12 major claims checked against the paper's own evidence: all adequately supported.
partialAbstract, DiscussionReviewer 1
IDiom can serve as a general platform for generative design of intrinsically disordered proteins and regions.
The paper demonstrates generation and post-training for localization, but does not yet show experimental validation (e.g., in vitro or in vivo testing) of designed sequences, which would be needed to fully support a 'platform' claim for design.
Evidence: The paper shows computational validation (pLDDT, sequence metrics, motif analysis) but no wet-lab experimental confirmation of designed sequences.
“These results establish IDiom as a general platform for the generative design of intrinsically disordered proteins and regions.”
partialAbstractReviewer 3
IDiom provides a general platform for generative design of intrinsically disordered proteins and regions.
The paper demonstrates generation and post-training for one function (subcellular localization). The claim that it is a 'general platform' is supported by the architecture but not yet shown for multiple diverse objectives (e.g., phase behavior, binding affinity, enzymatic activity). The paper does suggest extensibility but provides only one example.
Evidence: The paper states in the Discussion that IDiom can be combined with other objectives, but only demonstrates it with ProtGPS.
“These results establish IDiom as a general platform for the generative design of intrinsically disordered proteins and regions.”
supportedAbstract, Results (paragraphs 4-8)Reviewers 1, 2
IDiom generates diverse sequences that recapitulate biologically relevant sequence features of natural disordered regions.
The paper provides extensive evidence through distribution comparisons of FCR, κ, SHD, and complexity metrics showing close matches to DisProt and training set IDRs.
Evidence: Figures 1f, 2b–e show compositional and metric distributions comparing generated sequences to natural DisProt and training set IDRs.
“We demonstrate that the model generates diverse sequences that recapitulate biologically relevant sequence features of natural disordered regions.”
supportedAbstract, Results (paragraphs 9-10)Reviewer 1
The model can generate disordered region sequences conditioned on their surrounding structured context.
The paper demonstrates context-prompted generation using DisProt flanking contexts and provides evidence through Wasserstein distances and NPM1 case study showing conditioned generations are closer to natural IDRs.
Evidence: Figures 2b–e right subplots show lower W1 distances for prompted IDRs vs unprompted IDPs; Figure 3 shows κ distribution for NPM1-prompted generations.
“We demonstrate that post-training via reinforcement learning with a subcellular localization reward model produces sequences with features which are consistent with known sequence determinants of compartment-specific localization.”
supportedAbstract, Results (paragraphs 11-13)Reviewer 1
Post-training via reinforcement learning with a subcellular localization reward model produces sequences with features consistent with known determinants of compartment-specific localization.
The paper shows compartment-specific amino acid compositions, NLS enrichment, PTM motif density, and RNA-interaction motifs that align with known biology for each of the four compartments analyzed.
Evidence: Figures 5a–d (composition), 6a–e (κ, NLS, PTM, RNA motifs) for nucleolus, chromosomes, P-bodies, and stress granules.
“These results establish IDiom as a general platform for the generative design of intrinsically disordered proteins and regions.”
supportedResults, paragraph 9Reviewer 1
IDiom learns in-context: given flanking sequence context, it generates disordered spans whose sequence features are more appropriate for that context than unprompted generations.
The NPM1 case study (Figure 3) directly shows that NPM1-prompted generations have κ values closer to the WT NPM1 IDR than scrambled or DisProt sequences, and the Wasserstein distances (Figure 2b–e) quantitatively confirm that prompted generations are closer to natural IDRs than unprompted ones.
Evidence: Figure 3b (κ distributions), Figure 2e right subplot (W1 distances lower for prompted IDRs across metrics).
“We further show that IDiom learns in-context: given the flanking sequence context of a specific protein, the model generates disordered spans whose sequence features are more appropriate for that context than unprompted generations.”
supportedAbstractReviewer 2
Post-training via reinforcement learning with a subcellular localization reward model produces sequences with features consistent with known sequence determinants of compartment-specific localization.
The paper shows that post-trained models generate sequences with compartment-specific amino acid compositions, charge patterning, and motif enrichment (NLS, PTM, RNA-binding motifs) that are biologically interpretable and consistent with known biology.
Evidence: Figures 5 and 6 show compositional biases, κ values, NLS frequency, PTM motif density, and RNA-interaction motif enrichment for each compartment, all consistent with known biology.
“we demonstrate that post-training via reinforcement learning with a subcellular localization reward model produces sequences with features which are consistent with known sequence determinants of compartment-specific localization.”
supportedIntroduction, paragraph 4Reviewers 2, 3
IDiom learns in-context: given flanking sequence context, the model generates disordered spans whose sequence features are more appropriate for that context than unprompted generations.
The paper shows that Wasserstein distances for DisProt context-prompted IDRs are consistently lower than for unprompted IDPs across all metrics, and the NPM1 case study demonstrates context-specific charge patterning.
Evidence: Results section 'Conditioned generation recapitulates context-specific IDR sequence features' and Figure 3 show lower W1 distances for prompted generations and the NPM1 case study with κ distributions shifted toward the wild-type value.
“We further show that IDiom learns in-context: given the flanking sequence context of a specific protein, the model generates disordered spans whose sequence features are more appropriate for that context than unprompted generations.”
supportedAbstractReviewer 2
IDiom provides a general platform for the generative design of intrinsically disordered proteins and regions.
The paper demonstrates the platform's capabilities through pre-training, context-prompted generation, and post-training with a reward model, establishing its generality. The claim is appropriately scoped as a platform, not a finished product.
Evidence: The entire paper demonstrates the platform: data curation, model training, sequence generation, and post-training with ProtGPS. The Discussion outlines future applications with other reward models.
“These results establish IDiom as a general platform for the generative design of intrinsically disordered proteins and regions.”
supportedAbstractReviewer 3
IDiom generates diverse sequences that recapitulate the compositions, patterning, and motifs of natural intrinsically disordered regions.
The paper supports this claim with extensive analysis of amino acid enrichment, charge patterning (κ), hydrophobic decoration (SHD), and sequence complexity, showing that generated sequences closely match DisProt IDRs across all these metrics.
Evidence: Figures 1f, 2b–e, and the accompanying text in the Results section show distribution comparisons and Wasserstein distances.
“IDiom generates diverse sequences that recapitulate the compositions, patterning, and motifs of natural intrinsically disordered regions.”
supportedResults, paragraph 3Reviewer 3
The model generates diverse sequences that remain substantially dissimilar to training examples.
Figure 1d shows that the maximum sequence identity to training set IDRs peaks broadly around 60%, indicating that most generated sequences are substantially dissimilar to any training example.
Evidence: Figure 1d and the text: 'The distribution of maximum sequence identities to training set IDRs peaks broadly around 60%'.
“The distribution of maximum sequence identities to training set IDRs peaks broadly around 60%, indicating that most generated sequences are substantially dissimilar to any sequence seen during training.”
supportedAbstractReviewer 3
IDiom can be post-trained using reinforcement learning with a ProtGPS reward to design disordered sequences with targeted subcellular localization.
The paper demonstrates post-training for four compartments (nucleolus, chromosomes, P-bodies, stress granules) and shows that generated sequences exhibit compartment-specific compositional biases, motif enrichment, and sequence features consistent with known biology.
Evidence: Figures 4, 5, 6 show training curves, compositional biases, and motif enrichment for each compartment.
“We demonstrate that post-training via reinforcement learning with a subcellular localization reward model produces sequences with features which are consistent with known sequence determinants of compartment-specific localization.”