Protein generation with evolutionary diffusion: sequence is all you need
Curation statements for this article:-
Curated by eLife
eLife Assessment
This important study introduces EvoDiff, an order-agnostic diffusion foundation model for protein sequence generation, and demonstrates its potential across unconditional generation, motif scaffolding, MSA-conditioned design, and IDR inpainting. The evidence supporting the principal claims is solid, combining computational evaluation with experimental validation, although the comparative advantages and generality of some applications would benefit from further clarification and calibration.
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (eLife)
Abstract
Deep generative models are increasingly powerful tools for the in silico design of novel proteins. Recently, a family of generative models called diffusion models has demonstrated the ability to generate biologically plausible proteins that are dissimilar to any actual proteins seen in nature, enabling unprecedented capability and control in de novo protein design. However, current state-of-the-art diffusion models generate protein structures, which limits the scope of their training data and restricts generations to a small and biased subset of protein design space. Here, we introduce a general-purpose diffusion framework, EvoDiff, that combines evolutionary-scale data with the distinct conditioning capabilities of diffusion models for controllable protein generation in sequence space. EvoDiff generates high-fidelity, diverse, and structurally-plausible proteins that cover natural sequence and functional space. We show experimentally that EvoDiff generations express, fold, and exhibit expected secondary structure elements. Critically, EvoDiff can generate proteins inaccessible to structure-based models, such as those with disordered regions, while maintaining the ability to design scaffolds for functional structural motifs. We validate the universality of our sequence-based formulation by experimentally characterizing intrinsically-disordered mitochondrial targeting signals, metal-binding proteins, and protein binders designed using EvoDiff. We envision that EvoDiff will expand capabilities in protein engineering beyond the structure-function paradigm toward programmable, sequence-first design.
Article activity feed
-
eLife Assessment
This important study introduces EvoDiff, an order-agnostic diffusion foundation model for protein sequence generation, and demonstrates its potential across unconditional generation, motif scaffolding, MSA-conditioned design, and IDR inpainting. The evidence supporting the principal claims is solid, combining computational evaluation with experimental validation, although the comparative advantages and generality of some applications would benefit from further clarification and calibration.
-
Reviewer #1 (Public review):
Summary:
This manuscript presents a new foundation model, EvoDiff, for designing primary protein sequences. By leveraging an evolutionary-scale dataset, the model can be applied to evolution-guided sequence generation, sequence inpainting, and functional scaffolding.
Strengths:
The model provides an efficient approach for designing protein sequences and could be useful for developing protein therapeutics, engineering enzymes, designing biomaterials, and many other applications. The manuscript presents solid results showing that proteins designed by EvoDiff can achieve the same biological functions as their wild-type counterparts.
Weaknesses:
Compared with other sequence-generation models, EvoDiff does not substantially improve the success rate, suggesting that significant experimental effort is still …
Reviewer #1 (Public review):
Summary:
This manuscript presents a new foundation model, EvoDiff, for designing primary protein sequences. By leveraging an evolutionary-scale dataset, the model can be applied to evolution-guided sequence generation, sequence inpainting, and functional scaffolding.
Strengths:
The model provides an efficient approach for designing protein sequences and could be useful for developing protein therapeutics, engineering enzymes, designing biomaterials, and many other applications. The manuscript presents solid results showing that proteins designed by EvoDiff can achieve the same biological functions as their wild-type counterparts.
Weaknesses:
Compared with other sequence-generation models, EvoDiff does not substantially improve the success rate, suggesting that significant experimental effort is still required to screen and identify successful hits.
-
Reviewer #2 (Public review):
In this work, Alamdari et al. present EvoDiff, which provides the capability to generate protein sequences directly in sequence space, using a discrete diffusion model. There are several versions. EvoDiff-seq is trained on UniRef50 sequences (~42 million), and EvoDiff-MSDA operates instead by using sequence alignment methods to generate new members of protein families. The authors demonstrate many modes of sequence generation, including unconditional and conditional, inpainting of disordered regions, and also generating scaffolding of functional motifs. Their evaluation is also multifaceted, covering foldability, folding self-consistency, language embeddings, secondary structure distributions, and experiments for a set of different scenarios.
The paper has many notable strengths. It is comprehensive in …
Reviewer #2 (Public review):
In this work, Alamdari et al. present EvoDiff, which provides the capability to generate protein sequences directly in sequence space, using a discrete diffusion model. There are several versions. EvoDiff-seq is trained on UniRef50 sequences (~42 million), and EvoDiff-MSDA operates instead by using sequence alignment methods to generate new members of protein families. The authors demonstrate many modes of sequence generation, including unconditional and conditional, inpainting of disordered regions, and also generating scaffolding of functional motifs. Their evaluation is also multifaceted, covering foldability, folding self-consistency, language embeddings, secondary structure distributions, and experiments for a set of different scenarios.
The paper has many notable strengths. It is comprehensive in breadth, and the experimental component is distinctive, although I am not personally suited to review the rigor of that element.
I would suggest that the paper's results certainly support the conclusion that order-agnostic sequence generation can yield useful candidates for multiple conditional design tasks. I am not totally convinced that it necessarily establishes diffusion as a generally superior approach to other competitors, like the conventional protein language models- EvoDiff is certainly competitive, and I think that the demonstration of diffusion is nice. I also am not sure that it is fair to say that sequence alone is sufficient for the broad design capabilities claimed (other than the "in principle" statement).
I am overall quite supportive of the work and its demonstration, but I have a few comments for consideration in any revision.
(1) I did not work through all dates of everything, but it appears to me that there are several recent conceptual and methodological competitors. These include DPLM and ProtBFN - both of these seem to be after the first preprint of EvoDiff, but given the time gap, there probably deserves to be some additional discussion or comparison. I would say, ideally, they should offer direct benchmarking. If the authors are disinclined, then I would think they should just temper their claims of contemporary SOTA performance or general superiority. Instead, the paper would still remain valuable as an early and experimentally demonstrated sequence diffusion framework. I don't think it needs to be more than that.
(2) Related to the above, the manuscript should more carefully distinguish the demonstrated advantage of order-agnostic generation from the quality of unconditional generation. Regarding Figure 3, the authors argue that Evodiff's diffusion objective is necessary, but this does not seem to account for or address the LRAR baselines in Tables S1 and S3. Unless I am misunderstanding, the 640M LRAR model exhibits several better scores. The authors later suggest that EvoDiff's principal advantage is conditioning on arbitrary positions, which is valid, but that's a little different that what is being claimed. I suggest that the LRAR results should be shown or discussed alongside Figure 3, and then the authors revise to say that they have flexible conditional generation as the principal empirical benefit.
(3) I really like the IDR experiment, but I'm not sure it demonstrates that Evodiff can design functional IDRS generally. Cox15 is a favorable target because its mature sequence strongly identifies a conserved mitochondrial protein, and EvoDiff-MSA is supplied directly with its orthologous family. The eight tested sequences were also selected from hundreds of candidates using both DR-BERT and MitoFates, making the experiment a test of the full generation-and-prediction pipeline rather than of EvoDiff alone. Moreover, mitochondrial targeting is tested, but the disordered character of the generated sequences is not experimentally established. In any case, I think it would certainly be more convincing if there were other examples, with unrelated proteins or IDR functions. I would appreciate that this is again a step beyond what the authors might be compelled to do, but their claim could be simply more calibrated.
(4) The authors might benefit from explaining the advantages or complementarity of EvoDiff to other property-directed approaches for exploring sequence space. This has been done, for example, by using Bayesian optimization and genetic algorithms to tune properties of IDP condensates (DOI: 10.1021/acs.jpcb.8b03822). In my understanding, these are addressing a different problem from EvoDiff by optimizing sequences explicitly towards physical targets, while EvoDiff is a generative framework that can be used for sampling/inpainting/ etc. Is it clear how these strategies might be plausibly integrated? If so, that would be a relevant point of discussion and a potential advantage for EvoDiff.
-
