Comentários
Escrever um comentárioNenhum comentário foi publicado ainda.
VReceptor is a simple, interpretable virtual receptor. Position importance comes from a mixture of Gaussians over the sequence. Each substitution is scored by how far its amino acid is from the reference residue (z-scale distances: cosine, Chebyshev-like or L2). The scores are summed and noise is added to give a simulated log IC50.
Design-strategy demo: on a 31-mer (PrRP), an LSTM trained on 589 random triple mutants reached a test MSE of about 0.02. The same model trained on a full single-mutant scan (also 589 peptides) reached about 0.14 when tested on triple mutants.
Realism check: fitting about 6 parameters to PrRP GPR10 binding data gave test Pearson r of 0.83 (Chebyshev and L2) and 0.77 (cosine). A 155-parameter ridge regression got 0.80 on the same split. The fits were worse for NPFFR1 and NPFFR2.
It offers a transparent, controllable data generator for testing peptide library design, active learning and ML models before spending wet-lab effort. It is also a possible interpretable baseline. The contribution is modest and mostly conceptual: it is a useful tool idea, but the paper doesn't yet show that conclusions drawn on VReceptor data transfer to real peptide landscapes
The scan-vs-random result is close to circular. The data come from a purely additive, position-wise model with no epistasis. The test set is triple mutants, which a single-mutant scan never samples. It is a single run with no seeds or confidence intervals, and there is no simple additive baseline, such as ridge, to separate library design from LSTM generalization. This can't support claims about real experimental planning.
Realism is under-validated. The test uses one receptor, a filtered subset (relGPR10 < 2.0), about 250 training peptides, and a random 50:50 split, which likely leaks between near-identical analogs. The test MSE of 0.118 (RMSE about 0.34) is far above the stated assay SD of 0.1. Fitting mean affinity also doesn't show that the generated data are realistic in distribution, noise or multi-mutant behavior. Results for the two other receptors are only described qualitatively.
Overclaiming relative to ridge. The "enhanced generalization" claim rests on a 0.83 vs 0.80 Pearson gap on one split. The ridge baseline is weakly specified (alpha "empirically determined", with no stated tuning protocol), and the conclusion's "validated" contradicts the introduction's "plausibility only". The speculation about matching state-of-the-art ML is untested.
Model expressivity. Distances are nonnegative and measured only from the reference residue. A substitution is penalized for being different, not for being the wrong kind of residue at that position. Affinity-improving mutations can't be generated except through noise or a negative amplitude, and a negative amplitude would make every substitution at that position improve binding. No interactions are modeled, though this limitation is acknowledged.
No comparison to existing generators such as the Poli oracles. The claimed advantages in interpretability and realism are never tested.
Reproducibility. There is no GitHub URL, LSTM hyperparameters, training details, seeds, or the fitted parameter values. It is unclear whether the PrRP data are public. The 155 ridge parameters (31 × 5) suggest z-scale features rather than one-hot, but this isn't stated.
Fit parameters: it is unclear whether Gaussian widths are fixed or fitted (the text mentions only positions and amplitudes, yet 6 parameters are reported).
"Pharmacophore" usually means a 3D arrangement of interaction features. Here it is a 1D positional importance profile, which the authors should acknowledge.
The ridge test MSE in Table 1 (0.140) is identical to the scan-design test MSE in Fig. 4 (0.140), which are unrelated experiments. Please confirm this is not a copy error.
The author declares that they have no competing interests.
The author declares that they used generative AI to come up with new ideas for their review.
I read the manuscript thoroughly and formed my critical assessment independently. I then used an AI language model to help rephrase and structure my findings into this report. I checked all comments against the manuscript and take full responsibility for the content.
Nenhum comentário foi publicado ainda.