Target Audience

  • Researchers interested in applying EvoDiff, DPLM, and other protein generative AI models to personalized medicine and therapeutic development.
  • Entrepreneurs and startup founders exploring the commercialization of AI-driven protein design technologies.
  • Readers seeking a deeper understanding of EvoDiff's assumptions, experimental validation, technical limitations, and implications for practical drug discovery.

Executive Summary

  • EvoDiff is a protein generative model that learns probability distributions from naturally evolved protein sequences and uses their statistical regularities to generate novel sequences.
  • One of EvoDiff's fundamental implicit assumptions is that learning evolutionary constraints from natural sequences can help design functional proteins, including sequences not previously observed in nature.
  • The 2026 eLife Reviewed Preprint reports experimental validation involving protein expression, secondary structure, mitochondrial targeting signals, metal binding, and protein–protein interactions.
  • However, generating natural-like sequences does not guarantee that the resulting proteins will exhibit desired biological functions. EvoDiff's general superiority over alternative generative approaches has also not been established.
  • For applications in personalized medicine and therapeutic development, evaluation must extend beyond sequence novelty and predicted structural plausibility to include functional activity, specificity, stability, expression, developability, and experimental cost.

1. What Has EvoDiff Actually Demonstrated, and What Remains Unproven?

EvoDiff occupies an important position in the history of protein generative AI.

Introduced by Alamdari and colleagues in 2023, EvoDiff demonstrated an approach to generating amino acid sequences directly using discrete diffusion models trained on evolutionary protein sequence data.[1]

Unlike structure-first approaches, which typically begin by generating a protein's three-dimensional structure, EvoDiff emphasizes a sequence-first approach in which amino acid sequences themselves are the primary objects of generation.

However, the significance of EvoDiff goes beyond a simple transition from structure-first to sequence-first protein design.

EvoDiff is based on the hypothesis that learning the probability distribution of naturally evolved protein sequences enables a model to capture structural and functional constraints embedded in that distribution and use them to design new proteins.

This hypothesis rests on an important implicit assumption:

If a model learns the statistical regularities of naturally occurring protein sequences sufficiently well, that knowledge should also be useful for designing functional protein sequences that have not been observed in nature.

However, three different capabilities must be distinguished:

  1. Generating natural-like protein sequences.
  2. Generating proteins that actually fold and exhibit biological functions.
  3. Designing practically useful proteins that satisfy specific objectives.

These are not equivalent.

For example, a generated sequence may statistically resemble naturally occurring proteins without necessarily folding into a stable three-dimensional structure.

Even if the resulting protein folds, it may not bind a desired target or catalyze a particular reaction.

Furthermore, even if biological activity or target binding is demonstrated, the protein may not satisfy the specificity, safety, stability, or manufacturability requirements of a therapeutic product.

The 2026 EvoDiff study added experimental validation of protein expression, secondary structure, subcellular localization, metal binding, and protein–protein interactions.[2]

EvoDiff is therefore more than a purely computational proof of concept.

Nevertheless, the public eLife reviews acknowledge the importance of the work while raising questions about its comparative advantages over other generation methods and the generalizability of its results across protein design applications.[3]

The distinction can be summarized as follows.

Research question Current assessment
Can new protein sequences be generated using sequence information alone? Demonstrated
Can sequence diffusion support flexible conditional protein design? Demonstrated across several applications
Can desired functional proteins be designed reliably and with high success rates across diverse tasks? Experimental validation remains limited

This distinction is particularly important when considering EvoDiff for personalized medicine and therapeutic development.

2. The Implicit Assumptions Underlying EvoDiff

To understand EvoDiff's scientific significance and limitations, it is necessary to examine the assumptions underlying its design.

The following discussion distinguishes ideas explicitly presented in the original research from assumptions inferred from the model architecture, training objectives, and evaluation methods.

2.1. Assumption 1: The Distribution of Naturally Evolved Protein Sequences Is Useful for Protein Design

This is arguably EvoDiff's most fundamental assumption.

EvoDiff learns from approximately 42 million protein sequences derived from UniRef50 and generates new sequences that resemble the learned distribution.[1]

Conceptually, its objective can be expressed as:

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

Here:

  • $x$ represents an amino acid sequence.
  • $P_{\mathrm{natural}}$ represents the distribution of naturally observed protein sequences.
  • $P_{\theta}$ represents the distribution learned by the generative model.

However, the distribution we ultimately want in protein engineering may instead be:

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

where:

  • $f_{\mathrm{desired}}$ represents a desired function.
  • $c$ represents the environmental conditions and design constraints under which the protein must operate.

These two distributions are not necessarily equivalent.

Natural evolution did not directly optimize proteins for the specific requirements of drug development, such as:

  • High binding specificity against a therapeutic target.
  • Low immunogenicity.
  • High production yield.
  • Long-term storage stability.

Instead, natural selection has shaped proteins under multiple constraints associated with the survival and reproduction of organisms in their respective environments.

EvoDiff therefore implicitly assumes that:

The sequence space explored by natural evolution contains sufficient information to help solve artificial protein design problems.

This is a reasonable assumption.

Indeed, EvoDiff has demonstrated that statistical regularities learned from natural sequences can support the generation of functional candidates in several design tasks.[2]

However, these results do not establish that learning the natural protein sequence distribution is sufficient for designing arbitrary functions required in personalized medicine or drug discovery.

2.2. Assumption 2: Structural and Functional Constraints Can Be Recovered from Sequence Statistics

EvoDiff can generate protein sequences without necessarily requiring explicit three-dimensional structural information or information about a target molecule.

This approach assumes that structural and functional constraints can be learned indirectly from amino acid co-occurrence patterns, evolutionary conservation, and other sequence-level statistical relationships.

However, the following implication does not generally hold:

$$ \text{Sequence statistics} \not\Rightarrow \text{Guaranteed biological function} $$

Learning statistical regularities in protein sequences is not equivalent to guaranteeing biological function.

For example, a sequence may contain the amino acids required for metal binding without positioning them correctly in three-dimensional space.

Similarly, catalytic residues may be conserved while the resulting structure prevents substrates from accessing the active site.

Protein function also depends on factors beyond the amino acid sequence, including:

  • Temperature.
  • pH.
  • Solvent environment.
  • Cofactors.
  • Subcellular localization.
  • Post-translational modifications.
  • Molecular interaction partners.

What EvoDiff has demonstrated is that a model trained on sequence information alone can produce proteins that satisfy some of these constraints in particular experimental settings.

It has not demonstrated that these constraints can be controlled reliably under arbitrary conditions.

2.3. Assumption 3: Learning High-Density Regions of Natural Sequence Space Enables the Discovery of Novel Functional Proteins

This assumption is especially relevant to de novo protein design.

Naturally observed protein sequences occupy only a tiny fraction of the theoretically possible amino acid sequence space.

EvoDiff can generate novel sequences rather than merely reproducing known proteins.[1]

However, the following properties are fundamentally different:

  • A sequence differs from known sequences.
  • A sequence produces a novel three-dimensional structure.
  • A sequence enables a novel biological function.

Sequence novelty is not sufficient to establish functional novelty.

Furthermore, when generation moves far beyond the training distribution, there is no guarantee that statistical relationships learned from natural sequences will continue to preserve functional plausibility.

This is the general problem of out-of-distribution generalization.

For drug discovery applications involving new targets or previously unexplored biochemical functions, this distinction is particularly important.

A protein may be sequence-novel without possessing a genuinely novel or therapeutically useful function.

2.4. Assumption 4: Flexible Sequence Completion Translates into Greater Design Flexibility

One of EvoDiff's major technical contributions is flexible conditional generation using an Order-Agnostic Autoregressive Diffusion Model (OADM).[1]

Whereas conventional autoregressive models typically generate sequences from left to right, OADM allows sequence positions to be generated in different orders.

This enables applications such as:

  • Sequence inpainting.
  • Completing missing regions.
  • Preserving functional motifs while generating surrounding sequences.
  • Generating sequences conditioned on specified residues.

However, flexibility in conditioning is not equivalent to reliably satisfying biological design objectives.

For example, fixing the residues of an active site and generating the surrounding sequence does not guarantee:

  • Correct distances between catalytic residues.
  • Appropriate side-chain orientations.
  • Productive substrate interactions.
  • Specific binding to a target molecule.

The original EvoDiff study evaluated 17 motif-scaffolding tasks.[1]

Its evaluation used structural prediction metrics, including pLDDT and motif RMSD, to assess generated candidates.

This reveals an important distinction.

Although EvoDiff can perform sequence generation without explicitly generating structures, evaluating the resulting designs is not necessarily structure-free.

This distinction matters when assessing the actual capabilities and limitations of sequence-first protein design.

2.5. Assumption 5: Evolutionary Covariation in MSAs Represents Functional Constraints Useful for Design

EvoDiff-MSA uses evolutionary information contained in Multiple Sequence Alignments (MSAs).[1]

Covariation among related protein sequences can reflect structural or functional constraints that have been preserved during evolution.

However, MSAs do not exclusively reflect functional selection.

Their statistical properties are also influenced by:

  • Phylogenetic relationships.
  • Sequence sampling bias.
  • Alignment quality.
  • The number and diversity of homologous sequences.

As a result, performance may differ when the model is applied to:

  • Novel protein families with few known homologs.
  • Artificially designed proteins.
  • Rare or poorly characterized protein families.

The original EvoDiff study reported cases in which motif-scaffolding performance decreased when the number of sequences in an MSA fell outside the range encountered during training.[1]

Therefore, the availability and quality of evolutionary information should be considered when evaluating EvoDiff-MSA for practical protein design.

2.6. Assumption 6: Computational Evaluation Metrics Are Predictive of Experimental Success

This assumption is not unique to EvoDiff.

It is a fundamental issue throughout protein generative AI.

EvoDiff uses several computational metrics to evaluate the quality of generated sequences.[1][2]

Computational metric What it aims to measure What it does not directly measure
Perplexity How well the model represents natural sequence distributions Actual biological function
pLDDT Confidence in predicted protein structures Experimental stability or biological activity
Self-consistency Agreement between sequence and predicted structural properties Biological function
ProtT5 embedding distance Similarity to natural protein distributions in representation space Correspondence with the true distribution of biological functions
Sequence novelty Differences from known sequences Novel function or practical utility

One particularly important example is the Fréchet ProtT5 Distance.

This metric measures how similar generated and reference sequence distributions are in the embedding space of a protein language model.

It does not directly measure the distribution of experimentally verified biological functions.

Consequently, a claim that generated proteins cover a broad region of natural protein representation space should not be interpreted as proof that the model has experimentally produced a comparably broad range of biological functions.

This distinction is crucial when evaluating generative models for personalized medicine and therapeutic development.

3. Major Challenges and Limitations of EvoDiff

When evaluating EvoDiff's limitations, three different categories should be distinguished:

  1. Limitations arising from the model's underlying principles or architecture.
  2. Limitations associated with its current implementation.
  3. Open questions that cannot yet be resolved because experimental validation or comparative studies are insufficient.

The following sections examine these issues separately.

3.1. Experimental Success Rates Are Not Necessarily High

The 2026 eLife Reviewed Preprint includes experimental evaluation of proteins generated unconditionally by EvoDiff-Seq.[2]

The reported evaluation involved:

  • 1,000 initially generated candidates.
  • 25 candidates selected for experimental evaluation.
  • 4 candidates exhibiting expression and secondary structure consistent with expectations.

Thus, 4 out of 25 experimentally evaluated candidates, or 16%, satisfied the reported criteria.

However, this 16% figure should not be interpreted as the unconditional success probability of the model.

The 25 experimentally tested candidates were selected from the original 1,000 using computational criteria, including structural prediction metrics.

Moreover, the four successful candidates were evaluated for expression and secondary-structure consistency using circular dichroism (CD) spectroscopy.

These experiments did not establish specific biological activities for those four proteins.

Nevertheless, the result is important.

A protein sequence that appears computationally plausible does not necessarily produce an experimentally successful protein.

For applications in personalized medicine and therapeutic development, the ability to generate candidate sequences is only one part of the problem.

The probability of obtaining experimentally useful proteins is equally important.

3.2. EvoDiff Does Not Consistently Outperform Autoregressive Models

One of EvoDiff's main strengths is flexible conditional generation.

However, it does not outperform existing methods on every evaluation metric.

The original study reports that conventional left-to-right autoregressive models perform better on certain measures of natural sequence distribution modeling and structural plausibility.[1]

Furthermore, Reviewer #1 of the 2026 eLife public review noted that EvoDiff did not demonstrate a substantial improvement in experimental success rates over other sequence generation methods.[3]

This observation is important when interpreting EvoDiff's historical contribution.

EvoDiff's main achievement was not showing that it outperformed all previous protein language models across every metric.

Instead, its contribution was:

Combining order-agnostic conditional generation with large-scale learning from evolutionary protein sequences.

For practical protein engineering, model selection should therefore depend not only on general generation quality but also on conditional design flexibility and experimentally measured success rates.

3.3. Geometric Constraints Are Difficult to Control Directly

EvoDiff can preserve specified amino acid motifs during generation.

However, its standard sequence-first framework does not provide a general mechanism for directly specifying the three-dimensional positions and orientations of residues.

This distinguishes EvoDiff from structure-conditioned approaches such as RFdiffusion.

For example, when designing a binder for a specific epitope, success depends on more than whether the generated sequence resembles natural proteins.

Important constraints include:

  • Shape complementarity.
  • Electrostatic complementarity.
  • Chemical interactions.
  • Appropriate three-dimensional positioning.

Standard sequence-first generation does not directly optimize all these constraints.

Consequently, EvoDiff-based workflows may still depend on subsequent structure prediction, complex evaluation, and experimental screening.

3.4. Generalization to Design Objectives Not Represented in Natural Evolution Remains Uncertain

Even a model trained on millions of naturally occurring sequences cannot guarantee performance outside its training distribution.

Examples of challenging design objectives include:

  • Extreme stability in non-natural environments.
  • High selectivity for artificial substrates.
  • Precise discrimination among closely related molecules.
  • Engineered molecular switches.

EvoDiff has demonstrated utility across multiple protein design tasks.[2]

However, it has not systematically established general-purpose design capabilities for arbitrary novel functions.

This limitation becomes particularly important when developing therapies against new molecular targets or attempting to engineer functions that are uncommon or absent in nature.

3.5. A Gap Remains Between Computational and Experimental Evaluation

The unconditional generation experiments included cases in which candidates with favorable predicted pLDDT values did not achieve experimental success.[2]

This is not exclusively an EvoDiff problem.

It reflects a broader limitation of computational screening based on predicted protein structures.

Several distinct properties need to be evaluated independently:

  • Sequence likelihood.
  • Confidence in predicted structure.
  • Actual protein folding.
  • Protein expression.
  • Biological function.

Therefore, EvoDiff should not necessarily be treated as a complete protein design solution.

It may be more appropriate to use it as one component within a broader system combining generation, evaluation, candidate selection, experiments, and iterative improvement.

3.6. OADM Can Generate Positions in Arbitrary Orders but Does Not Freely Revise Previously Generated Residues

This is an easily overlooked architectural limitation.

Although OADM allows residues to be generated in different orders, once a residue is selected, it is generally fixed during the subsequent generation process.

Consequently, the basic approach is not naturally suited to repeatedly revising previously generated residues in response to feedback from a functional predictor.

The EvoDiff authors discuss the possibility that D3PM may be more compatible with function-guided generation because it permits residues to be edited during successive diffusion steps.[1][2]

In other words:

Order-agnostic generation in EvoDiff-OADM is not equivalent to iterative sequence optimization.

This distinction matters when integrating a generative model with search algorithms, functional predictors, or optimization systems.

3.7. Insertions, Deletions, and Sequence-Length Design Remain Challenging

In the standard EvoDiff-Seq generation and inpainting framework, the sequence length and regions to be completed are specified in advance.

However, natural protein evolution involves more than amino acid substitutions.

Insertions, deletions, and domain rearrangements also contribute to the evolution of protein structure and function.

EvoDiff-MSA can process alignments containing gaps, but subsequent work on EvoFlows reported difficulties in controlling insertion generation using EvoDiff-MSA.[4]

This presents a limitation when exploring designs that require flexible sequence lengths or changes in domain architecture.

3.8. Large-Scale Sequence Datasets Also Contain Biases

One of EvoDiff's main advantages is its ability to learn from large quantities of naturally occurring protein sequences.

However, these datasets do not represent all species or biological functions equally.

Even after sequence clustering, UniRef50 may retain biases associated with:

  • Organisms that are frequently studied.
  • Organisms that are easier to sample and sequence.
  • Protein families that are overrepresented in available data.

Therefore, diversity among generated sequences should not be confused with comprehensive coverage of biological function space.

Moreover, EvoDiff-MSA and single-sequence EvoDiff models use different forms of training information.

Consequently, differences in their performance cannot necessarily be attributed solely to how evolutionary information is represented.

4. What Has Been Experimentally Validated in 2026, and What Remains Unvalidated?

The 2026 eLife version of EvoDiff reports several wet-lab experiments.[2]

These are important for evaluating the practical utility of sequence-first protein generation.

However, each experiment supports a different level of scientific claim.

4.1. Summary of Experimental Results

Experimental application Reported results What the evidence supports What remains unproven
Unconditional generation 4 of 25 candidates showed expression and expected secondary-structure characteristics Novel sequences can yield experimentally viable protein candidates Broad functional diversity and high success rates
Mitochondrial targeting signals 8 selected candidates showed localization results consistent with Cox15 Sequence completion can reproduce a particular intrinsically disordered region (IDR) function Generalization to arbitrary IDRs or other cellular functions
Calcium-binding motifs 12 of 13 candidates were expressed; 1 showed a spectral change suggestive of Ca²⁺ binding Conditional sequence generation can produce potentially functional scaffolds General metal-binding capability and precise affinity control
p53–MDM2 binding 4 of 24 candidates exhibited $K_D<50$ nM; 8 showed evidence of binding Protein binders preserving a specific interaction motif can be generated Activation of p53 signaling in cancer cells or therapeutic efficacy

These experimental findings can be interpreted in two ways.

First, EvoDiff's sequence-first approach has demonstrated experimental utility across different biological phenomena, including IDR-related functions and molecular interactions.

This is a meaningful scientific achievement.

Second, these experiments involved relatively small sets of selected candidates.

They do not establish that comparable success rates can be achieved for arbitrary unknown targets or novel biological functions.

The p53–MDM2 experiment illustrates this distinction.

The experiments demonstrated binding to the target.

They did not demonstrate intracellular competitive inhibition, activation of p53 signaling, or antitumor efficacy.

4.2. The Performance of EvoDiff Alone Is Different from the Performance of the Entire Design Pipeline

When interpreting EvoDiff's experimental results, it is important to distinguish the generative model itself from the broader design workflow.

For example, the p53–MDM2 binding experiment involved a process conceptually summarized as follows:[2]

  1. EvoDiff generated 1,300 candidate sequences.
  2. Structural plausibility was evaluated using tools including OmegaFold, and candidates were selected.
  3. A subset of 24 candidates underwent experimental testing.
  4. Biolayer interferometry (BLI) measurements identified four candidates with strong binding.

This result should not be interpreted as evidence that EvoDiff alone can reliably predict or generate a specified binding affinity.

The outcome reflects a combination of sequence generation, structure prediction, candidate selection, and experimental validation.

Furthermore, because not all 1,300 generated candidates were experimentally evaluated, the ratio of 4/1,300 should not be treated as the model's intrinsic experimental success rate.

This distinction is particularly relevant when assessing the cost-effectiveness of AI-driven protein design in therapeutic development.

5. What Did the Public Reviewers Identify as EvoDiff's Main Weaknesses?

The 2026 eLife public reviews provide direct insight into the scientific limitations of EvoDiff.[3]

A shared theme in the comments from Reviewer #1 and Reviewer #2 is that, while EvoDiff is recognized as a useful contribution, the general comparative advantage of diffusion-based generation and the broad applicability of a sequence-only approach remain insufficiently established.

5.1. Comparative Advantages over Other Generative Methods

Reviewer #1 raised concerns that EvoDiff did not demonstrate a substantial improvement in experimental success rates compared with other sequence generation approaches.

This highlights the importance of distinguishing technical novelty from practical performance improvements.

The ability to fill arbitrary sequence positions using OADM is a meaningful architectural contribution.

However, whether that flexibility leads to higher experimental success rates, and under which design conditions, requires separate evaluation.

5.2. Generalizability of the Sequence-Only Approach

Reviewer #2 acknowledged that order-agnostic sequence generation can produce useful candidates for several conditional protein design tasks.

However, the reviewer expressed reservations about claims of a general advantage over conventional protein language models and the broader suggestion that sequence information alone is sufficient for a wide range of protein design problems.

These observations do not diminish EvoDiff's contributions.

Instead, they help distinguish demonstrated technical capabilities from broader claims that require additional evidence.

Demonstrating success in specific protein design tasks is not equivalent to demonstrating that sequence-only generation is sufficient for all protein engineering problems.

6. Important Research Questions That Remain Insufficiently Validated

Based on the original EvoDiff work, its 2026 experimental findings, the public reviews, and related research, several important questions remain unresolved.

The priorities below are an analytical assessment based on practical protein engineering and drug discovery considerations, rather than an official ranking from the EvoDiff authors.

Priority Open or insufficiently validated question Research required
High Does modeling natural sequence distributions improve actual functional success rates? Correlate sequence likelihood and novelty with experimentally measured activity
High Under what conditions does EvoDiff outperform alternative generation methods? Compare methods using matched data, model sizes, computational budgets, and experimental budgets
High Can EvoDiff generalize to unknown protein families? Family-held-out and low-homology evaluation
High Can target specificity be controlled? Binding experiments against multiple intended and unintended targets
High Can generated proteins satisfy therapeutic developability requirements? Integrated evaluation of expression, aggregation, thermal stability, and solubility
Medium Can diverse IDR functions be reproduced? Evaluate phase separation, signaling, binding regulation, and other functions beyond localization
Medium How dependent is performance on MSA quality and depth? Systematically evaluate MSA depth, phylogenetic bias, and noise
Medium Can EvoDiff handle long proteins, multidomain architectures, and indels? Experimental evaluation across sequence lengths and domain arrangements
Medium How much does combining sequence and structure information improve performance? Compare sequence-only and multimodal approaches on identical tasks
Medium Does iterative optimization during inference improve results? Evaluate property-guided generation and active learning

6.1. True Extrapolation to Novel Biological Functions

One of the most important research questions concerns functional extrapolation.

Generating sequences similar to known protein families and reproducing established functions is different from engineering entirely new functions.

Future research should examine capabilities such as:

  • Generalization to new protein families.
  • Design of previously unexplored enzymatic reactions.
  • Engineering new protein–protein interactions.

Validation should extend beyond similarity to training sequences and include experimental evidence of functional novelty.

6.2. Design Efficiency Including Experimental Costs

A second important question concerns overall design efficiency.

Generative model studies often emphasize whether at least some successful candidates can be obtained.

However, in therapeutic development, the more practical question is how many useful hits can be obtained from a fixed experimental budget.

For example, useful comparisons could include:

  • The fraction of experimentally successful proteins among 100 tested candidates.
  • The number of functional hits obtained per screening campaign.
  • The cost per experimentally validated useful candidate.

Comparisons between EvoDiff and other generative approaches should ideally use equivalent experimental budgets.

6.3. Separating the Contributions of Sequence, Structure, and Target Information

A third major research question is how much each type of information contributes to protein design performance.

These include:

  • Amino acid sequence information.
  • Three-dimensional structural information.
  • Information about the intended molecular target.

The value of sequence-only modeling cannot be determined simply by asking whether it outperforms multimodal models.

It is also necessary to consider the improvement in experimental success rates relative to the additional structural data, computational cost, and model complexity required.

For certain applications, sequence information alone may provide sufficient functional information.

For other tasks, such as binding to a specified epitope or positioning catalytic residues precisely, structural and interaction information may be particularly important.

Therefore, comparisons should ideally use identical design tasks, evaluation criteria, and experimental budgets.

7. How Should EvoDiff's Limitations Be Interpreted in the History of Protein Generative AI?

Later developments in protein generative AI should not simply be viewed as technologies that replaced EvoDiff.

Instead, the field has evolved along several complementary research directions.

7.1. Sequence-First Protein Generation

Models such as ProGen, EvoDiff, and DPLM focus on learning structural and functional constraints from protein sequence information.[1][5][6]

Major advantages of this direction include:

  • Access to large-scale evolutionary sequence datasets.
  • Direct learning of statistical regularities in protein sequences.
  • Potential applicability to intrinsically disordered proteins and regions.
  • Flexible conditional generation based on specified sequence information.

However, direct control over geometric constraints and desired biological functions remains challenging.

7.2. Structure-First Protein Generation

Models such as RFdiffusion and Chroma emphasize the generation of protein structures.[7][8]

Their strengths include:

  • Structure-conditioned generation.
  • Three-dimensional positioning of functional motifs.
  • Design under explicit geometric constraints.

However, these approaches face different challenges when dealing with dynamic proteins, intrinsically disordered regions, and functions that cannot be explained solely by a single static structure.

7.3. The Development of Multimodal and All-Atom Models

Subsequent work, including ESM3, DPLM-2, La-Proteina, and BoltzGen, has increasingly integrated sequence information with structure, function, atomic geometry, and molecular interactions.[9][10][11][12]

This development can be summarized conceptually as:

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

However, this trend does not prove that sequence-only approaches were fundamentally misguided.

Some protein design problems may be addressed effectively using sequence information alone.

Others may require explicit structural or molecular interaction constraints.

For example, EvoDiff-based IDR inpainting and target-conditioned binder design using a model such as BoltzGen address different objectives under different constraints.

Therefore, comparing these models requires evaluating them on equivalent design problems under comparable conditions.

7.4. EvoDiff Should Not Simply Be Described as the First Protein Sequence Diffusion Model

Another important consideration is historical accuracy.

Related approaches to diffusion-based protein sequence generation existed before 2023.

Therefore, EvoDiff should not be characterized simply as the world's first protein sequence diffusion model.

Its importance lies in combining large-scale evolutionary protein sequence learning, general-purpose discrete diffusion, and flexible conditional sequence generation.[1]

Historically, EvoDiff is better understood as an important contribution to establishing and advancing sequence diffusion as a practical research direction in protein generative AI.

8. Implications for Personalized Medicine and Therapeutic Development

The preceding analysis has several implications for researchers and entrepreneurs considering EvoDiff for personalized medicine and therapeutic development.

The central distinction is between generating protein candidates and designing proteins that meet the functional and practical requirements of therapeutic applications.

8.1. Generating Natural-Like Sequences Does Not Guarantee Therapeutic Efficacy

EvoDiff can generate new protein candidates by learning statistical regularities from evolutionary sequence data.

However, therapeutic development requires experimental evidence that generated proteins perform their intended biological functions.

Even demonstrating target binding is not sufficient to establish therapeutic efficacy.

The p53–MDM2 experiment illustrates this distinction.

Some EvoDiff-generated candidates exhibited strong binding, but activation of the p53 pathway in cells and antitumor efficacy were not established.[2]

The same principle applies to personalized medicine.

The ability to generate a candidate protein for a particular molecular context should not be confused with evidence of clinical benefit for an individual patient.

8.2. Target Specificity and Developability Must Be Evaluated

Standard sequence-first generation in EvoDiff does not directly guarantee target specificity or therapeutic developability.

Therefore, applications in drug development require evaluation of properties such as:

  • Activity or binding affinity against the intended target.
  • Specificity against unintended targets.
  • Protein expression.
  • Solubility.
  • Aggregation propensity.
  • Thermal stability.
  • Safety and immunogenicity, where relevant.

These properties are distinct from sequence likelihood or confidence in predicted protein structures.

8.3. The Entire Design System Matters More Than the Generative Model Alone

The 2026 EvoDiff experiments involved computational evaluation and candidate selection before experimental testing.[2]

This demonstrates that successful protein engineering does not necessarily depend on a generative model acting alone.

For personalized medicine and therapeutic development, a useful conceptual workflow is:

Generate → Evaluate → Select → Experiment → Improve

Such a system integrates:

  1. Candidate generation.
  2. Computational evaluation.
  3. Candidate selection.
  4. Experimental validation.
  5. Iterative improvement.

The effectiveness of this workflow depends not only on EvoDiff's generative capabilities but also on the quality of evaluation models, the efficiency of experimental screening, and the ability to use experimental results to guide future designs.

8.4. For Entrepreneurs, Design Efficiency May Matter More Than Generation Capability

When considering the commercialization of protein generative AI, technical novelty should be distinguished from practical research and development value.

EvoDiff provides a powerful approach to flexible protein sequence generation.

However, that capability alone does not establish a commercial advantage over existing design methods.

A more important question is how efficiently the system produces experimentally useful candidates.

For example:

  • Does EvoDiff produce more functional hits than alternative methods when the same number of candidates is tested?
  • Under what design conditions does its flexible conditional generation provide a meaningful advantage?
  • Does combining EvoDiff with structure-based models improve experimental success rates?
  • Can experimental results be incorporated into subsequent rounds of candidate generation?

These questions are not fully answered by the existing EvoDiff research.

For entrepreneurs developing protein design platforms, it is therefore important to define the intended therapeutic targets, required protein functions, and available experimental evaluation systems before determining whether EvoDiff provides a practical advantage.

9. Conclusion: EvoDiff's Greatest Contribution and Its Fundamental Limitations

One of the most important scientific questions raised by EvoDiff is:

To what extent can learning the distribution of naturally evolved protein sequences generalize to the design of proteins with artificially specified functions?

EvoDiff provides partially affirmative evidence.

In particular, its results involving IDR function and motif scaffolding without explicit structural inputs demonstrate the potential of sequence-first protein design.[2]

However, it remains unclear whether the information contained in natural evolutionary sequences is sufficient for arbitrary novel drug discovery tasks.

It also remains uncertain whether diffusion-based generation can use that information more efficiently than alternative methods across a broad range of protein design problems.

EvoDiff's greatest contribution is demonstrating that protein sequences can be treated not merely as outputs of structural design but as an independent generative space containing evolutionary information.

Within that space, flexible conditional generation becomes possible.

Its fundamental limitation is that learning the natural distribution of protein sequences does not directly guarantee optimization of desired biological functions, molecular interactions, or therapeutic developability.

Understanding this distinction changes how we interpret subsequent developments such as DPLM, ESM3, and BoltzGen.

These models are not simply making protein generation more complex.

Rather, they are addressing the gap between:

Being able to generate proteins and being able to design proteins that reliably satisfy specific objectives.

This gap is particularly important for personalized medicine and therapeutic development.

Generating a novel protein sequence is only the beginning.

The resulting protein must be experimentally shown to perform the intended function, exhibit appropriate target specificity, and satisfy the practical requirements of development.

EvoDiff did not prove that sequence information alone is sufficient to solve every protein design problem.

Instead, it demonstrated how evolutionary sequence information can be exploited for generative protein design and helped establish a research direction for investigating both the potential and limitations of sequence-first approaches.

This is EvoDiff's lasting scientific significance and a crucial perspective for evaluating its role in future personalized medicine and therapeutic development.

References

[1] 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

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

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

[4] “EvoFlows,” Related research on evolutionary sequence generation and insertion control in EvoDiff-MSA.

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

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

[7] Watson et al., “De novo design of protein structure and function with RFdiffusion,” Nature, 620, 1089–1100, 2023. https://www.nature.com/articles/s41586-023-06415-8

[8] 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

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

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

[12] “BoltzGen,” bioRxiv, revised preprint, 2026. https://www.biorxiv.org/content/10.1101/2025.11.20.689494v2.full

[13] Microsoft Research, “EvoDiff Model Card,” GitHub. https://github.com/microsoft/evodiff/blob/main/EvoDiff_modelcard.md

[14] “RAG-ESM,” PRX Life, 2025. https://journals.aps.org/prxlife/pdf/10.1103/db1b-hy16