Design and experimental characterization of specificity-switching mutational paths of WW domains

  1. Ahmed Rehan
  2. Eugenio Mauri
  3. Jorge Fernandez-de-Cossio-Diaz
  4. Pierre-Guillaume Brun
  5. Remi Monasson
  6. Marco Ribezzi-Crivellari
  7. Simona Cocco

  • Curated by eLife

    eLife logo

    eLife Assessment

    In this important study, the authors demonstrate that generative AI techniques (restricted Boltzmann machine) can be used effectively to design and characterize mutational pathways of WW domains with different binding specificities. The computational studies are complemented by experimental validations, and the results provide convincing evidence supporting the idea that sequence landscape holds significance in understanding protein evolution from a transition path perspective. The minor weakness of the study has been largely addressed in the revision. The work will likely have a major impact on research aimed at understanding how evolution navigates fitness landscapes as well as reconstructing ancestral sequences.

Reviewed by

Read the full article

Listed in

Log in to save this article

Abstract

Specific interactions between proteins and other biomolecules are ubiquitous in cellular processes. How specificity is encoded in the protein sequence and can be modified through a minimal set of concerted mutations is a complex issue. In this work, we focus on the WW protein domain, whose variants specifically bind to different classes of proline-rich peptides. Combining unsupervised learning of homologous WW sequence data with Restricted Boltzmann Machines (RBM) and path-sampling methods, we design mutational paths of putative WW domains interpolating between two natural WW domains with either distinct or similar specificities. Sequences along the designed paths are then experimentally validated with high-throughput in-vitro binding assays against 3 peptides of different classes. The vast majority (93%) of intermediate sequences along the designed paths are responsive to the initial or/and final peptides. On the contrary, domains along scrambled paths, in which the same mutations are introduced in random order are not functional, emphasizing how successful design crucially depends on the ability to model epistatic interactions. Switch in specificity between classes I and IV, whose representative peptides bind to different pockets on the WW domain takes place through intermediates displaying some level of binding cross-reactivity with the tested peptides, contrary to the transition from class I to II, which are associated with the same binding pocket. Lastly, we show that the RBM paths share a high identity with internal nodes obtained from ancestral sequence reconstruction based on the seed WW domains.

Article activity feed

  1. eLife

    eLife Assessment

    In this important study, the authors demonstrate that generative AI techniques (restricted Boltzmann machine) can be used effectively to design and characterize mutational pathways of WW domains with different binding specificities. The computational studies are complemented by experimental validations, and the results provide convincing evidence supporting the idea that sequence landscape holds significance in understanding protein evolution from a transition path perspective. The minor weakness of the study has been largely addressed in the revision. The work will likely have a major impact on research aimed at understanding how evolution navigates fitness landscapes as well as reconstructing ancestral sequences.

  2. eLife

    Reviewer #1 (Public review):

    Summary:

    The authors aim to study mutational paths connecting WW domains with different binding specificities. Their approach combines an unsupervised sequence generative model based on RBMs with a path-sampling algorithm. The key result is that most intermediate sequences along the designed transition paths retain measurable binding activity in wet-lab assays, whereas paths containing the same mutations introduced in a randomized order are largely non-functional. This difference is attributed to epistatic interactions captured by the RBM model.

    Strengths:

    Exploring mutational paths in high-dimensional protein sequence space is a challenging problem. The computational framework used here is state-of-the-art and is strengthened by systematic experimental characterization of binding activity. The study is comprehensive in scope, including multiple transition paths both within and across WW specificity classes, and the integration of modeling with high-throughput experimental validation is a clear strength.

  3. eLife

    Reviewer #2 (Public review):

    The manuscript is extremely clear and well written, the experimental evidence is strong and the methods are clearly described, so I do not have major issues to raise. A few minor issues are listed below. Some of these were presented in my first review but were not addressed by the authors.

    (2) In Section 3.3 the authors say that direct paths connecting Class I and Class IV behave similarly to indirect paths, despite having lower scores according to the RBM. How generic is this? Does it also happen for other classes? This might be an important point to address as direct paths are easier to sample.

    (3) The path shown in Fig.4 goes through a region of non-functionality around sequences 18-19. It seems that the sample path is basically exploring the functional regions for Class I and Class II/III separately trying to approach the other class, but then it can't really do the switch.

    By contrast the path going from Class I to Class IV seems able to perform the functional switch in a single step (20-21) without losing too much of the function.

    Perhaps the authors could better comment on this? Is this a limitation of the sampling method, or a fundamental biological fact?

    (4) On page 12 it is stated that temperature was chosen to 1/3 to maximize the score. This is important and should be mentioned earlier (I didn't notice it until that point).

    (5) On page 13 it is stated that "However, the scores of the ancestral sequences along the phylogenetic pathways assigned by the RBM are significantly lower than the ones of the RBM-designed sequences. This result is expected as ASR reconstruction does not take into account epistasis, differently from RBM, and we expect ASR sequences to generally be of lesser quality."

    I was very surprised by this result. My own experience with ASR shows that on the contrary, sequences found by ASR (via maximum likelihood) tend to have high scores in the (R)BM and tend to be more stable than extant sequences. I attribute this to the fact that ASR typically finds a "consensus" sequence that maximizes the contribution to the score coming from the fields (the profile), which is typically dominant over the epistatic signal, resulting in a bigger score. Maybe the authors did not use maximum likelihood in the ASR? Some clarification might be useful here.

  4. eLife

    Author response:

    The following is the authors’ response to the original reviews.

    Public Reviews:

    Reviewer #1 (Public review):

    Summary:

    The authors aim to study mutational paths connecting WW domains with different binding specificities. Their approach combines an unsupervised sequence generative model based on RBMs with a path-sampling algorithm. The key result is that most intermediate sequences along the designed transition paths retain measurable binding activity in wet-lab assays, whereas paths containing the same mutations introduced in a randomized order are largely non-functional. This difference is attributed to epistatic interactions captured by the RBM model.

    Strengths:

    Exploring mutational paths in high-dimensional protein sequence space is a challenging problem. The computational framework used here is state-of-the-art and is strengthened by systematic experimental characterization of binding activity. The study is comprehensive in scope, including multiple transition paths both within and across WW specificity classes, and the integration of modeling with high-throughput experimental validation is a clear strength.

    Weaknesses:

    A major concern is whether the stated goal of specificity switching is fully achieved. Along the sampled transition paths, most intermediate variants appear to retain specificity close to either the initial or the final class, rather than exhibiting gradually shifting specificity. For example, in Figure 4G (Class I to Class II/III), binding appears largely binary, with intermediates behaving similarly to one of the endpoints. A similar pattern is observed in Figure 3H for the Class I to Class IV transition, where binding responses are close to 0 or 1. In this sense, the specificity-switching objective is only partially realized by assigning two endpoints with different specificity. This raises a broader conceptual question: is it possible that different WW specificities evolved from a common ancestor without passing through intermediates that exhibit mixed or intermediate specificity? If so, then inferring specificity-switching pathways purely from extant natural sequences may be fundamentally challenging.

    We have added a sentence in the introduction to underline these points, in particular: Is the specificity switch happening through abrupt or more gradual specificity changes? Does it correspond to a bottleneck in sequence space?

    Reviewer #2 (Public review):

    This is an extremely important work that shows how one can use generative models to construct specificity-switching mutational paths in complex fitness landscapes. The experimental evidence is very clear, and the theoretical tools are innovative.

    The work will likely have a deep impact on future research aimed at understanding how evolution navigates fitness landscapes as well as reconstructing ancestral sequences.

    The manuscript is extremely clear and well written, the experimental evidence is strong, and the methods are clearly described, so I do not have major issues to raise. A few minor issues are listed below.

    (1) I consider the WW domain as an 'easy' case from the point of view of generative modelling. The domain is rather short, epistatic effects are not very strong (e.g. Boltzmann learning usually converges very quickly to a very paramagnetic state), and the resulting models are well interpretable (e.g. the hidden units of the RBM correlate well with subclasses).

    This is not always (not often?) the case, however. In more complex proteins, the learning procedures can be slower and the resulting models less interpretable. Just for completeness, perhaps the authors could comment on the generality of the results and what they would expect for other systems based on their experience.

    We have added a sentence in conclusion to discuss the application of our framework to more complex and longer proteins.

    (2) In Section 3.3, the authors say that direct paths connecting Class I and Class IV behave similarly to indirect paths, despite having lower scores according to the RBM. How generic is this? Does it also happen for other classes? This might be an important point to address, as direct paths are easier to sample.

    We have added a sentence in conclusion to emphasize the need for a more exhaustive study of non-direct vs direct paths.

    (3) The path shown in Figure 4 goes through a region of non-functionality around sequences 18-19. It seems that the sample path is basically exploring the functional regions for Class I and Class II/III separately, trying to approach the other class, but then it can't really make the switch.

    By contrast, the path going from Class I to Class IV seems able to perform the functional switch in a single step (20-21) without losing too much of the function.

    Perhaps the authors could better comment on this? Is this a limitation of the sampling method, or a fundamental biological fact?

    We now better stress and discuss the comparison between Class I to Class IV paths and Class I to Class II paths in the conclusion.

    (4) On page 12, it is stated that the temperature was chosen to 1/3 to maximize the score. This is important and should be mentioned earlier (I didn't notice it until that point).

    We have added a sentence in Section 3.2 on the lowering of the sampling temperature to 1/3 to obtain high-scores sequences.

    (5) On page 13, it is stated that: "However, the scores of the ancestral sequences along the phylogenetic pathways assigned by the RBM are significantly lower than the ones of the RBM-designed sequences. This result is expected as ASR reconstruction does not take into account epistasis, differently from RBM, and we expect ASR sequences to generally be of lesser quality."

    I was very surprised by this result. My own experience with ASR shows that, on the contrary, sequences found by ASR (via maximum likelihood) tend to have high scores in the (R)BM, and tend to be more stable than extant sequences. I attribute this to the fact that ASR typically finds a "consensus" sequence that maximizes the contribution to the score coming from the fields (the profile), which is typically dominant over the epistatic signal, resulting in a bigger score. Maybe the authors did not use maximum likelihood in the ASR? Some clarification might be useful here.

    We have updated Figure 5 with shuffled and ASR sequences, and added Suppl Figure 7 to compare the score of the consensus sequences and the ones of the most ancestral nodes along the path in the ASR reconstruction with the natural scores.

    Recommendations for the authors:

    Reviewer #1 (Recommendations for the authors):

    I suggest that the authors either more clearly acknowledge that smooth specificity switching is not fully demonstrated or reconsider the use of "specificity switching" in the title. An alternative framing could emphasize the study of mutational paths connecting different specificity classes, without implying a transition in specificity.

    One of the main questions motivating our study was precisely to investigate where along the path and how the transition in specificity happens, i.e. through intermediate that exhibit mixed or intermediate specificity (which we had defined as ‘smooth’) or more abruptly.

    For class I-IV paths, we observe a smooth specificity switch with some cross-specificity of the sequences designed in the switching-specificity region, with low but significant binding responses not only towards class I and IV ligands (the extant sequences specificity class) but also to ligand of class II. On the contrary, in the paths from class I to class II-III, we observe the abrupt switch, at the predicted specificity boundary, with sequences with a clear response to one or the other peptide. This abrupt switch requires the change of several amino acids determining the specificity class, and binding is loss when this change is incomplete.

    We understand that the term “smooth” to design the cross-reactivity of the few sequences in the predicted ‘switching boundary zone’ could be misleading, so we have removed it from the abstract.

    Moreover, we have added a sentence in introduction to motivate the study of where and how the specificity transition takes place, and two sentences in conclusion about the identification of where the specificity switching take place along the path, and to underline the difference between the paths from class I to class IV and the ones from class I to class II/III.

    It is striking that all scrambled paths are non-functional. Given that sequences adjacent to the end states differ by only a few mutations, one might expect at least some of these variants to remain functional, as small numbers of mutations often do not abolish activity. The authors should discuss this result in greater depth and more clearly explain how it uniquely supports the role of epistatic interactions.

    The sampling of the scrambled paths is not strictly random. Along the scrambled path the mutations are inserted in the reverse order; the scrambled paths are therefore chosen to maximize the sequence divergence between the unscrambled sequences (in which the mutation is predicted to be functional by the path sampling algorithm) and the one in which it is tested in the scrambled path. The fact that the background matters is a clear sign of epistasis.

    We have better explained this important point in Section 3.3, where we introduce scrambled paths.

    Specificity class annotations are inferred using RBM latent weights. However, as shown in Figure 2E, the boundaries between classes in latent space are not sharply defined. It would strengthen the analysis to validate these class assignments using alternative representations, such as clustering in the latent space of a VAE or embeddings from protein language models, to assess whether the inferred class boundaries are robust and consistent.

    Beyond RBM weights, the boundaries between clusters have been assessed by inferring two different RBMs (which we call class specific RBM) on class specific MSA, such models validated the robustness of the boundaries. Most importantly, we have shown that the classification is consistent with experiments.

    To further answer this comment, we have added Suppl. Figures 7 and 8, showing that a simple PCA approach, as well as VAE agree with our clustering.

    In Figure 3C and Figure 3F, variants 18 and 19 appear to represent transition states located in regions of latent space with little or no training data. Intuitively, such variants would be expected to have a lower probability than variants 22, which lie in a high-density region of the RBM model. However, this expectation is not reflected in Figure 3C. A similar issue appears in Figure 4. The authors should carefully re-examine these results and clarify whether this behavior arises from the model, the scoring procedure, or a potential error.

    We confirm that variants 18 and 19 in mutational path between classes I-IV have larger scores than the class IV wild-type domain. We have attributed the low score of this class IV wild-type (the extant sequence) to its smaller structural stability, as discussed in Sec 3.3 and with Supplementary Figure 3.

    The key point here is that the scores of Natural sequences are spread, as shown with the histogram in Fig. 5 (green distribution), corresponding to more or less stable natural protein. The RBM model (with a low sampling temperature favoring higher RBM scores) is able to generate artificial sequences far away in sequence space (i.e. with high Hamming distance with respect to natural one), while keeping high scores. The functionality of designed proteins with high RBM scores has been demonstrated in previous work by some of us, e.g. [Malbranke et al PLoS Comp Bio 2026], [Huot, Monasson, Cocco Cell system 2026], [Fernandez-de-Cossio-Diaz Nature Communication 2026], building up on similar results for BM models [Russ et al Science 2020].

    We have now clarified the point of the sampling at low temperature earlier in the manuscript in Section 3.2.

    Reviewer #2 (Recommendations for the authors):

    (1) In Figure 2 F-G, maybe a histogram of the scores of natural sequences could be added for comparison.

    We have added the histogram of scores of natural sequences in figure 5, where we discuss the scores of the generated sequences for the different paths.

    (2) In Section 3.4, "(Wild type 1YWI or formin binding protein)" is not clear; there might be a typo. Please check

    We do refer to Formin-binding protein 1, also known as FNBP1. We have reformulated the corresponding sentence.

    (3) On page 16 at the end of section 4, the authors could consider citing "Integrating experimental feedback improves generative models for biological sequences" by Calvanese et al. among the active learning methods.

    Thanks for suggesting to cite this very relevant work. We have added the reference.

    (4) Reference 78 is stated "in preparation (2019)", I guess this is a mistake and could be fixed.

    The reference has been updated.

  5. Version published to 10.7554/elife.110491.2 on eLife
  6. Version published to 10.7554/elife.110491 on eLife
  7. Version published to 10.64898/2025.12.08.693000v2 on bioRxiv
  8. eLife

    Author response:

    Public Reviews:

    Reviewer #1:

    Summary:

    The authors aim to study mutational paths connecting WW domains with different binding specificities. Their approach combines an unsupervised sequence generative model based on RBMs with a path-sampling algorithm. The key result is that most intermediate sequences along the designed transition paths retain measurable binding activity in wet-lab assays, whereas paths containing the same mutations introduced in a randomized order are largely nonfunctional. This difference is attributed to epistatic interactions captured by the RBM model.

    Strengths:

    Exploring mutational paths in high-dimensional protein sequence space is a challenging problem. The computational framework used here is state-of-the-art and is strengthened by systematic experimental characterization of binding activity. The study is comprehensive in scope, including multiple transition paths both within and across WW specificity classes, and the integration of modeling with high-throughput experimental validation is a clear strength.

    Weaknesses:

    A major concern is whether the stated goal of specificity switching is fully achieved. Along the sampled transition paths, most intermediate variants appear to retain specificity close to either the initial or the final class, rather than exhibiting gradually shifting specificity. For example, in Figure 4G (Class I to Class II/III), binding appears largely binary, with intermediates behaving similarly to one of the endpoints. A similar pattern is observed in Figure 3H for the Class I to Class IV transition, where binding responses are close to 0 or 1. In this sense, the specificityswitching objective is only partially realized by assigning two endpoints with different specificity. This raises a broader conceptual question: is it possible that different WW specificities evolved from a common ancestor without passing through intermediates that exhibit mixed or intermediate specificity? If so, then inferring specificity-switching pathways purely from extant natural sequences may be fundamentally challenging.

    This is a key question, which was one of the original motivations of our work. Both hypothesis of ‘abrupt switches’ (punctuated equilibria, corresponding to distinct specificities) and more gradual changes (smooth transition, through intermediate that exhibit mixed or intermediate specificity) are possible.

    Many natural specificity-switching events have probably resulted from the need to adapt to environmental change and selection for a different specificity, which can be compatible with an abrupt change in specificity. Others may reflect the gradual evolution of promiscuous ancestral sequences to more specialized ones, loosing cross-reactivity. A molecular mechanism that could allow abrupt switching is gene duplication, a frequent mechanism for WW domain diversification, beyond standard mutational-driven evolution processes.

    As for the specificity-switching paths for WW domains found in this work, the presence of weakly responsive cross-reactive intermediates along the designed paths for I<->IV, and their absence in the I<->II path, suggests that designing promiscuous domains is hard (see also related response to point 3 of Reviewer 2) and generally not selected by natural evolution (as seen from the clear clustering of extant proteins in different specificity classes).

    For a small domain such as WW, mutations that favor some specificity classes are known to have detrimental effects on fundamental properties, such as folding kinetics and stability, see Ref [72]. It is possible that larger, less constrained protein domains could allow for more crossreactive variants and smoother specifity switching. However, experiments on fluorescent proteins looking for interpolation between two wave-lengths have shown that the switch was abrupt [Poelwijk et al. Nature Communications (2019)].

    Our scope was to achieve a functional switch (imposed by the two extant end-points) through a path of designed, functional intermediates and to correctly predict, with our RBM model, the location of the specificity transition and of the cross-reactivity region (which we expected only along the I-IV path). This scope was successfully reached as demonstrated by experiments.

    Reviewer #2:

    This is an extremely important work that shows how one can use generative models to construct specificity-switching mutational paths in complex fitness landscapes. The experimental evidence is very clear, and the theoretical tools are innovative.

    The work will likely have a deep impact on future research aimed at understanding how evolution navigates fitness landscapes as well as reconstructing ancestral sequences.

    The manuscript is extremely clear and well written, the experimental evidence is strong, and the methods are clearly described, so I do not have major issues to raise. A few minor issues are listed below.

    (1) I consider the WW domain as an 'easy' case from the point of view of generative modelling. The domain is rather short, epistatic effects are not very strong (e.g. Boltzmann learning usually converges very quickly to a very paramagnetic state), and the resulting models are well interpretable (e.g. the hidden units of the RBM correlate well with subclasses).

    This is not always (not often?) the case, however. In more complex proteins, the learning procedures can be slower and the resulting models less interpretable. Just for completeness, perhaps the authors could comment on the generality of the results and what they would expect for other systems based on their experience.

    We agree with Reviewer 2 that WW sequences are short and simple to handle from a computational point of view, and was chosen for this reason to test the design of full mutational paths (after having benchmarked it to lattice-protein models, see Refs. [30] and [44]). Our work gives additional support to the effectiveness of generative models learned from sequence data. This said, from a biological point of view, WW is a highly constrained domain, see comment by Reviewer 1 above and our answer.

    In longer and more complex proteins, we expect it will be more difficult to disentangle specificityswitching latent units, see Fernandez-de-Cossio-Diaz et al., Physical Review X 2023 for a discussion and a possible computational approach to this issue. Notice that, while relating the latent units to specificity classes was convenient, it was not used to generate the paths themselves. Therefore, we believe that our method is quite robust and easily generalizable to applications to more complex and longer proteins. As an illustration, we have recently used it to sample viral trajectories (more precisely, variants of the Receptor Binding Domain of the SARSCoV-2 spike protein) capable of escaping antibody recognition, see Huot et al., PNAS 2026. In this recent work, we projected the paths onto the principal antigenic space, defined by the top two Principal Components of the viral variant binding affinities to 32 antibodies. In this representation, sampled paths displayed trends similar to natural paths, drawn from the sequences sampled during the pandemics. This finding supports the applicability and interpretation of our method for more complex proteins.

    (2) In Section 3.3, the authors say that direct paths connecting Class I and Class IV behave similarly to indirect paths, despite having lower scores according to the RBM. How generic is this? Does it also happen for other classes? This might be an important point to address, as direct paths are easier to sample.

    We think that this finding, true for paths connecting classes I and IV, is not general. In a previous paper we have benchmarked our path-designing approach on simple models of insilico lattice proteins and shown that indirect path led to gains in the overall fitness (computed according with the ground-truth model) [Mauri, Cocco, Monasson, Physical Review E 2023, fig. 9-12].

    In general, we would expect that indirect paths could explore alternative mutations, important to compensate for transitory destabilizing mutations that could occur along the path. We speculate that these stabilizing mutations happen for non-direct paths at its extremity near class-I wildtype. A slightly decrease in binding response to peptide C1 for direct path is nevertheless observed (see Suppl Table 4), but our experimental detection, focused on binding response, is not tailored to directly detect a difference in stability. When approaching the class-IV anchoring point, we observe that paths interpolating between classes I and IV are very constrained and show limited diversity, going through a funnel in sequence space corresponding to the direct path. We agree with Reviewer 2 that a more exhaustive comparison with direct paths would be interesting, and will add a sentence in conclusion.

    (3) The path shown in Figure 4 goes through a region of non-functionality around sequences 1819. It seems that the sample path is basically exploring the functional regions for Class I and Class II/III separately, trying to approach the other class, but then it can't really make the switch.

    By contrast, the path going from Class I to Class IV seems able to perform the functional switch in a single step (20-21) without losing too much of the function.

    Perhaps the authors could better comment on this? Is this a limitation of the sampling method, or a fundamental biological fact?

    Class I to Class IV paths and Class I to Class II paths fundamentally differ because the binding pocket in Class I WW domains is different from the one of Class IV WWs, while Classes I and II/III share the same binding region. This important difference may explain why class I specificity can switch to class IV specificity (steps 20-21), without completely loosing affinity to the peptide of class I. To investigate if the two binding regions are really independent or not, we have tested some additional specific mutations along the I-IV mutational paths. In our attempts to engineer cross-reactivity, we have observed that it is important to substantially lower affinity to class I peptide to acquire class IV specificity, in agreement with previous studies [72]. Moreover, the I to IV path seems to go through a funnel-like part in the region with no natural sequences, with the same transition intermediates obtained in several designed paths. This indicates that the Class I to Class IV functional switch is more constrained than the Class I to II switch. Let us also emphasize that our assessment of class specificity is based on one peptide for each class. It would be interesting to test multiple WW-binding peptides with similar biochemical properties to acquire a more complete view of the specificities.

    (4) On page 12, it is stated that the temperature was chosen to 1/3 to maximize the score. This is important and should be mentioned earlier (I didn't notice it until that point).

    Section 3.5 explains that RBM samples can be biased, by lowering the sampling temperature to 1/3 to obtain high-scores sequences, which are more likely to be functional as proven in [Russ et al., Science 2020]. We acknowledge (as also noted by Reviewer 1) that this section comes at the end of the manuscript, while differences in scores along the path are shown before, so the discussion of this important point is somewhat delayed. We will add a sentence earlier in Results to explain this point.

    (5) On page 13, it is stated that: "However, the scores of the ancestral sequences along the phylogenetic pathways assigned by the RBM are significantly lower than the ones of the RBMdesigned sequences. This result is expected as ASR reconstruction does not take into account epistasis, differently from RBM, and we expect ASR sequences to generally be of lesser quality."

    I was very surprised by this result. My own experience with ASR shows that, on the contrary, sequences found by ASR (via maximum likelihood) tend to have high scores in the (R)BM, and tend to be more stable than extant sequences. I attribute this to the fact that ASR typically finds a "consensus" sequence that maximizes the contribution to the score coming from the fields (the profile), which is typically dominant over the epistatic signal, resulting in a bigger score. Maybe the authors did not use maximum likelihood in the ASR? Some clarification might be useful here.

    We agree with Reviewer 2 that the consensus sequence is an atypical sequence for an independent model with a large RBM score. We will update Figure 5 of the manuscript to show that this is also happening in our case.

    We use Maximum Likelihood in ASR but our ASR path corresponds to all internal nodes of the reconstructed tree joining the two extant sequences, not only to the most ancestral node. Overall, the ancestral sequences along the ASR paths are different from the consensus sequence (mean identity of 76% and 60% respectively). The most ancestral nodes in the paths are also different from the consensus having 81% (paths between type I and IV domains) or 54%(paths between type I and II/III domains) similarity, and an RBM score of -21, or -58, respectively. We agree that some ASR internal-node sequence have a higher score than the natural wild-types (extant sequences). This is shown in Fig. 6: several points have larger RBM score than the two anchoring points at the extremities of the path, possibly due to the fact that natural sequences are not always the most stable ones. As discussed in conclusion, ASR nodes have moreover generally better scores than the sequences obtained by sampling an independent model. Phylogenetic reconstruction implicitly takes into account some degree of co-variation between sites in natural sequences, as shown by the success of the use of the phylogenetic distance of a mutated sequence to the wild-type for predicting the fitness effect of these mutations [Laine, Mol. Biol. Evol. 2019].

    To better show this effect we will update Figure 6, reporting also the scores of the « scrambled » sequences, which do not respect potential epistasis extracted by the RBM. It appears that ASR sequences generally have better scores than the scrambled sequences, and lower than RBM sequences (sampled at T=1/3). RBM models takes into account multiple-residues correlations, which could contribute to reaching better scores than ASR and BM models. Ongoing studies on larger proteins show that the score of sequences sampled from ASR reconstruction, including the Maximum Likelihood one, can still be improved according to the RBM score by a few mutations consistent with the ASR posterior probabilities (unpublished).

    Mistakes in the reference list will be amended in the updated version.

  9. Version published to 10.7554/elife.110491.1 on eLife
  10. eLife

    eLife Assessment

    In this important study, the authors demonstrate that generative AI techniques (restricted Boltzmann machine) can be used effectively to design and characterize mutational pathways of WW domains with different binding specificities. The computational studies are complemented by experimental validations, and the results provide solid evidence supporting the idea that sequence landscape holds significance in understanding protein evolution from a transition path perspective. The minor weakness of the study in the current form concerns limited success in designing variants with smoothly varying binding specificities. Nevertheless, the work will likely have a major impact on research aimed at understanding how evolution navigates fitness landscapes as well as reconstructing ancestral sequences.

  11. eLife

    Reviewer #1 (Public review):

    Summary:

    The authors aim to study mutational paths connecting WW domains with different binding specificities. Their approach combines an unsupervised sequence generative model based on RBMs with a path-sampling algorithm. The key result is that most intermediate sequences along the designed transition paths retain measurable binding activity in wet-lab assays, whereas paths containing the same mutations introduced in a randomized order are largely non-functional. This difference is attributed to epistatic interactions captured by the RBM model.

    Strengths:

    Exploring mutational paths in high-dimensional protein sequence space is a challenging problem. The computational framework used here is state-of-the-art and is strengthened by systematic experimental characterization of binding activity. The study is comprehensive in scope, including multiple transition paths both within and across WW specificity classes, and the integration of modeling with high-throughput experimental validation is a clear strength.

    Weaknesses:

    A major concern is whether the stated goal of specificity switching is fully achieved. Along the sampled transition paths, most intermediate variants appear to retain specificity close to either the initial or the final class, rather than exhibiting gradually shifting specificity. For example, in Figure 4G (Class I to Class II/III), binding appears largely binary, with intermediates behaving similarly to one of the endpoints. A similar pattern is observed in Figure 3H for the Class I to Class IV transition, where binding responses are close to 0 or 1. In this sense, the specificity-switching objective is only partially realized by assigning two endpoints with different specificity. This raises a broader conceptual question: is it possible that different WW specificities evolved from a common ancestor without passing through intermediates that exhibit mixed or intermediate specificity? If so, then inferring specificity-switching pathways purely from extant natural sequences may be fundamentally challenging.

  12. eLife

    Reviewer #2 (Public review):

    This is an extremely important work that shows how one can use generative models to construct specificity-switching mutational paths in complex fitness landscapes. The experimental evidence is very clear, and the theoretical tools are innovative.

    The work will likely have a deep impact on future research aimed at understanding how evolution navigates fitness landscapes as well as reconstructing ancestral sequences.

    The manuscript is extremely clear and well written, the experimental evidence is strong, and the methods are clearly described, so I do not have major issues to raise. A few minor issues are listed below.

    (1) I consider the WW domain as an 'easy' case from the point of view of generative modelling. The domain is rather short, epistatic effects are not very strong (e.g. Boltzmann learning usually converges very quickly to a very paramagnetic state), and the resulting models are well interpretable (e.g. the hidden units of the RBM correlate well with subclasses).

    This is not always (not often?) the case, however. In more complex proteins, the learning procedures can be slower and the resulting models less interpretable. Just for completeness, perhaps the authors could comment on the generality of the results and what they would expect for other systems based on their experience.

    (2) In Section 3.3, the authors say that direct paths connecting Class I and Class IV behave similarly to indirect paths, despite having lower scores according to the RBM. How generic is this? Does it also happen for other classes? This might be an important point to address, as direct paths are easier to sample.

    (3) The path shown in Figure 4 goes through a region of non-functionality around sequences 18-19. It seems that the sample path is basically exploring the functional regions for Class I and Class II/III separately, trying to approach the other class, but then it can't really make the switch.

    By contrast, the path going from Class I to Class IV seems able to perform the functional switch in a single step (20-21) without losing too much of the function.

    Perhaps the authors could better comment on this? Is this a limitation of the sampling method, or a fundamental biological fact?

    (4) On page 12, it is stated that the temperature was chosen to 1/3 to maximize the score. This is important and should be mentioned earlier (I didn't notice it until that point).

    (5) On page 13, it is stated that: "However, the scores of the ancestral sequences along the phylogenetic pathways assigned by the RBM are significantly lower than the ones of the RBM-designed sequences. This result is expected as ASR reconstruction does not take into account epistasis, differently from RBM, and we expect ASR sequences to generally be of lesser quality."

    I was very surprised by this result. My own experience with ASR shows that, on the contrary, sequences found by ASR (via maximum likelihood) tend to have high scores in the (R)BM, and tend to be more stable than extant sequences. I attribute this to the fact that ASR typically finds a "consensus" sequence that maximizes the contribution to the score coming from the fields (the profile), which is typically dominant over the epistatic signal, resulting in a bigger score. Maybe the authors did not use maximum likelihood in the ASR? Some clarification might be useful here.

  13. Mentioned by @aixbiobot.bsky.social on Bluesky

    Design and experimental characterization of specificity-switching mutational paths of WW domains [new] WW domain specificity shifts via designed mutations. Binding assays confirm epistatic interactions are key.

  14. Mentioned by @biorxiv-bioeng.bsky.social on Bluesky

    Design and experimental characterization of specificity-switching mutational paths of WW domains https://www.biorxiv.org/content/10.64898/2025.12.08.693000v1

  15. Mentioned by @biorxivpreprint.bsky.social on Bluesky

    Design and experimental characterization of specificity-switching mutational paths of WW domains https://www.biorxiv.org/content/10.64898/2025.12.08.693000v1

  16. Version published to 10.64898/2025.12.08.693000v1 on bioRxiv