The paper ID ECOG-06559 uses a recently developed joint species distribution model (HMSC) to explore the variation in abundance of 59 reef fish species in four oceanic islands in Brazil. Authors were particular interested on how species response to environmental factors depend on species traits and phylogeny. Although this is an interesting exercise, I couldn’t find enough advances and novelty in the paper that would support its publication in Ecography. For example, the title highlights that phylogenetic history drives species response to environmental conditions. This phylogenetic signal has been extensively found in the literature, at a point where some authors suggest that its presence should be the null hypothesis (see Losos 2008 https://doi.org/10.1111/j.1461-0248.2008.01229.x). I’m also concerned with the choice of the authors to focus on analyzing species abundance conditioned to the presence (ACP). Zeros are omnipresent in community matrices and omitting them may blur our understanding of natural ecosystems. My own experience with these models indicates that ACP models are more unstable and have low predictive power, the letter also observed by the authors (L. 620-622). Below, I listed some specific comments that may help authors to improve the ms.

L. 100 Replace “hard (Tethys and Panama) barriers “ with “hard barriers (Tethys and Panama)”
L. 123 According to the previous sentence, the total number of sampling sites is eleven, not ten.
L. 174 Add a comma after “values”.
L. 188 Replace “codes” with “package”.
L. 192-196 HMSC uses residual co-occurrence matrices to infer the role of biotic interactions. However, these matrices can still contain patterns associated with other factors, including unmeasured environmental covariates and model misspecifications (Poggiato et al. 2021). Thus, authors should be careful in interpreting the outputs of this modeling approach. A sentence highlighting this limitation should be provided by the authors.
L. 219-224 If the authors did not interpret the results associated with presence-absence, then it shouldn’t be considered a hurdle model approach.
L. 243-245 Rabosky et al. super tree is for ray-finned fishes. How about Chondrichthyes? Were they removed from analyses?
L. 274 I suggest authors provide the predictive power as well. This can be easily calculated in the HMSC package using cross-validation.
L. 289-290 “the posterior probability > 0.9 was used to analyze how species respond to the environmental covariates” Is that right? I think the authors are referring to the level of support used to consider an effect statistically significant. Please, revise.
L.298-302 HMSC outputs include exploratory power for each species individually. So, I assume the values reported are averages across all species. Please, provide the standard deviation associated with the average values as well.
L. 304-320 These are relative contributions, right? Authors should make this clearer in the text. In addition, authors should indicate when the values represent averages.
L. 329-331 The addition of second-order terms is not described in the methods.
L. 351-352 How so? The significance of second-order polynomials may indicate lower abundances (or probability of occurrence) at intermediate values.
L. 400-402 The significance of these average values is hard to assess without presenting the credibility intervals.
L. 402-403 See my comment above. Authors should be much more careful in how interpreting residual co-occurrence matrices.
L. 420-422 Why not explore diversity gradients?
L. 422-427 But only the relationship presented in panel 6d has significant statistical support (prob >0.9)
L. 439-441 Significant support seems weak, especially for 7A, which is not significant (Prob < 0.9)
L. 475-485 This text fits better the introduction, not the discussion. The first paragraph of the discussion should highlight the main findings of the study.
L. 527 This is the relative contribution of the variable (note that the variance proportion presented in figure 2 is equal to 1 for all species). Thus, the predictor is explaining a lower amount of abundance variation.
L. 620-622 This may indicate model overfitting. It would be helpful if the authors provide the number of sampling units used for each model created. Looking to figure 6, 7, and 8, the sample size doesn’t seem to be large enough to accommodate four environmental variables with polynomial terms.
L. 634-637 That makes sense, but it is important to highlight that SDM and JSDM (such as HMSC) capture the effect of both abiotic and biotic filters (realized niche). Thus, the signal of species interaction that remains in the residuals after accounting for covariates is expected to be weak (Poggiato et al. 2021).
Figure 1 – This figure is very similar to Figure 3 from Ovaskainen et al. 2017 (https://doi.org/10.1111/ele.12757). Authors should either remove this figure or mention that it was adapted from Ovaskainen et al. 2017. Personally, I don’t think this figure is necessary. Instead, a study area map with the sampling points would be much appreciated in the main text.
Figure 5. Names of the species are impossible to read. Please revise.

Overall, the authors did a good job of addressing the issues I raised in the previous round of revisions. However, I`m still skeptical about the R2 values provided by the authors. In the response letter, they state:
“We are aware of that risk (overdispersion) as we discussed this issue with Otso Ovaskainen… We
kept environmental variables at a minimum using only those with known influence on
abundances. We tried removing mixed layer depth from the list of environmental covariates but
it did not result in a better performance”
In my point of view, that is not sufficient. Overfitting is a serious problem and may indicate that authors are using an overly complex model to explain a limited set of sampling points. As a result, the model is explaining random noise and, consequently, performs poorly when modeling unseen data. Authors should be transparent with the readers and provide the predictive performance of the models, which was only presented in the response letter. There, it is possible to see that most R2 values drop to less than 0.2. In many instances, the values go below 0, which indicates that the model performs worse than random predictions.

Specific comments
L. 72 Replace “Otherwise” with “Conversely”.
L. 303. What is adequate exploratory power? Not clear.
L. 320 “04, 49 and 36 respectively” Not clear what the values represent.
L.404 Remove ,-
L. 479 “Area Under the Curve, AUC > 0.9 and 0.7 <” Do the authors mean AUC values between 0.7 and 0.9? Confusing.
L. 528 “Halichoeres (maximum R2 = 0.98, see Table S6), Holacanthus (maximum R2 = 0.95)” and L. 533 “Stegastes (maximum R2 = 0.96)” Something seems off here. These numbers are really high, especially when referring to abundance models. My main concern is that these values are associated with overfitting.
L. 637 What does “unsatisfactory results” mean? Please, be more specific.
Figure 2. It is still hard to read the axis labels and legend in figure 2. The resolution of the figure could be improved.