Target Audience#

• Researchers and professionals who want to understand how EvoDiff, DPLM, RFdiffusion, ESM3, BoltzGen, and related models fit into the history of protein generative AI.
• Readers interested in the scientific assumptions underlying sequence-based, structure-based, and multimodal protein generation.
• Scientists, drug discovery professionals, and biotechnology researchers seeking to distinguish demonstrated capabilities from unresolved challenges in AI-driven protein design.

Summary#

• EvoDiff, introduced in 2023, is a discrete diffusion model that directly learns from evolutionary protein sequence data. At a time when structure-based protein generation was advancing rapidly, EvoDiff demonstrated the potential of sequence-first protein design.
• Its fundamental assumption is that the statistical regularities accumulated through natural evolution encode useful constraints on protein structure and function. However, generating natural-looking sequences is not equivalent to designing proteins with desired biological functions.
• The 2026 eLife Reviewed Preprint reported experimental validation involving protein expression, secondary structure, mitochondrial targeting signals, metal binding, and protein–protein interactions. Nevertheless, questions remain regarding comparative performance, generalization to novel functions, and practical design efficiency.
• Subsequent research, including DPLM, DPLM-2, ESM3, La-Proteina, Proteina-Complexa, and BoltzGen, has expanded protein generation toward integrating sequence, structure, function, and molecular interactions.
• EvoDiff's historical significance is not that it proved "sequence is all you need." Rather, it established the value of evolutionary sequence space as an independent generative domain and highlighted the fundamental distinction between generating proteins and designing proteins to fulfill specific objectives.

1. Introduction: What Did EvoDiff Change in Protein Generative AI?#

Protein generative AI has developed rapidly in recent years.

Alongside protein language models (PLMs) that generate amino acid sequences from large sequence datasets, researchers have developed diffusion models that directly generate three-dimensional protein structures, multimodal models that jointly represent sequence and structure, and all-atom generative models that incorporate molecular interactions with target molecules.

Within this research landscape, EvoDiff, introduced in 2023, occupies an important position.

EvoDiff was proposed by Alamdari and colleagues in their study, "Protein generation with evolutionary diffusion: sequence is all you need." The original bioRxiv preprint was released on September 12, 2023. [5]

At the time, models such as RFdiffusion and Chroma were driving major advances in three-dimensional protein structure generation.

Against this background, EvoDiff raised a different question:

Why should protein generation necessarily begin with generating a three-dimensional structure?

Nature contains an enormous collection of protein sequences accumulated through evolutionary history.

If generative models could learn directly from these sequences, perhaps they could produce useful proteins without explicitly generating their three-dimensional structures.

This was the central idea behind EvoDiff.

However, the approach raises an even more fundamental question:

To what extent does learning the statistical regularities of naturally occurring protein sequences enable us to design proteins with artificially specified functions?

This question extends beyond EvoDiff to protein generative AI as a whole.

In this article, we first review the historical development of protein generative models before EvoDiff and explain its technical contributions.

We then examine the assumptions underlying EvoDiff, the capabilities validated experimentally through 2026, and the limitations that remain unresolved.

Finally, we revisit EvoDiff's historical significance in light of subsequent research, including DPLM, ESM3, La-Proteina, and BoltzGen.

2. Before EvoDiff: Learning Proteins as a Language#

To understand EvoDiff, we must first examine how researchers came to treat amino acid sequences as objects that could be learned and generated.

Protein sequences encode numerous constraints accumulated over evolutionary time.

Some amino acid substitutions are tolerated without compromising function, while others disrupt folding or biological activity.

Different positions within a sequence may also evolve in a coordinated manner.

These observations motivated an important hypothesis:

By learning the statistical regularities of naturally occurring protein sequences, could a model also acquire information about protein structure and function?

2.1. DeepSequence: Learning Evolutionary Sequence Distributions#

One influential development around 2017–2019 was DeepSequence.

DeepSequence used a variational autoencoder (VAE) to model the distribution of protein sequences within a multiple sequence alignment (MSA).

Its primary objective was predicting the functional effects of mutations.

However, its historical significance extends beyond mutation effect prediction.

DeepSequence demonstrated that a generative model trained on evolutionarily observed sequences could capture statistical constraints associated with protein function. [1]

This idea became foundational to subsequent sequence-based generative approaches.

Rather than explicitly encoding every biophysical rule, a model could potentially infer useful biological constraints from the distribution of sequences that had survived natural evolution.

2.2. UniRep: Learning General-Purpose Protein Representations#

Subsequent studies, including UniRep, explored self-supervised learning from large collections of protein sequences.

UniRep demonstrated that representations learned directly from amino acid sequences could capture information useful for predicting protein properties and supporting protein engineering. [2]

Whereas DeepSequence primarily focused on individual protein families, UniRep pursued representations transferable across diverse proteins.

This marked an important step toward modern protein foundation models.

2.3. ProGen, ProGen2, and ProtGPT2: Direct Sequence Generation#

Between approximately 2020 and 2023, transformer-based protein language models developed rapidly.

Representative examples include ProGen, ProGen2, and ProtGPT2.

These models treat amino acid sequences analogously to natural language and generate new sequences by predicting subsequent amino acids.

A conventional autoregressive model factorizes the probability of a protein sequence as follows:

$$ P(x_1,\ldots,x_L)

\prod_{i=1}^{L} P(x_i\mid x_{<i}) $$

Here, $x_i$ denotes the amino acid at position $i$.

An important milestone was the ProGen study, which experimentally demonstrated the generation of functional artificial proteins using a 1.2-billion-parameter model trained on approximately 280 million protein sequences. [3]

This result established that useful protein generation does not necessarily require explicit three-dimensional structure generation.

Sequence information alone can support powerful generative models.

Therefore, EvoDiff did not originate the broader idea that sequence information could be sufficient for useful protein generation.

Its novelty lay in combining three elements:

• Sequence language modeling: Treating amino acid sequences as the generative domain.
• Diffusion modeling: Learning to reconstruct sequences progressively from corrupted states.
• Evolutionary-scale sequence data: Leveraging the enormous volume of naturally occurring protein sequences.

Together, these elements provided an alternative to conventional left-to-right autoregressive protein generation.

3. A Parallel Development: Structure-First Protein Generation#

While sequence-based generative models were advancing, a parallel line of research focused on generating and designing proteins through three-dimensional structures.

3.1. ProteinMPNN: Designing Sequences from Structures#

One influential development was ProteinMPNN.

ProteinMPNN is an inverse-folding model that designs amino acid sequences conditioned on a specified protein backbone. [15]

Conceptually, the process is:

3D backbone → Amino acid sequence

Rather than generating a sequence without structural constraints, the model seeks sequences compatible with a given backbone geometry.

This became a fundamental component of structure-first protein design workflows.

3.2. RFdiffusion and Chroma: Generating Three-Dimensional Structures#

In 2023, RFdiffusion and Chroma demonstrated substantial advances in three-dimensional protein generation.

RFdiffusion uses representations derived from RoseTTAFold to generate protein backbones through diffusion in three-dimensional space. [16]

The approach enabled progress in applications including:

• De novo backbone generation
• Motif scaffolding
• Binder design

Chroma emerged during the same period as a programmable generative model for conditional three-dimensional protein design.

Its research was published in Nature in November 2023. [4]

A representative structure-first pipeline can be summarized as:

Random noise → 3D backbone → ProteinMPNN → Amino acid sequence

A diffusion model first generates a backbone, after which a sequence-design model produces amino acids compatible with that structure.

The central object of generation is therefore the three-dimensional structure.

However, sequence-first and structure-first methods are not inherently competing paradigms.

They approach the protein design problem using different sources of information and different representations.

EvoDiff was important because it demonstrated a compelling alternative at a time when structure-based diffusion models were attracting substantial attention.

4. EvoDiff: Making Evolutionary Sequence Space the Generative Domain#

EvoDiff, introduced by Alamdari and colleagues in 2023, is a discrete diffusion approach for directly generating protein sequences. [5]

Its central idea is to learn the probability distribution of amino acid sequences rather than explicitly generating protein structures.

4.1. Why Learn from Evolutionary Sequences?#

A central motivation for EvoDiff was the enormous difference between the availability of protein sequence data and experimentally determined structural data.

Around 2023, the number of experimentally determined protein structures was on the order of hundreds of thousands.

By contrast, vastly larger collections of protein sequences were available.

This asymmetry was an important motivation for EvoDiff. [6]

Structure-based generative models conceptually learn distributions such as:

$$ P(\text{structure}) $$

EvoDiff instead directly models:

$$ P(\text{sequence}) $$

By using evolutionary sequence data as the primary learning resource, the approach can exploit substantially larger training datasets.

EvoDiff-Seq was trained using approximately 42 million protein sequences derived from UniRef50.

The researchers also introduced EvoDiff-MSA, which models multiple sequence alignments directly. [5][14]

This shift—from experimentally determined structures to evolutionary sequences as the primary generative domain—is one of EvoDiff's defining characteristics.

4.2. Generation Without a Fixed Left-to-Right Order#

Autoregressive protein language models such as ProGen typically generate sequences from left to right:

$$ 1\rightarrow2\rightarrow3\rightarrow\cdots\rightarrow L $$

EvoDiff instead uses diffusion-based approaches that learn to reconstruct protein sequences from masked or corrupted states.

A particularly important component is the Order-Agnostic Autoregressive Diffusion Model (OADM).

OADM is not restricted to generating residues in a fixed left-to-right order.

Conceptually, it enables operations such as:

Partially specified sequence:
???A????G???

Generated completion:
MKTAV...G...

This simplified illustration represents the ability to preserve specified residues while generating missing regions.

The generated positions may appear in the middle, near the termini, or across multiple separated regions.

4.3. Flexible Conditional Generation#

This capability supports several protein design tasks:

• Unconditional generation: Generating new protein sequences without specifying a target sequence.
• Sequence inpainting: Preserving parts of an existing sequence while filling in missing regions.
• Motif scaffolding: Fixing functionally important motifs and generating surrounding residues.
• N- and C-terminal conditioning: Preserving terminal regions while generating internal sequences.
• MSA-conditioned generation: Using evolutionary information from related sequences to guide generation.

EvoDiff's contribution was not simply demonstrating that protein sequences could be generated.

Its distinctive contribution was combining large-scale evolutionary sequence learning with flexible, position-independent conditional generation.

This capability became an important theme in subsequent diffusion protein language model research.

However, order-agnostic generation should not be confused with unrestricted iterative optimization.

In OADM, residues are generally fixed once generated.

Consequently, repeatedly revising previously generated residues in response to a functional predictor requires additional mechanisms.

The EvoDiff authors also discussed how approaches such as D3PM, which permit iterative residue editing, may be more compatible with certain types of function-guided generation. [14]

4.4. The Role of Multiple Sequence Alignments#

EvoDiff-MSA leverages evolutionary information contained in multiple sequence alignments.

MSAs can reveal conservation patterns and correlated substitutions across residue positions.

Such information may reflect structural and functional constraints.

However, not every statistical pattern in an MSA directly represents a functional relationship.

Phylogenetic relatedness, sequence sampling biases, and alignment quality also influence the observed patterns.

Moreover, sufficiently informative MSAs may not be available for novel proteins or rare protein families.

The original EvoDiff study also reported cases in which motif-scaffolding performance declined when the number of sequences in an MSA fell outside the range encountered during training. [5][6]

Thus, evolutionary information is a valuable resource, but its reliability depends on the quality and availability of the underlying sequences.

5. An Important Advantage: Generating Proteins Without a Single Stable Structure#

One of EvoDiff's distinctive contributions is its ability to model proteins and regions that do not adopt a single stable three-dimensional structure.

Structure-generating models such as RFdiffusion are particularly effective when the design objective involves a well-defined geometry.

However, not all biologically important protein regions are characterized by a single stable structure.

Examples include:

• Intrinsically disordered proteins and regions (IDPs/IDRs)
• Flexible linkers
• Targeting peptides
• Signal sequences

In these regions, structural flexibility or disorder may contribute directly to biological function.

For example, some targeting signals depend strongly on amino acid composition, charge, hydrophobicity, and local sequence properties rather than a single stable tertiary structure.

Because EvoDiff generates sequences directly, such regions can, in principle, be treated as ordinary generative targets.

This opens possibilities beyond conventional design frameworks centered on stable structure–function relationships. [5]

However, generating an intrinsically disordered region and controlling its biological function are different tasks.

Similarly, demonstrating that a generated targeting sequence directs a protein to a cellular compartment does not necessarily establish that its conformational ensemble or dynamic properties have been accurately reproduced.

This distinction becomes particularly important when interpreting EvoDiff's experimental results.

6. Six Underlying Assumptions of EvoDiff#

EvoDiff's technical contributions are important, but understanding its historical significance also requires examining the assumptions on which the approach depends.

The following six assumptions are derived from the model's design, evaluation strategy, and broader scientific context.

They should not all be interpreted as assumptions explicitly enumerated by the original authors.

6.1. Assumption 1: Natural Evolutionary Sequence Distributions Are Useful for Artificial Protein Design#

The most fundamental assumption behind EvoDiff is that naturally occurring protein sequences encode information useful for designing new proteins.

Conceptually, EvoDiff attempts to learn:

$$ P_{\theta}(x)\approx P_{\mathrm{natural}}(x) $$

Here, $x$ denotes a protein sequence, and $P_{\mathrm{natural}}(x)$ represents the distribution of naturally occurring sequences.

In practice, the model learns an approximation based on a collected dataset rather than the complete distribution of proteins in nature.

However, the distribution relevant to an engineering objective may instead be expressed as:

$$ P(x\mid f_{\mathrm{desired}},c) $$

Here, $f_{\mathrm{desired}}$ represents the desired biological function, and $c$ represents the environmental or design constraints.

These distributions are not necessarily identical.

Natural evolution did not directly optimize proteins for the requirements of modern drug development, such as high target specificity, low immunogenicity, manufacturability, or long-term storage stability.

Instead, evolutionary selection operates under constraints associated with survival and reproduction in particular biological environments.

EvoDiff therefore relies on a broad underlying assumption:

The protein sequence space explored by natural evolution contains sufficient information to support at least some artificially defined protein design objectives.

There is substantial evidence supporting the usefulness of this assumption.

However, it has not been established that the assumption holds for arbitrary drug discovery objectives.

6.2. Assumption 2: Statistical Sequence Regularities Encode Structural and Functional Constraints#

Protein function depends strongly on amino acid sequence.

Consequently, learning patterns of conservation, co-occurrence, and evolutionary substitution may reveal constraints associated with structure and function.

Studies such as DeepSequence, UniRep, ProGen, and EvoDiff support this general idea. [1][2][3][5]

However, two capabilities must be distinguished:

• Learning statistical patterns associated with biological function.
• Guaranteeing that a specified biological function will be realized.

For example, a sequence may contain residues associated with metal binding without positioning them in the correct three-dimensional geometry.

Likewise, a catalytic motif may be conserved without providing the structural environment needed for substrate access or catalytic activity.

Protein function also depends on temperature, pH, solvent conditions, cofactors, post-translational modifications, cellular localization, and interacting molecules.

Thus, learning functional constraints indirectly from sequence statistics does not guarantee that those functions can be controlled under arbitrary conditions.

6.3. Assumption 3: Learning Natural Sequence Distributions Enables Exploration of Novel Functional Sequences#

EvoDiff can generate sequences not present in its training dataset.

However, three forms of novelty should be distinguished:

• Sequence novelty
• Structural novelty
• Functional novelty

A sequence can be substantially different from known proteins while retaining a familiar structure or function.

Conversely, a novel sequence may fail to fold or function properly.

The set of naturally observed protein sequences occupies only a tiny fraction of the theoretically possible amino acid sequence space.

Therefore, when generation moves far beyond the training distribution, learned statistical regularities may no longer reliably preserve functional validity.

This is an example of the broader problem of out-of-distribution generalization.

6.4. Assumption 4: Flexible Sequence Completion Supports Flexible Functional Design#

OADM allows different sequence regions to be generated or reconstructed.

This enables applications such as motif scaffolding, in which important functional residues are preserved while the surrounding sequence is generated.

However, flexibility in choosing sequence positions is not equivalent to flexibility in specifying biological function.

Preserving catalytic residues does not guarantee their correct spatial arrangement.

Similarly, preserving a binding motif does not guarantee favorable interactions across the complete protein–target interface.

The EvoDiff study evaluated 17 motif-scaffolding tasks computationally, using structure-based metrics such as pLDDT and motif RMSD. [5][6]

Thus, although EvoDiff can generate sequences without explicit structural input, its evaluation is not necessarily independent of structural information.

6.5. Assumption 5: Evolutionary Covariation in MSAs Represents Useful Design Constraints#

EvoDiff-MSA uses information from evolutionarily related sequences.

Conservation and correlated substitutions may reflect constraints associated with folding and function.

However, MSAs also contain patterns resulting from phylogeny, sequence sampling, and alignment procedures.

Furthermore, models conditioned on MSAs may struggle when sufficiently informative homologous sequences are unavailable.

Success within well-represented protein families does not necessarily imply equivalent performance on entirely novel families or artificial proteins.

The availability and reliability of evolutionary information therefore impose important limitations.

6.6. Assumption 6: Computational Metrics Predict Experimental Success#

EvoDiff uses several computational metrics to assess generated sequences.

These metrics are valuable, but each measures only certain aspects of protein quality.

Metric	What It Primarily Measures	What It Does Not Directly Measure
Perplexity	Agreement with statistical patterns in natural protein sequences	Actual biological function
pLDDT	Local confidence in predicted protein structure	Experimental stability or activity
Folding self-consistency	Agreement between a generated sequence and its predicted structure	Biological function in cells
Fréchet ProtT5 Distance	Similarity of sequence distributions in a protein language model embedding space	Experimental equivalence of functional distributions
Sequence novelty	Difference from known protein sequences	Novel biological function
Motif RMSD	Structural agreement of a motif with a reference arrangement	Catalytic activity or binding specificity

These metrics can help select promising candidates.

However, none directly measures all aspects of biological function.

For example, a favorable Fréchet ProtT5 Distance may indicate similarity to natural protein representations, but it does not demonstrate that a generated library experimentally covers the diversity of naturally occurring functions.

This distinction between computational plausibility and experimental validation is central to evaluating EvoDiff and protein generative AI more broadly.

7. Was "Sequence Is All You Need" Correct?#

EvoDiff's title makes a deliberately strong statement: "sequence is all you need."

Looking back from 2026, it is useful to distinguish at least two interpretations.

7.1. Sequence Data Alone Can Support Powerful Generative Models#

This interpretation is supported by EvoDiff and subsequent protein language model research.

Evolutionary protein sequences contain extensive information about biological constraints.

Large-scale sequence learning can capture properties associated with protein structure and function.

EvoDiff demonstrated one way of exploiting these properties through discrete diffusion.

It also suggested that sequence-first generation may be useful for proteins and regions that lack a single stable structure.

In this sense, the claim that sequences alone can support useful generative models has substantial empirical support.

7.2. Protein Design Never Requires Information Beyond Sequence#

This stronger interpretation has not been established.

Protein design requires more than generating sequences that resemble natural proteins.

An engineered protein may need to fold reliably, bind a particular target, avoid unintended interactions, remain stable under specific conditions, and meet practical manufacturing requirements.

Structural information, functional measurements, target properties, and environmental constraints may all be important.

Subsequent research has increasingly explored ways to integrate these sources of information.

Conceptually, the expansion can be expressed as:

$$ \text{Sequence} \rightarrow \text{Sequence + Structure} \rightarrow \text{Sequence + Structure + Function + Interactions} $$

However, this should not be interpreted as a linear replacement of older methods by newer ones.

Sequence-only and structure-conditioned approaches continue to develop in parallel and may be appropriate for different problems.

EvoDiff's title is therefore best understood as a challenge to the assumption that protein generation must be centered on three-dimensional structures.

It did not establish that structural information is universally unnecessary for protein design.

8. EvoDiff in 2026: What Has Been Experimentally Validated?#

An important and sometimes overlooked aspect of EvoDiff is that the research did not end with the original 2023 preprint.

On August 24, 2026, EvoDiff was released as an eLife Reviewed Preprint. [14]

This version included experimental investigations alongside computational evaluations.

The investigated properties and applications included:

• Expression of generated proteins
• Protein folding and secondary structure
• Intrinsically disordered mitochondrial targeting signals
• Metal-binding proteins
• Protein binders

These experiments demonstrate that EvoDiff is not merely a computational proof of concept.

However, the strength of the evidence differs substantially across applications.

8.1. Unconditional Generation: Can Novel Sequences Produce Real Proteins?#

In the unconditional generation experiment, EvoDiff-Seq generated 1,000 candidate sequences.

From these, 25 candidates were selected using computational criteria, including structure-prediction-related metrics.

Four of the 25 experimentally evaluated candidates showed expression and secondary-structure characteristics consistent with the intended properties. [14]

The proportion among the experimentally tested candidates was:

$$ \frac{4}{25}=16% $$

However, this 16% should not be interpreted as the unconditional success probability across all generated sequences.

The 25 candidates had already passed a computational selection process.

Furthermore, the experimental evidence concerned protein expression and secondary-structure characteristics, including circular dichroism measurements.

It did not demonstrate a specified biological activity for these proteins.

The result supports EvoDiff's ability to generate novel sequences that can yield experimentally viable protein candidates.

At the same time, it illustrates that computationally promising sequences do not necessarily succeed experimentally.

8.2. Mitochondrial Targeting Signals: Can EvoDiff Design IDR-Associated Functions?#

One of EvoDiff's distinctive experiments involved mitochondrial targeting signals.

The study used sequence inpainting to generate targeting sequences associated with Cox15.

For eight selected candidates, the experiments reported results consistent with mitochondrial localization of Cox15. [14]

This is an important demonstration that sequence-based generation can reproduce a biological function associated with a region that does not depend on a single stable tertiary structure.

However, the candidate selection process matters.

The eLife review noted that Cox15 is an evolutionarily conserved mitochondrial protein and that EvoDiff-MSA was supplied with homologous sequence information.

The eight experimentally evaluated sequences were also selected from hundreds of generated candidates using prediction tools including DR-BERT and MitoFates. [17]

Consequently, the experiment demonstrates the effectiveness of a pipeline combining generation and computational prediction, rather than isolating EvoDiff's contribution.

Moreover, mitochondrial localization was experimentally observed, but the structural disorder of the generated regions was not directly demonstrated.

Further experiments would be required to establish general-purpose design capabilities across different IDR functions.

8.3. Calcium-Binding Motifs: Can EvoDiff Generate Functional Scaffolds?#

The study also explored motif scaffolding for metal-binding proteins.

In an experiment involving calcium-binding motifs, 12 of 13 candidates were expressed, and one candidate exhibited a spectral change suggestive of Ca²⁺ binding. [14]

This supports the possibility that preserving a functionally important motif while generating the surrounding sequence can yield candidates with relevant biological properties.

However, it does not establish general-purpose metal-binding protein design.

Important questions remain regarding binding affinity, specificity, stoichiometry, and structural responses to binding.

Furthermore, success with one motif does not necessarily generalize to other metals or protein families.

8.4. p53–MDM2 Binding: Can EvoDiff Generate Protein Binders?#

EvoDiff also evaluated motif scaffolding for the p53–MDM2 interaction.

The experiment generated 1,300 candidate sequences.

Computational evaluation and selection, including the use of OmegaFold, reduced this set to 24 experimentally tested candidates.

Among these, four candidates exhibited dissociation constants below 50 nM, and eight candidates in total showed evidence of binding. [14]

These findings support the ability to generate sequences that preserve an interaction motif and exhibit measurable binding to a target.

However, several distinctions are necessary.

First, the 24 tested candidates had been selected from 1,300 generated sequences.

Therefore, the ratio of four successes among 24 tested candidates does not represent the success probability across all generated sequences.

Similarly, four divided by 1,300 should not be treated as the intrinsic success rate of EvoDiff alone.

Second, binding to MDM2 is not equivalent to demonstrating therapeutic activity.

Additional experiments would be needed to establish whether the designed proteins:

• Disrupt the relevant interaction inside cells.
• Restore p53 transcriptional activity.
• Produce the desired biological effects in cancer cells.
• Avoid unwanted interactions with other molecules.
• Meet requirements for stability, delivery, and safety.

The experiment is therefore an encouraging demonstration of binder generation, but it does not establish therapeutic efficacy.

8.5. What the Experimental Results Do—and Do Not—Demonstrate#

The 2026 experimental evidence can be summarized as follows.

Application	Reported Result	Supported Conclusion	What Remains Unproven
Unconditional generation	1,000 sequences generated; 25 selected; four showed expression and expected secondary-structure characteristics	Novel sequences can produce experimentally viable protein candidates	General functional success rates and success probability across all generated candidates
Mitochondrial targeting signals	Eight selected candidates showed results consistent with Cox15 localization	Sequence inpainting can reproduce a particular IDR-associated function	Generalization across IDR functions and direct confirmation of disorder
Calcium-binding motif	12 of 13 candidates expressed; one exhibited a change suggestive of Ca²⁺ binding	Motif-conditioned generation can yield potentially functional scaffolds	General metal-binding capability and precise control of affinity or specificity
p53–MDM2 binding	Four of 24 tested candidates had $K_D<50$ nM; eight showed binding evidence	Motif-guided sequence generation can yield binders	Cellular activity, target specificity, and therapeutic efficacy

Together, these experiments support the usefulness of sequence-first protein generation across several biological tasks.

However, they involve relatively small sets of selected candidates.

Demonstrating that EvoDiff can generate functional proteins is not the same as demonstrating that it can reliably design proteins for arbitrary objectives.

9. Current Challenges and Technical Limitations of EvoDiff#

To evaluate EvoDiff accurately, we must distinguish between architectural constraints, experimentally observed limitations, and unresolved questions about generalization.

9.1. Generating Natural-Looking Sequences Does Not Guarantee Desired Function#

The most fundamental limitation is that learning natural sequence distributions is not equivalent to optimizing a specified function.

This distinction can be understood through three levels of design.

Level 1: Generating natural-looking protein sequences

The model reproduces statistical patterns found in naturally occurring protein sequences.

Level 2: Generating proteins that fold and function experimentally

The resulting sequences produce proteins with appropriate structural or biological properties.

Level 3: Designing proteins that meet an artificial engineering objective

The proteins satisfy requirements such as target specificity, binding affinity, stability, expression, safety, and manufacturability.

These levels are not automatically connected.

EvoDiff's 2026 experiments provide evidence that selected candidates can reach aspects of Level 2.

However, broad and reliable achievement of Level 3 remains unproven.

9.2. EvoDiff Does Not Consistently Outperform Other Generative Models#

EvoDiff offers flexible conditional generation.

However, it does not outperform conventional autoregressive models across every metric.

The study reports cases in which left-to-right autoregressive models perform better on measures related to natural sequence distributions or structural plausibility. [5][14]

Furthermore, Reviewer #1 in the 2026 eLife review noted that the approach did not produce a large improvement in success rates relative to other sequence generation methods. [17]

This does not diminish EvoDiff's methodological contribution.

Rather, its contribution should be characterized precisely.

EvoDiff's main advance was combining evolutionary-scale sequence learning with flexible, order-agnostic conditioning—not demonstrating universal superiority over existing protein language models.

9.3. Direct Control of Geometric Constraints Is Limited#

EvoDiff can preserve functionally important amino acid motifs.

However, explicitly specifying three-dimensional distances, side-chain orientations, or spatial complementarity with a target molecule is not straightforward within standard sequence-first generation.

This contrasts with structure-conditioned approaches such as RFdiffusion.

For example, effective binder design depends not only on natural-looking amino acid sequences but also on shape complementarity, electrostatics, hydrophobic interactions, and interface geometry.

Consequently, EvoDiff-generated candidates often require additional structural prediction, complex modeling, and experimental screening.

9.4. Generalization to Novel Protein Families and Functions Is Insufficiently Established#

Although EvoDiff learns from a large collection of natural sequences, performance outside its training distribution is not guaranteed.

Examples of challenging design objectives include:

• Artificial proteins that differ substantially from known families.
• Enzymes acting on unusual substrates.
• Proteins that function under extreme temperatures or pH conditions.
• Binders that precisely discriminate among closely related targets.
• Molecular switches responding to artificial inputs.
• Proteins combining functions rarely or never observed together in nature.

These problems require testing whether sequence models can generalize beyond the functional patterns represented in their training data.

Generating new members of a known protein family is not equivalent to inventing a genuinely new function.

9.5. OADM Has Limitations for Iterative Residue Editing#

OADM supports generation in different residue orders.

However, residues are generally retained once generated.

This makes its standard formulation different from optimization procedures that repeatedly revise the entire sequence in response to functional predictions.

Alternative diffusion formulations, property-guided generation, and optimization methods may be better suited to some iterative design workflows.

9.6. Sequence Length, Insertions, Deletions, and Domain Organization Remain Challenging#

Natural evolution involves more than amino acid substitutions.

Insertions, deletions, domain recombination, and changes in sequence length also contribute to protein diversity.

In standard EvoDiff-Seq generation and inpainting, sequence length and masked regions are typically specified in advance.

Although EvoDiff-MSA handles alignment gaps, this does not automatically provide unrestricted control over insertions and deletions.

Subsequent work such as EvoFlows has also discussed limitations associated with controlling insertion generation in EvoDiff-MSA. [18]

These issues matter when designing proteins with flexible lengths or novel domain architectures.

9.7. Training Data Bias May Affect Generated Proteins#

Large-scale sequence datasets are a major advantage of EvoDiff.

However, UniRef50 does not represent every naturally occurring protein family or function equally.

Biases may arise from the organisms that have been extensively studied, the environments that have been sampled, and the protein families that are particularly abundant in sequence databases.

MSAs may also contain phylogenetic biases and alignment errors.

Therefore, diversity in generated sequences should not automatically be interpreted as comprehensive coverage of biological function space.

10. How Did the eLife Reviewers Evaluate EvoDiff?#

The 2026 eLife Reviewed Preprint includes an editorial assessment and public peer reviews. [14][17]

The assessment recognized EvoDiff as an important order-agnostic diffusion foundation model for protein sequence generation.

Its applications to unconditional generation, motif scaffolding, MSA-conditioned generation, and IDR inpainting were viewed as valuable.

The combination of computational evaluation and experimental validation also provided evidence supporting several principal findings.

However, reviewers raised questions about comparative advantages and the generality of certain claims.

10.1. Reviewer #1: Limited Improvement in Success Rates#

Reviewer #1 acknowledged the value of EvoDiff's design methodology and potential applications.

However, the review noted that success rates were not dramatically improved compared with other sequence generation methods, and substantial experimental screening was still needed to identify successful candidates. [17]

This distinction is particularly important for practical applications.

A model capable of producing at least some functional proteins is not necessarily a model that discovers functional proteins efficiently under a limited experimental budget.

10.2. Reviewer #2: General IDR Design Has Not Been Demonstrated#

Reviewer #2 recognized the breadth of EvoDiff's generative capabilities but questioned how broadly the IDR-related claims could be generalized.

In particular, Cox15 is an evolutionarily conserved mitochondrial protein, and informative homologous sequences were available.

The eight tested candidates were also selected using computational predictors.

Furthermore, the experiment demonstrated mitochondrial localization but did not directly establish the structural disorder of the generated regions. [17]

Thus, the results should not be interpreted as proof that EvoDiff can generally design arbitrary IDR functions.

10.3. The Fundamental Lesson from Peer Review#

The reviews highlight two conclusions that should be considered simultaneously.

First, EvoDiff introduced an important method and demonstrated experimentally useful outcomes across several design tasks.

Second, universal superiority over alternative methods and reliable generalization to unfamiliar design objectives remain unproven.

This distinction is not unique to EvoDiff.

It reflects a broader challenge in protein generative AI: connecting model expressiveness, sequence novelty, computational quality metrics, and actual experimental success rates.

11. From EvoDiff to DPLM: The Development of Sequence Diffusion#

Following EvoDiff, diffusion-based protein sequence models continued to evolve.

One important development was DPLM, the Diffusion Protein Language Model.

11.1. DPLM: From Sequence Generation to Foundation Models#

DPLM was proposed by Wang and colleagues and presented at ICML 2024. [7]

It extended sequence diffusion beyond generation toward general-purpose protein representation learning.

Its applications include:

• Unconditional generation
• Motif scaffolding
• Structure-conditioned inverse folding
• Property-guided generation
• Downstream prediction

DPLM's importance lies in treating diffusion not merely as a sequence generator but as a foundation-model framework supporting both generation and representation learning.

Conceptually, the research progression can be described as:

EvoDiff → DPLM

This does not imply that DPLM directly inherited EvoDiff's implementation.

Rather, EvoDiff demonstrated the value of large-scale protein sequence diffusion, while DPLM further explored its potential as a general-purpose protein language modeling framework.

11.2. DPLM-2: Joint Modeling of Sequence and Structure#

DPLM-2 was presented at ICLR 2025. [8]

It extends the diffusion protein language model framework to jointly represent amino acid sequences and structural information.

Protein structures are converted into discrete structural tokens, allowing sequence and structure to be modeled together.

Conceptually, the model learns a joint distribution:

$$ P(\text{sequence},\text{structure}) $$

This supports tasks such as:

• Sequence → Structure
• Structure → Sequence
• Sequence + Structure → Completion
• Joint sequence–structure generation
• Motif scaffolding

The development from EvoDiff to DPLM and DPLM-2 illustrates an important pattern.

EvoDiff demonstrated the usefulness of sequence diffusion.

DPLM expanded sequence diffusion toward foundation-model learning.

DPLM-2 integrated structural information into the same general framework.

Rather than rejecting sequence-first design, this line of research builds upon its strengths while incorporating additional information.

12. ESM3: Integrating Sequence, Structure, and Function#

Another important development is ESM3.

The ESM3 research was published in Science in 2025. [9]

ESM3 jointly models protein sequence, structure, and function.

$$ \text{Sequence}+\text{Structure}+\text{Function} $$

This represents a broader approach than conventional sequence-only or structure-only generation.

12.1. Generating the Fluorescent Protein esmGFP#

One of the best-known results of ESM3 was the generation of the fluorescent protein esmGFP.

The generated protein exhibited approximately 58% sequence identity to known fluorescent proteins while retaining experimentally verified fluorescence.

The researchers described the evolutionary distance involved as approximately 500 million years. [9]

This figure should not be interpreted as experimentally reproducing 500 million years of evolution.

Rather, it represents an estimated evolutionary timescale associated with the observed sequence divergence.

The importance of the result lies in demonstrating the potential of generative models to explore protein sequences that differ substantially from known examples while retaining biological function.

12.2. What EvoDiff and ESM3 Have in Common#

EvoDiff and ESM3 share an underlying philosophy.

Both seek to leverage information accumulated through biological evolution to generate new proteins.

However, they differ in the types of information modeled.

EvoDiff focuses on amino acid sequences.

ESM3 jointly models sequence, structure, and function.

The emergence of ESM3 does not demonstrate that EvoDiff's sequence-first approach was incorrect.

Instead, it reflects a broader effort to combine the strengths of sequence learning with additional structural and functional constraints.

13. 2025–2026: Flow Matching and All-Atom Protein Generation#

Between 2025 and 2026, protein generative models increasingly moved toward more detailed molecular representations.

A simplified view of this development is:

• Around 2023: Backbone generation
• Around 2024–2025: Backbone and sequence generation
• Around 2025–2026: All-atom generation incorporating sequence and target interactions

These categories overlap, and the corresponding research directions have evolved in parallel.

Nevertheless, they reflect an expanding effort to model not only backbone geometry but also amino acid identities, side-chain atoms, and molecular interactions.

13.1. Proteina: Scaling Structural Generation with Flow Matching#

One representative model is NVIDIA's Proteina.

Proteina uses flow matching, a generative modeling framework related to diffusion, for protein backbone generation.

The researchers expanded training using tens of millions of synthetic structures and reported the generation of designable backbones containing up to approximately 800 residues.

Proteina was selected as an Oral presentation at ICLR 2025. [10]

The research illustrates how large-scale training data and newer generative formulations can improve protein backbone generation.

13.2. La-Proteina: Joint Generation of Sequence, Backbone, and Side Chains#

La-Proteina was presented at ICLR 2026. [11]

It addresses all-atom protein generation by jointly modeling:

• Amino acid sequence
• Protein backbone
• Side-chain atoms

La-Proteina represents Cα backbone coordinates explicitly while using latent representations for amino acid identities and side-chain geometry.

It employs partially latent flow matching to jointly generate sequence and atomic structure.

The study reports fully atomistic joint sequence–structure generation for proteins of up to 800 residues, as well as active-site motif scaffolding. [11]

The contrast with EvoDiff is clear.

EvoDiff primarily operates in:

$$ \boxed{\text{Sequence only}} $$

La-Proteina instead explores:

$$ \boxed{\text{Sequence + Every atom}} $$

This illustrates the expansion of protein generative modeling toward increasingly detailed molecular representations.

However, it does not establish that all-atom models are universally superior.

The appropriate representation depends on the design objective, available training data, computational resources, and constraints.

14. An Important Trend in 2026: Binder Design and Post-Generation Optimization#

One of the major application areas in protein generative AI is binder design.

Binder design aims to create proteins that bind specified molecular targets.

Conceptually:

$$ \text{Target}\rightarrow\text{Protein binder} $$

This has direct relevance to drug discovery and therapeutic protein development.

However, designing a useful binder requires more than producing a natural-looking sequence.

Binding-interface geometry, affinity, specificity, stability, and other properties are all important.

14.1. Proteina-Complexa: Integrating Generation and Search#

Proteina-Complexa is a 2026 study of protein binder design using fully atomistic generative modeling.

It integrates candidate generation with test-time search and optimization informed by computational evaluation, including structure-prediction-related metrics. [12]

A conventional generative model may primarily sample candidates from a learned distribution.

Proteina-Complexa emphasizes a broader process:

Generate → Evaluate → Search → Improve

Rather than ending at candidate generation, the design process incorporates evaluation and further exploration.

The researchers position generative design and structure-predictor-based hallucination or search as complementary rather than mutually exclusive approaches.

Proteina-Complexa was selected as an Oral presentation at ICLR 2026. [12]

A separate September 2026 preprint also reported large-scale experimental investigations using latent generative search. [19]

These developments illustrate a shift toward treating protein design as an integrated system of generation, evaluation, and optimization.

14.2. BoltzGen: Integrating Structure Prediction and Protein Design#

Another important study is BoltzGen.

A revised BoltzGen bioRxiv preprint was released in June 2026. [13]

BoltzGen uses an all-atom diffusion framework to integrate structure prediction and protein design.

Conceptually:

$$ \boxed{\text{Structure Prediction + Protein Design}} $$

It extends binder design across different molecular targets and contexts, including:

• Protein
• Peptide
• Nanobody
• Small molecule
• Disordered protein region

One notable feature of BoltzGen is its emphasis on experimental validation.

The study reports eight experimental campaigns spanning 26 targets.

It also demonstrates binders with nanomolar-range affinity against targets with limited similarity to previously characterized protein complexes. [13]

This reflects a broader change in how protein generative models are evaluated.

Around 2023, an important question was whether a generated sequence was predicted by AlphaFold or related tools to adopt the intended fold.

Increasingly, researchers also ask whether the protein can be produced experimentally and whether it demonstrates the desired binding or biological activity.

The emphasis is expanding from:

Does the generated protein appear structurally plausible in a prediction model?

to:

Can the generated protein be produced and experimentally validated for its intended interaction or function?

This shift is essential for bringing protein generative AI closer to practical drug discovery applications.

15. Reconsidering EvoDiff's Fundamental Limitations in Historical Context#

Viewed in the context of subsequent research, EvoDiff's limitations are not simply a consequence of its age or its focus on sequences.

The deeper issue concerns the gap between learning natural sequence distributions and optimizing artificially specified functions.

15.1. Natural Sequence Distributions Are Not the Same as Desired Functional Distributions#

EvoDiff begins with the distribution of sequences produced by natural evolution.

Drug discovery and protein engineering, however, require proteins that satisfy artificial design objectives.

A naturally functional protein may still fail to meet requirements such as:

• High target specificity
• Sufficient binding affinity or biological activity
• Good solubility
• Thermal and storage stability
• Adequate expression and manufacturability
• Low aggregation propensity
• Acceptable immunogenicity and toxicity profiles

Reproducing natural sequence statistics does not automatically optimize these properties.

Consequently, practical protein design requires methods for evaluating, selecting, and improving generated candidates according to the intended application.

15.2. Flexible Generation Is Not the Same as Flexible Functional Optimization#

EvoDiff can generate different regions of a protein sequence under various constraints.

However, this does not imply direct control over arbitrary physicochemical properties or biological functions.

Even when a functional motif is preserved, the resulting protein may still need to satisfy multiple conditions simultaneously:

• Correct three-dimensional geometry
• Appropriate side-chain interactions
• Complementarity with a target molecule
• High specificity
• Good expression
• Adequate stability

This distinction between generative flexibility and functional controllability is central to understanding EvoDiff's limitations.

15.3. Sequence-Only and Multimodal Approaches Are Not Simply Better or Worse#

Subsequent models increasingly integrate structural, functional, and interaction information.

However, this should not be interpreted as evidence that sequence-only approaches have failed.

Different protein design problems require different information.

Designing certain IDR-associated functions may benefit strongly from sequence-based representations.

Designing a binder against a complex molecular surface may require precise structural and interaction constraints.

Additional information also comes with costs, including requirements for training data, computational resources, and model complexity.

A more useful question is therefore:

For a specific protein design task, how much additional performance is gained by incorporating structural, functional, or interaction information?

This question cannot be answered merely by comparing the release dates or complexity of different models.

16. What Important Research Questions Remain Unanswered?#

EvoDiff's results and the broader history of protein generative AI suggest several important directions for further investigation.

16.1. Does Agreement with Natural Sequence Distributions Predict Functional Success?#

This is perhaps the most direct way to test EvoDiff's foundational hypothesis.

Are sequences with higher likelihood under a learned natural protein distribution more likely to be functional experimentally?

Does this relationship depend on the protein family or desired function?

Addressing this question requires measuring sequence likelihood, novelty, structural metrics, and experimental activity across the same candidate libraries.

It is also important to control for family membership and sequence similarity rather than relying on simple correlations.

16.2. Under What Conditions Does EvoDiff Outperform Alternative Models?#

EvoDiff's main architectural advantage is flexible, order-agnostic conditional generation.

However, it remains important to determine when this advantage translates into higher experimental success rates.

Comparisons with autoregressive protein language models, DPLM, RFdiffusion, and other methods should control, as far as possible, for:

• Training data
• Model size
• Training and inference compute
• Number of generated candidates
• Computational selection procedures
• Number of experimentally tested candidates
• Success criteria

Such comparisons would help isolate the contribution of the generative method itself.

16.3. Can EvoDiff Generalize to Unseen Protein Families and Novel Functions?#

Generalization beyond familiar protein families remains a major challenge.

Evaluations should control the relationship between training and test sequences.

Approaches such as family-held-out benchmarks and low-sequence-identity splits may be useful.

It is also necessary to distinguish between generating new variants around known functional motifs and designing genuinely novel functions.

16.4. Can Target Specificity and Developability Be Controlled?#

Practical binder design requires more than binding to a desired target.

It may also require avoiding closely related off-target molecules.

Likewise, experimentally observed binding does not establish therapeutic suitability.

Expression, solubility, aggregation, stability, immunogenicity, and other developability properties must be evaluated.

Whether sequence-first generative models can reliably optimize these properties simultaneously remains an important research question.

16.5. Can EvoDiff Design Diverse IDR Functions?#

EvoDiff's results involving mitochondrial targeting signals are particularly interesting because they demonstrate the potential of sequence-based design for regions without a single stable structure.

However, IDRs support many other biological functions, including:

• Phase separation
• Regulation of molecular interactions
• Signal transduction
• Post-translational modification-dependent regulation
• Functions involving dynamic conformational ensembles

Whether sequence diffusion can reliably design these functions remains largely unresolved.

Suitable experimental evaluation methods will also be essential, particularly where a single predicted structure is insufficient.

16.6. How Much Can Integrated Generation, Evaluation, Search, and Experimentation Improve Performance?#

EvoDiff's experiments already demonstrate the importance of combining generation with computational selection.

Future workflows may increasingly integrate structure predictors, functional predictors, experimental feedback, and iterative optimization.

A representative process is:

Generate → Predict → Select → Experiment → Update

This idea is closely related to the generation-and-search approaches explored in studies such as Proteina-Complexa.

However, improvement is not guaranteed.

Optimization may exploit errors in predictive models, producing candidates that score well computationally without performing well experimentally.

Careful experimental validation remains essential.

16.7. Prioritizing the Remaining Research Questions#

The major unresolved questions can be organized by their relevance to practical applications.

Priority	Research Question	Required Validation
High	Relationship between natural sequence likelihood and functional success	Compare likelihood, novelty, structural metrics, and experimental activity
High	Comparative advantage over alternative generative models	Controlled comparisons under matched data, compute, and experimental budgets
High	Generalization to unseen protein families	Family-held-out and low-sequence-identity evaluations
High	Control of target specificity	Binding measurements against intended and unintended targets
High	Therapeutic developability	Expression, solubility, aggregation, thermal stability, and immunogenicity assessments
Medium	Generalization across IDR functions	Experiments covering IDR functions beyond localization
Medium	Dependence on MSA quality and depth	Systematic evaluations of MSA depth, phylogenetic bias, and noise
Medium	Flexible sequence length and indel design	Experiments involving different lengths and domain organizations
Medium	Contributions of sequence, structure, and function information	Controlled sequence-only versus multimodal comparisons
Medium	Effectiveness of iterative optimization	Property guidance, active learning, and experimental-feedback evaluations

Among these questions, three are particularly important:

Experimental functional success rates, generalization to unfamiliar design problems, and design efficiency under realistic experimental budgets.

A useful generative model should not be evaluated solely by whether it can produce at least one successful candidate.

It should also be evaluated by how efficiently it can identify useful candidates under practical constraints.

17. A Historical Overview of Protein Generative AI#

The research developments discussed so far can be organized into several major directions.

However, the following sequence does not imply that each generation of models replaced the previous one.

Sequence-based and structure-based approaches have developed in parallel, with increasing efforts to combine their respective strengths.

17.1. Learning from Evolutionary Sequences#

DeepSequence / UniRep

These studies demonstrated that evolutionary sequence data can be used to learn biologically meaningful statistical constraints and representations.

17.2. Sequence Generation with Protein Language Models#

ProGen / ProGen2 / ProtGPT2 / ESM-family models

Research expanded toward large-scale protein language modeling, direct sequence generation, and general-purpose representation learning.

Not every model in the ESM family uses the same generative formulation, and some primarily focus on representation learning.

17.3. Structure-Conditioned Protein Design#

ProteinMPNN

Inverse folding made it practical to design amino acid sequences conditioned on specified three-dimensional structures.

17.4. Structure Diffusion#

RFdiffusion / Chroma (2023)

Diffusion models enabled substantial advances in generating three-dimensional protein backbones.

17.5. Sequence Diffusion#

EvoDiff (2023)

EvoDiff demonstrated that evolutionary sequence data could directly support diffusion-based protein generation and flexible conditional sequence design.

17.6. Diffusion Protein Foundation Models#

DPLM (2024)

Sequence diffusion expanded beyond generation toward general-purpose protein representation learning.

17.7. Multimodal Protein Generation#

ESM3 / DPLM-2 (2024–2025)

Generative models increasingly integrated protein sequence, structure, and function.

17.8. Flow Matching and All-Atom Generation#

Proteina / La-Proteina (2025–2026)

Research advanced toward large-scale backbone generation and joint generation of sequences with detailed atomic structures.

17.9. Target-Conditioned Design and Integrated Search#

BoltzGen / Proteina-Complexa (2026)

All-atom interaction modeling, target-conditioned binder design, test-time search, and experimental validation became increasingly prominent.

17.10. Comparison of Major Research Directions#

Period	Representative Models	Primary Generation or Learning Target	Main Research Direction
Approximately 2017–2019	DeepSequence, UniRep	Protein sequences	Statistical modeling and representation learning from evolutionary data
Approximately 2020–2023	ProGen, ProGen2, ProtGPT2, ESM family	Protein sequences	Protein language modeling and direct generation
Approximately 2022–2023	ProteinMPNN	Structure to sequence	Inverse folding
2023	RFdiffusion, Chroma	3D backbone	Structure-based diffusion generation
2023	EvoDiff	Evolutionary sequences	Sequence-first diffusion
2024	DPLM	Protein sequences	Diffusion protein foundation models
2024–2025	ESM3, DPLM-2	Sequence, structure, function	Multimodal generation
2025–2026	Proteina, La-Proteina	Backbone, sequence, atomic structure	Flow matching and all-atom generation
2026	BoltzGen, Proteina-Complexa	Atomic structures and target interactions	Target-conditioned design, integrated search, and experimental validation

Several important developments emerge from this history.

First, learning the statistical regularities of evolutionary protein sequences expanded from mutation effect prediction into direct protein generation.

Second, structure-centered generative models enabled explicit control of three-dimensional geometry.

Third, researchers increasingly explored integrating sequence, structure, function, and molecular interactions.

Finally, experimental success rates and practical design efficiency became increasingly important alongside computational novelty and structural plausibility.

18. Reassessing EvoDiff's Historical Significance#

EvoDiff's historical importance extends beyond applying diffusion models to protein sequences.

It demonstrated that evolutionary sequence distributions could serve as a meaningful generative space and that flexible conditional generation was possible within that space.

At the same time, it brought a more fundamental scientific problem into focus.

18.1. What EvoDiff Has Demonstrated#

The available results support the following conclusions:

• Large-scale evolutionary sequence data can be used to generate novel protein sequences directly.
• Order-agnostic generation enables flexible completion of different sequence regions.
• Functional motifs and MSA information can be incorporated into sequence generation.
• Sequence-based approaches can be useful for certain regions that lack a single stable three-dimensional structure.
• Selected generated candidates have experimentally demonstrated expression, secondary-structure properties, localization, metal-binding-related behavior, and protein–protein binding.

These are important results supporting the potential of sequence-first protein design.

18.2. What EvoDiff Has Not Demonstrated#

The following claims remain unproven:

• That accurately learning natural protein sequence distributions enables reliable design of arbitrary desired functions.
• That sequence diffusion is universally superior to other generative methods.
• That successful design within familiar protein families generalizes equally well to novel functions and targets.
• That sequence novelty guarantees functional novelty or practical utility.
• That EvoDiff alone can adequately optimize specificity, stability, expression, immunogenicity, and other developability requirements.
• That structural, functional, or target-specific information is generally unnecessary for protein design.

These limitations should not be viewed as failures of EvoDiff.

Instead, they help define the next scientific questions that protein generative AI must address.

18.3. What Subsequent Research Is Trying to Achieve#

DPLM explored the integration of sequence generation and representation learning.

DPLM-2 and ESM3 incorporated structural and functional information.

La-Proteina advanced joint generation of sequences and atomic-level structures.

BoltzGen emphasized target-conditioned design and experimental validation.

Proteina-Complexa integrated generation with search and optimization.

These developments are not merely attempts to build increasingly complex versions of EvoDiff.

They represent efforts to move from learning the distribution of naturally occurring proteins toward designing proteins that satisfy explicitly specified objectives.

19. Conclusion: Did EvoDiff Prove That "Sequence Is All You Need"?#

EvoDiff can be positioned historically as follows:

EvoDiff was an important 2023 study that demonstrated the potential of sequence-first protein design by treating evolutionary protein sequences—not three-dimensional structures—as the primary generative space, at a time when structure-based diffusion models were rapidly advancing.

Its significance can be summarized in three main contributions.

First, EvoDiff emphasized the value of using large-scale evolutionary sequence data rather than restricting protein generation to experimentally determined structural datasets.

Second, it combined order-agnostic diffusion with flexible conditional generation, enabling applications such as sequence inpainting and motif scaffolding.

Third, it demonstrated the potential of sequence-based generation for protein regions that do not necessarily adopt a single stable three-dimensional structure.

The experimental results reported in 2026 provided further evidence that these approaches can produce candidates with experimentally observable biological properties.

However, the central assumption underlying EvoDiff remains only partially validated.

That assumption is:

To what extent can learning protein sequence distributions shaped by natural evolution be generalized to the realization of artificially specified biological functions?

Learning statistical regularities from naturally occurring proteins is not necessarily equivalent to designing proteins for the requirements of drug discovery and protein engineering.

EvoDiff did not completely bridge this gap.

Nor does generating sequences that resemble natural proteins guarantee experimental success, target specificity, or efficient design under realistic budgets.

From this perspective, the history of protein generative AI cannot be described simply as a progression from sequence-only models toward multimodal models.

Sequence-based and structure-based methods have distinct advantages and limitations.

Recent research—including DPLM, ESM3, La-Proteina, BoltzGen, and Proteina-Complexa—has increasingly explored how these different sources of information can be combined to control desired structures, functions, and interactions more directly.

For protein generative AI to become a reliable engineering technology, evaluation must extend beyond sequence novelty and computational structural plausibility.

It must also consider experimental functional success rates, target specificity, stability, expression, practical screening costs, and other properties relevant to real-world applications.

EvoDiff's historical contribution was not proving that sequence information is always sufficient.

Its contribution was establishing evolutionary sequence space as a meaningful generative domain and making the question of how far evolutionary information can support artificial protein design both concrete and experimentally testable.

That question remains open.

The continuing development of models integrating sequence, structure, function, and molecular interactions can be understood as part of the scientific effort to answer it.

EvoDiff therefore represents both an important milestone in the history of protein generative AI and a starting point for investigating a fundamental distinction:

The difference between being able to generate a protein and being able to design a protein that reliably performs a desired function.

References#

[1] Riesselman et al., “Deep generative models of genetic variation capture the effects of mutations,” Nature Methods, 15, 816–822, 2018. https://www.nature.com/articles/s41592-018-0138-4.pdf

[2] Alley et al., “Unified rational protein engineering with sequence-based deep representation learning,” Nature Methods, 16, 1315–1322, 2019. https://www.nature.com/articles/s41592-019-0598-1

[3] Madani et al., “Large language models generate functional protein sequences across diverse families,” Nature Biotechnology, 41, 1099–1106, 2023. https://preview-www.nature.com/articles/s41587-022-01618-2.pdf

[4] Ingraham et al., “Illuminating protein space with a programmable generative model,” Nature, 623, 1070–1078, 2023. https://www.nature.com/articles/s41586-023-06728-8

[5] Alamdari et al., “Protein generation with evolutionary diffusion: sequence is all you need,” bioRxiv, 2023. https://www.biorxiv.org/content/10.1101/2023.09.11.556673v1

[6] Alamdari et al., “Protein generation with evolutionary diffusion: sequence is all you need,” bioRxiv, revised preprint, 2023. https://www.biorxiv.org/content/10.1101/2023.09.11.556673v2.full

[7] Wang et al., “DPLM: A Diffusion Protein Language Model,” Proceedings of the 41st International Conference on Machine Learning (ICML), 2024. https://arxiv.org/abs/2402.18567

[8] Wang et al., “DPLM-2: A Multimodal Diffusion Protein Language Model,” International Conference on Learning Representations (ICLR), 2025. https://proceedings.iclr.cc/paper_files/paper/2025/hash/57c30b677add9aa78e1745f0643104d0-Abstract-Conference.html

[9] Hayes et al., “Simulating 500 million years of evolution with a language model,” Science, 387, 850–858, 2025. https://pubmed.ncbi.nlm.nih.gov/39818825/

[10] “Proteina: Scaling Flow-based Protein Structure Generative Models,” International Conference on Learning Representations (ICLR), 2025. https://proceedings.iclr.cc/paper_files/paper/2025/hash/f4e9121ad30cd4e5528042fbfd835b3f-Abstract-Conference.html

[11] Geffner et al., “La-Proteina: Atomistic Protein Generation via Partially Latent Flow Matching,” International Conference on Learning Representations (ICLR), 2026. https://research.nvidia.com/publication/2026-01_la-proteina-atomistic-protein-generation-partially-latent-flow-matching

[12] Didi et al., “Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute,” International Conference on Learning Representations (ICLR), 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/639d992f819c2b40387d4d5170b8ffd7-Abstract-Conference.html

[13] Stark et al., “BoltzGen: Toward Universal Binder Design,” bioRxiv, revised preprint, 2026. https://www.biorxiv.org/content/10.1101/2025.11.20.689494v2.full

[14] Alamdari et al., “Protein generation with evolutionary diffusion: sequence is all you need,” eLife, Reviewed Preprint, 2026. DOI: 10.7554/eLife.112029.1. https://elifesciences.org/reviewed-preprints/112029.pdf

[15] Dauparas et al., “Robust deep learning–based protein sequence design using ProteinMPNN,” Science, 378, 49–56, 2022. https://doi.org/10.1126/science.add2187

[16] Watson et al., “De novo design of protein structure and function with RFdiffusion,” Nature, 620, 1089–1100, 2023. https://doi.org/10.1038/s41586-023-06415-8

[17] eLife, “Peer review of Protein generation with evolutionary diffusion: sequence is all you need,” 2026. https://elifesciences.org/reviewed-preprints/112029v1/reviews

[18] “EvoFlows,” Subsequent research on evolutionary generative modeling of protein MSAs and the modeling of insertions and deletions. Referenced as related research.

[19] Didi et al., “Latent generative search unlocks de novo design of untapped biomolecular interactions at scale,” bioRxiv, 2026. https://www.biorxiv.org/content/10.64898/2026.09.12.751118v1