MPRA: Implicit Assumptions, Limitations, and Open Questions

Target Audience#

• Researchers and readers interested in Massively Parallel Reporter Assays (MPRAs), their underlying principles, historical development, and current research challenges.
• Those interested in machine learning for genomics, including predicting gene expression and regulatory activity from DNA sequences.
• Those exploring the potential applications of AI in biological simulation, gene therapy, drug discovery, and synthetic biology.

Summary#

• MPRA enables large-scale, quantitative measurements of the transcriptional regulatory activity of DNA sequences under controlled experimental conditions, providing functional training data for genomic AI.
• MPRA relies on several implicit assumptions, including the idea that short DNA sequences retain measurable regulatory activity outside their native genomic contexts and that reporter activity is related to endogenous gene regulation.
• MPRA measurements depend on experimental design, sequence length, promoter choice, chromatin environment, and cellular state. Reporter activity therefore cannot be treated as equivalent to endogenous regulatory function.
• A large-scale MPRA study published in 2025 measured more than 680,000 candidate regulatory sequences and developed machine learning models for predicting regulatory activity and variant effects. However, important questions remain about predicting endogenous target-gene regulation and generalizing to untested cellular conditions.
• Future progress will require understanding when and how regulatory potential measured by MPRA translates into functional effects in native genomic environments, integrating MPRA with CRISPR-based experiments and long-range genomic models.

1. Introduction: What Can MPRA Reveal, and What Remains Unknown?#

Massively Parallel Reporter Assays (MPRAs) are experimental technologies that enable the simultaneous measurement of transcriptional regulatory activity for thousands to hundreds of thousands of DNA sequences or more.

In recent years, large-scale MPRA datasets have increasingly been used to train machine learning models that predict transcriptional regulatory activity from DNA sequences and design sequences with desired regulatory properties.

A representative example is the 2025 Nature paper by Agarwal, Inoue, and colleagues, "Massively parallel characterization of transcriptional regulatory elements" [1].

In this study, the researchers measured more than 680,000 candidate cis-regulatory elements using lentiMPRA and used the resulting data to develop models that predict regulatory activity and the effects of sequence variants.

This development has established MPRA as more than an experimental measurement technique. It has become an important source of functional training data for genomic AI.

However, a fundamental question remains:

Is the activity of a DNA sequence measured under artificial experimental conditions equivalent to its regulatory function in a living biological system?

The fundamental purpose of MPRA is not to reproduce endogenous gene regulation in its entirety.

Instead, MPRA isolates selected components of a complex regulatory environment, making it easier to measure the contribution of individual DNA sequences.

In this sense, the strengths and limitations of MPRA are closely connected.

Conceptually, there is an important distinction between what MPRA measures and the biological functions researchers ultimately want to understand.

What MPRA measures

DNA sequence + Reporter construct + Experimental cell

↓

MPRA activity

↓

RNA output observed under experimental conditions

What researchers ultimately want to understand

DNA sequence + Native genomic location + Chromatin + 3D genome structure + Cellular state

↓

Endogenous gene regulatory function

↓

Effects on target genes, cellular responses, and phenotypes

A major unresolved problem is how to generalize from the first system to the second.

Five questions are particularly important:

1. Sequence autonomy: Does a short DNA sequence have regulatory activity that can be measured independently, at least to some extent, of its surrounding genomic environment?
2. Contextual reproducibility: Does activity measured using artificial promoters or random genomic integration sites reflect activity at the sequence's native genomic location?
3. Meaning of the measurement: How accurately does the RNA/DNA ratio represent transcriptional regulatory activity?
4. Machine learning generalization: Can models trained on MPRA measurements accurately predict regulatory activity in previously unseen cell types, disease states, or biological conditions?
5. Causality and therapeutic relevance: Can the ability of a sequence to activate a reporter be translated into reliable predictions of endogenous target-gene regulation or disease phenotypes?

These are not merely technical limitations. They represent fundamental scientific questions that genomic AI must address.

Comparative studies of MPRA technologies have demonstrated systematic differences in measured activity depending on assay design and sequence length [2].

Recent reviews have also identified the inability to fully reproduce native chromatin environments and long-range regulatory interactions as major limitations of MPRA [3].

This article examines the historical development of MPRA, its implicit assumptions, its experimental and biological limitations, and the questions that remain insufficiently validated.

It also considers future research directions for biological simulation and drug discovery.

2. The Historical Position of MPRA: What Problem Was It Designed to Solve?#

Understanding MPRA requires examining not only the development of AI models such as DeepSEA and Enformer but also the broader history of experimental efforts to establish causal relationships in gene regulation.

2.1 From Conventional Reporter Assays to MPRA#

Since the 1980s, reporter assays have been widely used to investigate how specific DNA sequences affect gene expression.

In a conventional reporter assay, a DNA sequence of interest is combined with a reporter gene, and the resulting expression level is measured.

For example, researchers can compare reporter expression with and without a particular DNA sequence to determine whether that sequence has the capacity to increase or decrease transcription.

However, conventional reporter assays generally require sequences to be tested individually or in relatively small groups.

During the 2000s, technologies such as ChIP-seq and analyses of evolutionary conservation enabled researchers to identify large numbers of candidate regulatory regions across the genome.

The challenge then became determining which of these candidates actually possessed regulatory activity.

Testing such large numbers of sequences using conventional reporter assays was impractical.

MPRA emerged as a solution to this scalability problem.

By linking DNA sequences to identifiable barcodes and using next-generation sequencing, MPRA allows the activities of many sequences to be quantified in parallel.

2.2 2012–2013: Scaling Functional Measurements with MPRA#

One of the foundational MPRA studies was published by Melnikov and colleagues in Nature Biotechnology in 2012 [4].

The researchers analyzed more than 27,000 enhancer variants, enabling large-scale quantitative measurements of the relationship between DNA sequence and transcriptional regulatory activity.

They also used the experimental results to construct quantitative models and optimize regulatory sequences.

This historical point is important.

The integration of MPRA, computational modeling, and regulatory sequence design did not begin with recent advances in deep learning.

The general research strategy of measuring large numbers of sequences, learning regulatory rules from experimental data, and using those rules to design improved sequences was already being explored in 2012.

2.3 2015–2017: Deep Learning and the Emergence of lentiMPRA#

Around 2015, deep learning models such as DeepSEA demonstrated the ability to predict transcription factor binding, chromatin accessibility, and histone modifications directly from DNA sequences [5].

At the same time, MPRA research increasingly focused on how experimental conditions influenced the measured activity of regulatory elements.

A particularly important issue was that conventional MPRA constructs could remain outside chromosomes.

In such episomal assays, DNA sequences may be evaluated in chromatin environments that differ substantially from those found at endogenous genomic locations.

To address part of this problem, lentiMPRA was developed.

LentiMPRA uses lentiviral delivery to integrate reporter constructs into the genome.

In 2017, Inoue and colleagues compared the activity of 2,236 candidate enhancers in episomal and chromosomally integrated contexts [6].

They found substantial differences between the two measurement systems.

This study provided important evidence that measured regulatory activity depends on the experimental environment in which a sequence is tested.

2.4 2020–2021: Assay Context Dependencies and Long-Range Genomic Models#

In 2020, Klein and colleagues published a systematic comparison of nine MPRA designs [2].

The study evaluated 2,440 sequences and investigated how assay configuration and sequence length affected measured regulatory activity.

Although substantial agreement was observed across many assay designs, the researchers also identified systematic differences.

For example, comparing enhancer fragments of 192, 354, and 678 base pairs revealed that measured activity could depend strongly on the length of the tested sequence.

Meanwhile, genomic machine learning advanced toward modeling longer-range regulatory interactions.

In 2021, Enformer introduced an approach for predicting gene expression and chromatin-related measurements while incorporating long-range DNA sequence context [7].

Two complementary research directions were therefore developing in parallel:

• MPRA: Experimentally measuring the regulatory activity of individual DNA sequences under controlled conditions and investigating how measurement results depend on assay context.
• Long-range models such as Enformer: Incorporating broader genomic sequence context to predict more complex regulatory outcomes.

2.5 2025: Large-Scale MPRA Meets Machine Learning#

In 2025, Agarwal, Inoue, and colleagues published a large-scale lentiMPRA study examining more than 680,000 candidate regulatory sequences [1].

The study used three human cell systems:

• HepG2
• K562
• WTC11

The researchers measured sequence-specific regulatory activity and used the resulting dataset to train machine learning models that predict regulatory activity and variant effects.

The significance of this work was not the invention of MPRA itself.

Its contribution was the creation of a large-scale functional dataset and the demonstration that relationships between DNA sequence and transcriptional regulatory activity could be learned from these measurements.

2.6 2026: From Prediction and Design Toward Validation in Endogenous Contexts#

By 2026, research on MPRA-based sequence design and long-range genomic modeling had continued to advance.

At the same time, important questions remained concerning enhancer–target gene relationships, enhancer–promoter compatibility, and regulatory function at native genomic locations [8][9].

This historical perspective highlights three important points.

First, MPRA was not created as a consequence of deep learning.

It originated as a massively parallel extension of conventional reporter assays.

Second, not all MPRA limitations emerged as experiments became larger.

The question of whether reporter activity reflects endogenous gene regulation is an inherent challenge of reporter-based experimental systems.

Third, modern MPRA research is not concerned exclusively with increasing the number of sequences measured.

It also seeks to determine what MPRA measurements actually represent and under which conditions they correspond to biological function.

Historically, MPRA is therefore best understood not simply as a technology for directly measuring endogenous regulatory function, but as an experimentally controlled, quantitative platform for investigating relationships between regulatory DNA sequences and transcriptional activity.

3. The Implicit Assumptions Underlying MPRA#

The implicit assumptions of MPRA can be examined from three perspectives:

• Experimental biology
• Measurement theory
• Machine learning

Here, an "implicit assumption" does not mean that researchers necessarily accept the assumption without criticism.

Rather, it refers to a condition that must hold, at least to some extent, when MPRA measurements are used to infer or predict regulatory functions in native genomic environments.

3.1 Assumption 1: DNA Sequences Have Measurable Regulatory Activity Independent of Their Surrounding Context#

One of the central ideas behind MPRA is that a short DNA sequence, for example approximately 200 base pairs, can retain at least part of its regulatory capacity when removed from its native genomic location.

This property is commonly described as autonomous regulatory activity.

However, endogenous enhancer activity is influenced by much more than the local DNA sequence.

Relevant factors include:

• Neighboring DNA sequences
• Chromatin state
• Target promoters
• Three-dimensional genome contacts
• Transcription factor abundance and activity

MPRA therefore does not demonstrate that regulatory function is completely independent of genomic context.

Instead, it demonstrates that reproducible differences in regulatory activity can be observed among sequences when selected experimental conditions are controlled.

In the 2020 study by Klein and colleagues, enhancer fragments of 192, 354, and 678 base pairs produced substantially different activity measurements in some cases [2].

This illustrates that the boundaries chosen for a regulatory sequence can affect the measured result.

Open question: To what extent can the regulatory capacity of a DNA sequence be defined as an intrinsic property independent of its genomic environment?

3.2 Assumption 2: Activity in an Artificial Reporter System Is Related to Endogenous Gene Regulation#

In a typical enhancer MPRA, the sequence of interest is combined with an artificial minimal promoter and a reporter construct.

What the experiment directly measures is how much the tested sequence changes reporter expression when paired with that particular promoter.

However, enhancer–promoter relationships are not necessarily universal.

An enhancer that strongly activates one promoter may not produce the same effect when paired with another.

Research reported in 2026 has also investigated how enhancer–promoter compatibility measurements can depend on reporter configuration [8].

Open question: If a DNA sequence strongly activates an artificial promoter, will it also strongly activate a different promoter in its native genomic environment?

3.3 Assumption 3: Random Integration Effects Can Be Averaged Out in lentiMPRA#

LentiMPRA integrates reporter constructs into chromosomes using lentiviral delivery.

By measuring the activity of a sequence across many independent integration events, researchers aim to estimate its average regulatory activity while reducing the influence of any particular integration site.

This is a useful approach, but averaging does not mean that genomic position effects disappear.

Consider two hypothetical sequences:

• Sequence A: Strongly active only in accessible chromatin environments.
• Sequence B: Consistently active across a broad range of chromatin environments.

The two sequences might have similar average reporter activity while exhibiting fundamentally different biological properties.

Furthermore, if integration sites or copy numbers are unevenly distributed, the resulting average may not represent a universal regulatory activity.

Averaging may also obscure how sensitive individual sequences are to their genomic environment.

Open question: Can MPRA characterize not only the average activity of a sequence but also the variability of its activity and its sensitivity to genomic context?

3.4 Assumption 4: The RNA/DNA Ratio Accurately Represents Transcriptional Regulatory Activity#

MPRA activity is generally quantified by normalizing reporter RNA abundance to the corresponding DNA abundance.

Conceptually, this can be expressed 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 MPRA activity.

The RNA/DNA ratio is a practical measure for comparing regulatory activities while accounting for differences in DNA representation.

However, observed RNA abundance is not determined by transcription rate alone.

Other possible influences include:

• RNA stability
• RNA processing and splicing
• Barcode sequence and placement
• Reporter construct design
• DNA copy number
• Technical biases associated with the measurement system

In STARR-seq-based assays, where the tested regulatory sequence is included in the transcript, distinguishing transcriptional activation from post-transcriptional effects is particularly important.

The 2020 comparative study demonstrated that RNA stability and splicing-related factors can influence results in some assay configurations [2].

Open question: What proportion of the observed changes in RNA abundance reflects transcriptional activation itself, rather than other processes affecting RNA production or stability?

3.5 Assumption 5: The Relationship Between DNA Sequence and Regulatory Activity Can Be Learned Under Controlled Cellular Conditions#

For MPRA measurements to serve as machine learning training data, there must be reproducible relationships between DNA sequences and regulatory activity under the experimental conditions used.

However, transcriptional regulation depends on cellular factors such as:

• Transcription factor concentrations
• Cell cycle state
• Cellular differentiation
• External stimulation
• Drug treatment
• Culture conditions

Changing these factors may alter the regulatory activity of the same DNA sequence.

Consequently, the learned relationship is not necessarily a universal function of DNA sequence alone:

[ f(\mathrm{DNA}) ]

Instead, it may represent a particular slice of a more general context-dependent relationship:

[ f(\mathrm{DNA},\ \mathrm{cell\ state},\ \mathrm{assay}) ]

Here, cell state represents the cellular environment, while assay represents the experimental configuration.

Open question: How well can regulatory rules learned under a limited set of cellular conditions generalize to previously untested cell states, stimuli, or drug treatments?

3.6 Assumption 6: Regulatory Rules Learned from Short DNA Sequences Can Explain More Complex Gene Regulation#

Large-scale MPRA datasets enable machine learning models to identify DNA sequence patterns associated with transcriptional activity.

These patterns may include:

• Types of transcription factor binding motifs
• Motif positions
• Motif orientations
• Spacing between motifs
• Combinations of multiple motifs

Such sequence-dependent rules are often described as regulatory grammar.

However, MPRA primarily provides information about local regulatory grammar within the tested DNA fragments.

Endogenous gene regulation also depends on broader mechanisms, including:

• Cooperative interactions among multiple enhancers
• Functional redundancy between enhancers
• Selection of target promoters
• Higher-order chromatin structure
• Temporal changes in cellular state

Furthermore, identifying important motifs through model attribution does not establish their causal molecular mechanisms.

A machine learning model may achieve strong predictive performance without necessarily learning a biologically accurate representation of the underlying regulatory process.

Open question: Do the sequence rules discovered by machine learning models represent genuine biological mechanisms, or are they statistical features that are useful primarily for predicting a particular experimental dataset?

4. A Systematic Overview of MPRA Limitations and Challenges#

MPRA limitations can be organized into three broad categories:

• Experimental constraints
• Biological limitations
• Consequences for machine learning

The following priorities reflect their relevance to biological simulation and drug discovery.

Challenge	Specific limitation	Priority
Loss of native genomic context	Native genomic position, chromatin environment, and 3D contacts are not fully reproduced.	Critical
Target-gene identification	Reporter activity alone does not identify which endogenous genes are regulated.	Critical
Limited cellular conditions	Activity measured in a particular cultured cell system may not generalize to other cells or tissues.	Critical
Enhancer–promoter dependence	Measured activity depends on promoter selection and reporter construct design.	High
Sequence length constraints	Long-range interactions and contributions from neighboring DNA sequences may be missed.	High
Assay-specific biases	Episomal MPRA, lentiMPRA, and STARR-seq can produce different results.	High
Confounding in RNA measurements	RNA processing, copy number, and barcode-related effects can influence measurements.	High
Candidate selection bias	Tested sequences are often enriched for candidates identified using ATAC-seq, DNase-seq, or similar methods.	High
Measurement depth and noise	Increasing the number of tested sequences may reduce measurement precision per sequence.	Medium–High
Generalization to organisms and diseases	Predictive accuracy in animals, patients, and diseased tissues may remain unestablished.	Critical

It is important to distinguish between limitations inherent to MPRA as an experimental methodology and questions that remain unresolved in particular large-scale MPRA studies.

4.1 Differences Between Assay Designs Have Already Been Demonstrated#

In 2017, Inoue and colleagues compared the activity of 2,236 candidate enhancers in episomal and chromosomally integrated reporter systems [6].

They found substantial differences between the two contexts.

In 2020, Klein and colleagues compared nine MPRA designs using 2,440 sequences [2].

Although many designs produced partially consistent measurements, systematic differences were observed.

Sequence length was also found to have a substantial influence on measured activity.

A comparative study published in 2025 further reported disagreements in enhancer activity classification across MPRA datasets, partly associated with differences in experimental and analytical procedures [10].

These results suggest that MPRA activity should not be treated as a single, assay-independent physical quantity.

When comparing results across studies or combining MPRA datasets for machine learning, differences in assay design must be considered.

4.2 A 200-Base-Pair Fragment Is Not Necessarily a Complete Enhancer#

MPRA experiments frequently examine relatively short DNA sequences, sometimes around 200 base pairs.

However, detecting regulatory activity in a short fragment does not establish that the fragment contains all the information required for endogenous enhancer function.

Longer enhancers may include:

• Multiple transcription factor binding sites
• Cooperative sequence elements
• Repressive motifs
• Interactions with neighboring sequences

Short fragments may capture only part of these regulatory interactions.

Conversely, a sequence that is repressed in its native genomic environment may become active in an artificial reporter system.

Two important consequences follow:

• A negative MPRA result does not necessarily mean that a sequence lacks regulatory function in its native genomic context.
• A positive MPRA result does not demonstrate that a sequence contributes to endogenous target-gene expression.

These limitations are also highlighted in reviews of MPRA technologies [3].

4.3 Having Regulatory Capacity Is Not the Same as Being Functionally Necessary#

This distinction is particularly important.

If a DNA sequence increases reporter expression, the result indicates that the sequence has the capacity to promote transcription under the tested conditions.

This can be considered evidence related to sufficiency.

However, it does not establish that the sequence is necessary for gene regulation at its native genomic location.

For example, multiple enhancers may provide redundant regulatory functions.

In such cases, removing one enhancer may produce little or no detectable change in target-gene expression.

Conversely, a sequence that does not activate a reporter on its own may play an important role within the native genomic architecture.

Different experimental technologies therefore address different causal questions.

Technology	Primary question addressed
MPRA	Can a DNA sequence alter reporter expression under specified experimental conditions?
CRISPRi, genomic deletion, or base editing	Does the sequence function at its native genomic location, and how does perturbing it affect endogenous gene expression?
Phenotypic analysis	Do changes in gene regulation produce biological consequences at the cellular or organismal level?

MPRA and CRISPR-based experiments are therefore not competing technologies.

They are complementary approaches that provide evidence about different aspects of regulatory causality.

Integrating MPRA with CRISPR-based functional validation is important because reporter activity alone cannot fully explain the necessity of a sequence or its relationship with endogenous target genes [3].

5. What Was Validated in the 2025 Large-Scale MPRA Study, and What Remains Unresolved?#

The 2025 Nature study by Agarwal, Inoue, and colleagues provides an important case study for distinguishing experimentally validated results from unresolved questions [1].

5.1 Major Validated Results#

The study used lentiMPRA to measure regulatory activity for more than 680,000 candidate regulatory sequences.

Its main quantitative results can be summarized as follows.

Item	Result
Candidate regulatory sequences tested	More than 680,000
Proportion classified as active	Approximately 41.7%
Human cell systems	HepG2, K562, and WTC11
Shared sequences compared across cell systems	Approximately 60,000

The reported 41.7% represents the proportion classified as active under the study's experimental criteria.

It does not mean that 41.7% of all human regulatory sequences are functional.

The study also evaluated:

• Reproducibility of MPRA measurements
• Agreement with known regulatory annotations
• Cell-type-specific regulatory activity
• Prediction of regulatory activity from DNA sequence
• Prediction of variant effects on MPRA activity

These findings demonstrate that large-scale functional data can be used to learn relationships between DNA sequence and transcriptional regulatory activity.

However, evaluating their relevance to endogenous gene regulation reveals important limitations.

5.2 Problem 1: Adding MPRA Activity Produced Limited Improvements in Predicting Endogenous Enhancer–Promoter Relationships#

The study investigated whether MPRA activity could improve predictions of endogenous enhancer–promoter relationships using CRISPRi experimental results [1].

The researchers found that sequences with higher MPRA activity were more likely to be supported as functional enhancers in CRISPRi experiments.

In the highest-activity group, the proportion of CRISPRi-supported enhancers was approximately twice that of the low-activity group.

However, when MPRA activity was added to the existing Activity-by-Contact (ABC) model, improvements in discriminative performance were small in two of the three CRISPRi studies examined.

One possible explanation discussed in the paper is that H3K27ac information already used by the ABC model may capture not only local regulatory activity but also aspects of the surrounding regulatory environment.

This is an important result.

Although MPRA can measure regulatory activity, those measurements do not necessarily provide substantial additional predictive information about endogenous target-gene relationships beyond what is already captured by genomic context features.

In other words, collecting more detailed information about the regulatory capacity of individual sequences does not automatically lead to substantially better predictions of endogenous gene regulation.

5.3 Problem 2: Variant Effect Prediction Improved but Remained Imperfect#

The same study evaluated variant effect predictions produced by MPRALegNet [1].

Predictions were compared against previously measured MPRA saturation mutagenesis data for four genomic regions.

The reported Pearson correlations were:

Region	Pearson correlation (r)
SORT1	0.49
PKLR	0.65
LDLR	0.66
F9	0.51

These results indicate that the model captured some of the effects of DNA sequence variants on MPRA activity.

However, correlations ranging from 0.49 to 0.66 also indicate that substantial variation remained unexplained.

More importantly, this evaluation primarily concerned the prediction of variant effects measured by MPRA.

It does not demonstrate equivalent predictive accuracy for endogenous target-gene expression changes or disease risk.

Predicting variant effects on reporter activity and predicting their consequences in native genomic environments are distinct problems.

5.4 What the Study Did Not Fully Establish#

The following table distinguishes outcomes evaluated in the 2025 study from questions that remain insufficiently validated.

Validation target	What can be concluded from the study
Reporter activity across large numbers of sequences	Measured and evaluated at large scale.
Measurement reproducibility	Evaluated within and across experimental replicates.
Differences between cell systems	Partially validated across three human cell systems.
Variant effects on MPRA activity	Evaluated using selected existing experimental datasets.
Effects on endogenous target genes	Compared with CRISPRi and related evidence, but not comprehensively established.
Generalization to unseen primary or disease-related cells	Not sufficiently validated.
Therapeutic effects in living organisms	Not directly evaluated by this study.
Prediction of organism-level disease phenotypes	Not established.

The study did include comparisons relevant to endogenous biological function.

It would therefore be incorrect to say that it completely ignored the relationship between MPRA measurements and native genomic regulation.

However, the scope of these evaluations was limited.

The study's central contribution was the construction of a large-scale functional regulatory dataset.

Determining how far that dataset can support predictions of endogenous gene regulation, cellular responses, or disease phenotypes remains an important research challenge.

6. What Is Missing for Genomic AI and Biological Simulation?#

Combining MPRA with long-range genomic models such as Enformer and AlphaGenome does not automatically produce a comprehensive biological simulator.

The reason is that predicting regulatory activity, predicting endogenous target-gene expression, and predicting cellular phenotypes are different levels of biological modeling.

6.1 From DNA Sequence to Disease Phenotype: A Multilevel Problem#

Conceptually, biological prediction can be divided into several levels.

Level 1: Regulatory capacity of DNA sequences

This is a domain in which MPRA is particularly useful.

MPRA measures the transcriptional regulatory activity of DNA sequences under artificial reporter assay conditions.

↓

Level 2: Activity in the native genomic environment

Regulatory activity depends on chromatin state, cellular state, and genomic location.

↓

Level 3: Target-gene identification and expression changes

Predicting endogenous gene expression requires understanding enhancer–promoter relationships, long-range genomic contacts, and functional redundancy among regulatory elements.

↓

Level 4: Responses of intracellular molecular networks

Changes in gene expression influence proteins, signaling pathways, metabolism, and feedback mechanisms.

↓

Level 5: Tissue, organismal, and disease phenotypes

Biological outcomes depend on interactions among multiple cell types, developmental processes, tissue environments, and drug responses.

The ability of MPRA to generate large amounts of information at Level 1 represents a major advance.

However, accurate prediction at Level 1 does not necessarily imply accurate prediction at Level 5.

Long-range genomic models such as AlphaGenome have expanded the types of information that can be predicted from DNA sequences, including features relevant to Levels 2 and 3 [11].

Nevertheless, these models are not complete mechanistic models of molecular systems or comprehensive biological simulators.

6.2 Connecting the Levels Is a Central Research Challenge#

MPRA provides large-scale experimental measurements of relationships between DNA sequences and reporter activity.

Long-range genomic models incorporate broader sequence context.

CRISPR-based functional experiments can test the consequences of perturbing sequences at their endogenous genomic locations.

Integrating these sources of information may help answer questions such as:

• Does a sequence with high MPRA activity also function strongly at its native genomic location?
• Under which chromatin conditions is its regulatory activity enhanced or suppressed?
• Which endogenous genes does it regulate?
• How does its activity change across cellular states?
• How do changes in endogenous gene expression affect cellular and organismal phenotypes?

Learning and experimentally validating the connections between these biological levels is a central challenge for developing more comprehensive biological simulations.

7. Future Research Priorities: What Should Be Validated Next?#

Further progress in MPRA research should involve more than simply increasing the number of tested DNA sequences.

An important goal is to systematically determine under which conditions MPRA-based predictions generalize to endogenous biological function.

The following priorities represent proposed research directions based on the limitations discussed above.

7.1 Priority Research Questions#

Priority	Research question	Proposed validation approach
1	How accurately does MPRA activity predict endogenous regulatory function?	Compare the same sequences using MPRA, CRISPRi, and base editing.
2	Under which genomic contexts do MPRA-based predictions fail?	Analyze prediction errors according to chromatin state, 3D contacts, and promoter distance.
3	Can predictive models generalize across cell types?	Evaluate models in unseen primary cells, differentiated cells, and patient-derived cells.
4	How important are enhancer–promoter combinations?	Measure the same regulatory sequences with multiple promoters under controlled conditions.
5	Do AI-designed regulatory sequences function in vivo?	Progress from independent reporter assays to validation of endogenous gene regulation and tissue-specific function.
6	Does MPRA-derived information provide additional value to long-range genomic models?	Incorporate MPRA information into models such as AlphaGenome and compare performance against baseline models using independent experimental data.

7.2 A Particularly Important Direction: Modeling the Difference Between MPRA Activity and Endogenous Function#

One especially important research direction is to treat discrepancies between MPRA predictions and endogenous regulatory function as an object of investigation.

Consider a DNA sequence with the following properties:

• It exhibits strong regulatory activity in MPRA.
• It exhibits weak regulatory activity at its native genomic location.

Traditionally, such a discrepancy might be treated simply as a limitation of MPRA.

However, systematically collecting these discrepancies and relating them to chromatin state, promoter interactions, transcription factor environments, and sequence length could reveal the conditions under which regulatory potential becomes functional activity.

Conceptually, this could be represented using a model such as:

[ \begin{aligned} A &= f_{\mathrm{MPRA}}(S,C,R)\ Y &= g(A,G,P,X) \end{aligned} ]

The variables are defined as follows.

Variable	Meaning
(S)	DNA sequence
(C)	Cellular state
(R)	Reporter assay conditions
(A)	Reporter activity under experimental conditions
(G)	Native genomic and chromatin environment
(P)	Relationship with the target promoter
(X)	Other intracellular and environmental factors
(Y)	Effects on endogenous gene regulation

This formulation is not an experimentally established biological law.

It is a conceptual framework for investigating relationships between MPRA measurements and endogenous regulatory function.

In particular, a single MPRA activity value may not contain sufficient information to fully represent endogenous regulatory behavior.

Nevertheless, treating MPRA activity as an intermediate experimental measurement rather than the final biological ground truth could enable more powerful models of context-dependent gene regulation.

Research published in 2026 has already explored the use of MPRA information to improve predictions of context-specific enhancer–gene regulatory interactions [9].

This suggests that the general research direction is beginning to develop into a concrete area of investigation.

7.3 From Larger Datasets to Understanding Context Dependence#

Increasing the number of DNA sequences that can be measured has been one of the major advances in MPRA research.

However, expanding a dataset from hundreds of thousands to millions of sequences does not automatically resolve the challenge of predicting regulatory function in native genomic environments.

Future research may need to emphasize not only dataset size but also diversity across biological and experimental conditions.

Examples include:

• Different promoter combinations
• Different chromatin environments
• Different cell types and cellular states
• Different sequence lengths
• Functional validation at native genomic locations
• Responses to drug treatment or disease-related conditions

Systematically varying these conditions may help separate sequence-intrinsic effects from interactions with the surrounding environment.

It is also important to recognize that the model with the highest predictive accuracy on MPRA measurements may not be the model that most accurately predicts endogenous gene regulation.

Machine learning evaluation should therefore consider not only performance on MPRA activity prediction but also performance against independent CRISPR-based functional experiments and endogenous gene expression measurements.

8. Conclusion: How Should MPRA Be Positioned in Genomic Research?#

MPRA is a technology that enables large-scale, quantitative measurements of the transcriptional regulatory capacity of DNA sequences under artificial experimental conditions.

It has made relationships between DNA sequence and regulatory activity accessible to large-scale computational learning.

Its contribution to genomic AI is therefore not primarily an innovation in neural network architecture.

Rather, it is the creation of functional experimental datasets that can support machine learning approaches to understanding gene regulation.

At the same time, MPRA inherits a fundamental limitation from conventional reporter assays.

Removing a DNA sequence from its native genomic environment results in the loss of important contextual information.

LentiMPRA partially addresses this issue by integrating reporter constructs into chromosomes.

However, it does not fully reproduce the regulatory function of a sequence at its native genomic location.

Based on the implicit assumptions and limitations discussed in this article, several conclusions can be drawn.

First, MPRA measures the regulatory capacity of DNA sequences, but it does not directly provide a complete measurement of endogenous gene regulatory function.

Second, MPRA measurements depend on experimental design, promoter selection, sequence length, and cellular state.

Reporter activity should therefore not be treated as a universal property entirely independent of experimental conditions.

Third, even when machine learning models achieve high accuracy in predicting MPRA activity, that performance does not necessarily translate into accurate predictions of endogenous target-gene expression or disease phenotypes.

Fourth, MPRA, CRISPR-based functional validation at native genomic loci, and long-range sequence models such as Enformer and AlphaGenome are not competing technologies.

They provide complementary information about different levels of gene regulation.

Fifth, future research should focus not only on increasing the scale of MPRA measurements but also on explicitly modeling and understanding the differences between reporter activity and endogenous biological function.

From this perspective, the central research objective should evolve beyond:

"Can we accurately predict MPRA activity from a DNA sequence?"

Toward a broader question:

"Under which cellular states, genomic environments, target-gene relationships, and biological conditions does the regulatory potential measured by MPRA become functional activity in living systems?"

For applications in biological simulation and drug discovery, improving MPRA measurement precision remains important.

However, it is equally important to explicitly identify what MPRA can measure and what lies beyond its observational scope.

Developing models and experiments that connect these different levels of biological regulation will be an important direction for the future 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] Klein, J. C. 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

[3] "Applications of high-throughput reporter assays to gene regulation studies," Current Opinion in Structural Biology, 2025. https://doi.org/10.1016/j.sbi.2025.103105

[4] Melnikov, A. 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

[5] 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

[6] Inoue, F. 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

[7] 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

[8] "Enhancer–promoter compatibility study," 2026. Full bibliographic details require verification.

[9] DeGroat, E., and Kreimer, A., "Massively parallel reporter assay–informed modeling improves prediction of context-specific enhancer–gene regulatory interactions," Nucleic Acids Research, 2026. https://doi.org/10.1093/nar/gkag554

[10] "Comprehensive evaluation of diverse massively parallel reporter assays to functionally characterize human enhancers genome-wide," Genome Biology, 2025. https://doi.org/10.1186/s13059-025-03828-8

[11] 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