associated with Levels 2 and 3 by incorporating broader genomic contexts.

Borzoi advances prediction of RNA production and processing.

However, explaining how gene expression changes affect protein abundance, metabolism, signaling, cell–cell interactions, tissue function, and disease phenotypes requires additional information.

Therefore, increasing the number of MPRA measurements or improving prediction accuracy for MPRA activity is not sufficient to create a complete in vivo simulator.

12.2 High Accuracy for MPRA Activity Does Not Necessarily Imply High Accuracy In Vivo

When evaluating a machine learning model, it is essential to define what is being used as the ground truth.

Suppose a model predicts MPRA activity with high accuracy.

It may perform well on previously unseen sequences measured using the same MPRA configuration.

However, several other questions remain:

  • Can the model predict MPRA activity in an untested cell type?
  • Can it predict regulatory function at the native genomic locus?
  • Can it predict changes in endogenous target-gene expression?
  • Can it predict how noncoding variants influence disease risk?
  • Can it predict whether a designed regulatory sequence will work in vivo for therapeutic purposes?

Each of these questions requires separate validation.

The 2025 study by Agarwal, Inoue, and colleagues demonstrated associations between MPRA activity and endogenous enhancer function, while also showing that MPRA activity does not necessarily produce substantial improvements in target-gene prediction [1].

This illustrates why different levels of prediction must be distinguished.

12.3 Using MPRA Measurements as Intermediate Functional Information

The following is a conceptual research perspective based on the findings discussed above.

Rather than treating MPRA measurements as the final ground truth for in vivo regulatory function, we can consider them as intermediate experimental information describing the regulatory capacity of DNA sequences.

Conceptually, the relationship can be expressed as:

[ A = f_{\mathrm{MPRA}}(S,C,R) ]

[ Y = g(A,G,P,X) ]

where:

  • (S): DNA sequence
  • (C): Cellular state
  • (R): Reporter assay conditions
  • (A): Reporter activity measured by MPRA
  • (G): Native genomic and chromatin context
  • (P): Relationship with the target promoter
  • (X): Other cellular and environmental factors
  • (Y): Effect on gene regulation in the native genomic environment

These equations are not established biological laws. They are conceptual models intended to clarify the research problem.

Furthermore, a single MPRA activity value may not contain all the information required to explain endogenous regulatory function.

Nevertheless, explicitly modeling the distinction between MPRA-measured activity and native regulatory function could provide a valuable direction for future research.

13. Future Research Questions: What Has MPRA Not Yet Established?

Further progress in MPRA will require more than simply increasing the number of sequences measured.

A central research goal should be to determine the biological conditions under which MPRA measurements generalize to native genomic functions.

13.1 Validate the Relationship Between MPRA Activity and Endogenous Regulatory Function

One of the most important priorities is to compare MPRA measurements with functional effects at native genomic loci.

For example, researchers could test whether sequences showing strong MPRA activity also affect endogenous target-gene expression when perturbed by CRISPRi.

They could also examine whether variant effects predicted from MPRA are reproduced when the same nucleotide changes are introduced into native genomic loci using base editing.

Such experiments could help identify conditions under which MPRA-based predictions are reliable and conditions under which they fail.

13.2 Learn the Conditions Under Which MPRA Predictions Fail

Another important direction is to study MPRA prediction errors themselves.

For example, a sequence might show strong reporter activity in MPRA but have little measurable effect on a target gene at its native genomic locus.

Possible explanations include:

  • Chromatin state
  • Three-dimensional genome structure
  • Contact with the target promoter
  • Cooperation or redundancy among neighboring regulatory elements
  • Cellular state
  • Sequence length

By collecting such discrepancies systematically, it may become possible to learn the conditions under which a sequence's potential regulatory capacity becomes an actual biological function.

This approach would go beyond correcting MPRA limitations.

It could contribute to a more general understanding of context-dependent gene regulation.

13.3 Test Generalization to Unmeasured Cell Types and Disease Conditions

The large-scale 2025 MPRA study used three cell lines [1].

However, biological organisms contain many cell types and cellular states.

Regulatory networks may also differ between healthy and diseased cells.

Future validation should therefore examine conditions such as:

  • Previously untested cell types
  • Primary cells
  • Differentiating cells
  • Patient-derived disease cells
  • Drug-treated cells
  • Organoids

It is particularly important to evaluate prediction accuracy using independent data from cell types and disease conditions not represented during training.

13.4 Investigate Enhancer–Promoter Compatibility

MPRA often uses artificial minimal promoters.

However, enhancer–promoter compatibility may influence regulatory activity in native genomes.

An important research direction is therefore to measure the same enhancer sequences in combination with multiple promoters.

Such experiments could help distinguish regulatory capacities shared across promoters from functions that depend on specific enhancer–promoter combinations.

13.5 Determine the Added Value of MPRA Data for Long-Range Sequence Models

Models such as Enformer and AlphaGenome predict genomic functions from long DNA sequence contexts.

MPRA, in contrast, provides large-scale experimental information about the regulatory capacity of shorter sequences.

Combining these approaches may improve prediction.

However, the size and nature of any improvement must be evaluated experimentally.

The key question is not simply whether MPRA activity can be predicted more accurately.

It is whether including MPRA-derived information improves predictions of endogenous gene expression, regulatory interactions, and variant effects.

13.6 Validate the In Vivo Function of AI-Designed Regulatory DNA

MPRA-based machine learning can be used to design DNA sequences with desired regulatory activity.

Potential applications include promoters or enhancers that are strongly active in particular cell types.

However, successful activity in a reporter assay does not guarantee equivalent performance in vivo.

For therapeutic applications, important questions include:

  • Does the sequence function in the intended cell type?
  • Does it cause unwanted activity in other cell types?
  • Can it maintain stable expression over time?
  • How does it respond to changes in the biological environment?
  • Does it present any safety concerns?

13.7 Prioritizing Future Research

From the perspective of in vivo simulation and drug discovery, the following priorities are particularly important.

Priority Research Question Proposed Validation
1 How well does MPRA activity predict endogenous regulatory function? Comparing MPRA, CRISPRi, and base editing
2 Under which conditions do MPRA predictions fail? Integrating chromatin, 3D contact, and surrounding sequence information
3 Can predictions generalize to new cell types and disease conditions? Validation in primary cells, patient-derived cells, and organoids
4 How important is enhancer–promoter compatibility? MPRA experiments testing multiple enhancer–promoter combinations
5 Do AI-designed sequences function as intended in vivo? Validation in independent assays, tissues, and organisms
6 Does MPRA improve long-range genomic models? Evaluating the added value of MPRA information on independent functional prediction tasks

These priorities are a research-oriented interpretation emphasizing progress toward in vivo simulation. They are not a ranking established by the cited studies.

14. Applications in Drug Discovery and Synthetic Biology

14.1 From Genome Annotation to Sequence Design

The objectives of genomic research have expanded over time.

An early focus was:

Genome → Annotation

The goal was to identify genes and regulatory regions within DNA sequences.

With advances in machine learning, another direction became increasingly important:

Genome → Prediction

This includes predicting transcription factor binding, chromatin states, gene expression, and related functional measurements.

MPRA then made it possible to learn relationships such as:

Genome sequence → Experimentally calibrated function

using large-scale functional measurements.

Today, researchers are increasingly investigating the inverse problem:

Desired function → Sequence design

The goal is no longer only to predict the function of a given DNA sequence, but also to design sequences with desired biological properties.

14.2 Designing DNA Sequences with Desired Expression Characteristics

A traditional research question might be:

"How strongly does this DNA sequence promote gene expression?"

Sequence design introduces a different type of question:

"Can we design an enhancer that is strongly active in hepatocytes but inactive in other cell types?"

Other possibilities include promoters that respond to specific stimuli or regulatory sequences that activate only under particular cellular conditions.

Studies such as PARM are advancing this direction [9].

However, regulatory activity or cell-type specificity demonstrated in a reporter assay must still be validated in relevant tissues and organisms.

14.3 Potential Applications in Drug Development

Advances in MPRA and genomic AI could contribute to several fields, including:

  • Promoter and enhancer design for gene therapy
  • Development of cell-type-specific vectors
  • Synthetic biology
  • Therapeutic control of gene expression
  • Functional evaluation of noncoding variants

Agarwal, Inoue, and colleagues discussed the potential value of large-scale regulatory activity datasets for regulatory sequence design and therapeutic delivery applications [1].

The PARM study also investigated changes in promoter regulation associated with cellular stimulation and drug treatment [9].

In the future, these technologies may support studies using patient-derived cells and organoids to better understand disease progression, treatment responses, and potential drug targets.

However, potential applications must be distinguished from demonstrated clinical efficacy.

High predictive accuracy for MPRA activity does not directly establish the ability to predict therapeutic outcomes.

Applications in drug discovery will require additional validation of endogenous regulatory function, cellular responses, disease causality, and safety.

15. Conclusion: MPRA Has Transformed Functional Training Data, but Connecting It to In Vivo Biology Is the Next Challenge

The 2025 Nature study "Massively parallel characterization of transcriptional regulatory elements" by Agarwal, Inoue, and colleagues represents an important advance in genomic AI.

Its significance lies not simply in scaling up neural network architectures, but in expanding the scale and quality of experimentally measured functional data from which models can learn relationships between DNA sequences and regulatory activity.

However, a historical examination of MPRA reveals a fundamental limitation inherited from conventional reporter assays.

This limitation concerns how closely the regulatory capacity measured under engineered experimental conditions corresponds to the function of the same sequence in its native genomic environment.

MPRA simplifies the regulatory environment to make sequence-dependent effects easier to compare.

This simplification enables large-scale learning of relationships between DNA sequence and regulatory activity.

At the same time, information about chromatin structure, enhancer–promoter interactions, cellular state, and surrounding regulatory elements may be lost or altered.

Comparative studies published in 2017 and 2020 demonstrated that assay configuration and sequence length can influence the measured activity [13][14].

The large-scale 2025 study substantially advanced the measurement and prediction of regulatory sequence activity, but also showed that the additional value of MPRA data for predicting endogenous target-gene relationships can be limited in some settings [1].

These findings do not diminish the importance of MPRA.

Instead, they highlight the need to distinguish the biological quantity MPRA measures from the endogenous regulatory functions we ultimately want to understand.

At least two major directions are driving the current development of genomic AI.

The first direction is predicting diverse genomic functions while incorporating long-range sequence context.

Enformer, Borzoi, and AlphaGenome represent advances along this direction.

The second direction is learning regulatory grammar from functional experiments and using that knowledge to design DNA sequences with desired activities.

Large-scale MPRA studies, MPRALegNet, and PARM are advancing this direction.

These approaches are complementary, and connecting them will be increasingly important.

However, the central challenge is not simply to combine long-range sequence models with MPRA datasets.

It is to understand how the regulatory potential measured by MPRA becomes an actual biological function under specific genomic, cellular, and promoter contexts.

Addressing this challenge will require integrating MPRA with CRISPR-based perturbation experiments at native genomic loci, chromatin measurements, gene expression data, and functional studies at cellular and tissue levels.

It will also be important to systematically investigate not only when MPRA predictions succeed, but also when and why they fail.

From this perspective, the limitations of MPRA are more than technical shortcomings.

They point toward a fundamental question in gene regulation: under what conditions does the regulatory potential encoded in DNA become an actual biological function?

MPRA is not a technology that fully reproduces gene regulation in vivo.

Nevertheless, it provides an essential experimental foundation for understanding how DNA sequences contribute to regulatory function.

For future progress in in vivo simulation and drug discovery, an important research direction will be connecting large-scale functional measurements from MPRA with causal validation of gene regulation in native biological contexts.

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] Agarwal, V., Inoue, F., Schubach, M. et al., "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

[12] Melnikov, A., Murugan, A., Zhang, X. et al., "Systematic dissection and optimization of inducible enhancers in human cells using a massively parallel reporter assay," Nature Biotechnology, 30, 271–277, 2012. https://doi.org/10.1038/nbt.2137

[13] Inoue, F., Kircher, M., Martin, B. et al., "A systematic comparison reveals substantial differences in chromosomal versus episomal encoding of enhancer activity," Genome Research, 27, 38–52, 2017. https://doi.org/10.1101/gr.212092.116

[14] Klein, J. C., Agarwal, V., Inoue, F. et al., "A systematic evaluation of the design and context dependencies of massively parallel reporter assays," Nature Methods, 17, 1083–1091, 2020. https://doi.org/10.1038/s41592-020-0965-y