acceptodds
Under review as a conference paper at ICLR 2027

ProtGround: Functional Phrase Grounding in Local Protein 3D Structural Regions

Abstract

Protein functions often depend on local 3D environments, making it important to localize functional descriptions to their supporting structural regions. However, existing protein–text pretraining methods mainly operate at the whole-protein level, leaving these local correspondences unresolved. To this end, we introduce ProtGround, a fine-grained multimodal pretraining framework that integrates protein sequences, local protein 3D structural regions, and functional descriptions to ground functional phrases in the regions that support them. Specifically, we first construct ProtGround-Align, a large-scale dataset of fine-grained local region–function pairs. Building on this dataset, we propose ProtGround-Hier, a pretraining method that localizes the local protein regions supporting a functional phrase through joint modeling of protein–region–phrase relationships. It uses multi-positive region–phrase alignment, residue–token optimal transport, and multi-level masked modeling. Extensive experiments show that ProtGround consistently outperforms strong baselines on fine-grained grounding and protein understanding tasks. These results show that grounding functional language in local regions enables function-aware, spatially localized protein understanding and improves transfer across tasks. The data and code will be made publicly available.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein representation learning, protein-text learning, fine-grained protein grounding, local protein 3D structure, structure-function alignment, multimodal pretraining, protein language models

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_LHu5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

ProtGround learns to associate functional phrases with local protein structures. The authors construct region–phrase pairs from curated residue annotations and semantic matching, then combine contrastive learning, residue–token optimal transport, and masked modeling. They evaluate protein retrieval, candidate and blind localization, and downstream protein prediction.

Critical strengths

The local supervision is the part I find most valuable. Most pairs come directly from curated residue annotations, giving the dataset a firmer basis than associations generated from protein descriptions alone. The resource could be useful beyond this particular architecture. The reported localization gains are also substantial: U30-hard Core mAP increases from 0.2670 for ProTrek to 0.5129. These results make the approach worth investigating.

Critical weaknesses

I found the related work too thin to judge the novelty properly. Section 5 offers only a brief contrast with whole-protein modeling. DeepFRI is cited as a benchmark, but its residue-level localization is not discussed; COLLAPSE’s local structure–function representations are also relevant. Neither solves exactly the same phrase-grounding task, but both matter for establishing what language conditioning adds beyond existing functional-site learning. That distinction is currently missing.
I also could not reconstruct the blind-grounding procedure. Section 3 starts from a specified local region, whereas Level 3 requires searching the complete protein. How are whole-protein residue scores obtained, converted into five predicted sites, and matched to the annotations? The appendix does not supply this missing inference procedure. Since blind grounding is a central result, these details are necessary to assess Table 2. The adaptation of the grounding baselines also needs to be specified.

Questions for authors
  1. Please give the complete Level 3 inference procedure and the exact SiteRecall@5 matching rule.
  2. Were the grounding baselines adapted using ProtGround-Align? What local supervision and prediction procedure did each receive?
  3. It is better to provide an anonymized codebase for review rather than deferring the release until after acceptance.
5 comments
AuthorsOpenReview ↗

Response to Reviewer Comments

For W1. Distinction from Prior Local Function Modeling Methods. We thank the reviewer for highlighting DeepFRI and COLLAPSE. Due to space limitations, our original Related Work section provided only a brief discussion of existing local function modeling methods. We have expanded this discussion in the revised manuscript.

DeepFRI predicts protein functions labels using sequence and applies Grad-CAM to generate residue-level importance maps for GO and EC function labels. Although it can identify relevant residues, its localization is derived from post-hoc interpretation of predefined function-label predictions, rather than explicit supervision of functional phrase–region correspondences. Therefore, DeepFRI focuses on identifying individual functionally important residues rather than directly grounding free-form functional descriptions in local 3D structural regions, where protein function depends on both the residues and their surrounding structural microenvironments.

COLLAPSE learns representations of local sites through evolutionary self-supervision. It provides function representations of protein sites but does not incorporate natural-language descriptions or explicitly align functional phrases with local structural regions. Consequently, its learned site embeddings are not directly conditioned on functional phrase queries.

In contrast, ProtGround explicitly connects natural-language functional phrases with local 3D structural regions. Rather than focusing solely on individual functional residues, we consider both these residues and their surrounding 3D microenvironments, where spatially close but sequence-discontinuous residues can jointly contribute to protein function. We construct large-scale, explicitly annotated region–phrase pairs that capture many-to-many functional relationships. ProtGround-Hier further learns these correspondences through geometry-aware region encoding, multi-positive alignment, and residue–token optimal transport. Given a functional phrase, the model can directly localize its corresponding 3D functional regions and identify their core residues, rather than relying on post-hoc residue importance maps or text-independent structural site embeddings.

Thus, our main contribution is not simply residue-level localization or local structural representation learning, but language-conditioned grounding of distinct functional descriptions in their corresponding local 3D microenvironments.

We have now substantially expanded Section 5 (Related Work) of the revised manuscript, systematically reviewing the development of protein representation learning from whole-protein modeling to fine-grained local function modeling. We discuss representative methods, including ProtST, ProtCLIP, DeepFRI, COLLAPSE, MMSite, and the recent ProLoc, highlighting their methodological contributions and limitations. This expanded discussion clarifies the distinctions between existing approaches and ProtGround, particularly our explicit alignment of functional phrases with local 3D structural microenvironments, and better establishes the novelty and advantages of our work.

AuthorsOpenReview ↗

Response to Reviewer Comments

For W2 and Q1-Q2. Blind Grounding Implementation and Baseline Adaptation. We thank the reviewer for raising these questions. We have clarified the complete Level 3 inference procedure, evaluation metrics, and baseline adaptation protocols in the revised manuscript.

1. Level 3 Blind Grounding Procedure. Unlike Level 2, which ranks predefined candidate regions, Level 3 performs localization across the entire protein without requiring known functional sites or predefined candidate regions. Given a complete protein sequence, its 3D structure, and a functional phrase query, we construct a 10 Å spatial neighborhood around every canonical residue, using minimum interatomic distances. Each neighborhood includes its center residue and surrounding spatial neighbors, including residues that may be distant in sequence.

ProtGround-Hier encodes each local neighborhood using its sequence and 3D geometric information. For each functional phrase, we compute cosine similarity between the phrase embedding and every local region embedding. The resulting similarity score is assigned to the corresponding center residue, producing a query-dependent score for every residue across the protein. We rank all residue centers by these scores and select the five highest-scoring centers. Their corresponding 10 Å neighborhoods constitute the top-5 predicted functional regions. No ground-truth functional sites or annotated core residues are required during inference.

2. Core mAP and SiteRecall@5. Core mAP evaluates how highly the annotated functional core residues are ranked among all residue centers. For each functional phrase query q, Average precision (AP) averages precision values at the ranks of all annotated core residues, where precision at each rank is the fraction of true core residues among the residues ranked up to that position, average precision is computed as:

AP(q)=(1/∣Cq∣)∑k=1LqPq(k)⋅I[rk∈Cq]AP(q) = (1/|C_q|) ∑_{k=1}^{L_q} P_q(k) · I[r_k ∈ C_q]

Pq(k)=(1/k)∑j=1kI[rj∈Cq]P_q(k) = (1/k) ∑_{j=1}^{k} I[r_j ∈ C_q]

Here, CqC_q is the set of annotated core residues, LqL_q is the number of canonical residues, rkr_k is the residue at rank k, and I[·] is the indicator function. Core mAP is the mean AP across all functional phrase queries.

SiteRecall@5 evaluates whether the top five predicted neighborhoods recover the annotated functional sites. For SiteRecall@5, we select the five highest-scoring local regions per query and consider a ground-truth functional site recovered if their union covers any of its annotated core residues., Let Uq5U_q^5 denote the union of the five highest-scoring predicted regions:

Uq5=Rq,1pred∪Rq,2pred∪Rq,3pred∪Rq,4pred∪Rq,5predU_q^5 = R_{q,1}^{pred} ∪ R_{q,2}^{pred} ∪ R_{q,3}^{pred} ∪ R_{q,4}^{pred} ∪ R_{q,5}^{pred}

SiteRecall@5 is calculated as:

SiteRecall@5=(1/N)Σq=1N(1/∣Sq∣)Σs∈SqI[Cq,s∩Uq5≠∅]SiteRecall@5 = (1/N) Σ_{q=1}^{N} (1/|S_q|) Σ_{s ∈ S_q} I[C_{q,s} ∩ U_q^5 ≠ ∅]

Here, N is the number of functional phrase queries, SqS_q is the set of ground-truth functional sites for query q, and Cq,sC_{q,s} is the annotated core-residue set of site s.

A ground-truth site is considered recovered if at least one of its annotated core residues is covered by the union of the top five predicted regions. We calculate the fraction of recovered sites for each query and average it across queries. Thus, Core mAP measures residue-center ranking quality, while SiteRecall@5 measures functional-site recovery by the highest-scoring local 3D regions.

3. Baseline Adaptation for Local Grounding. We evaluate ProteinCLAP, ProtCLIP, Prot2Text-V2, and ProTrek using a shared candidate-generation and evaluation protocol. For ProteinCLAP, ProtCLIP, and Prot2Text-V2, we first encode the complete protein sequence to obtain contextualized residue representations. For each candidate 3D region, we identify its constituent residues and pool their corresponding full-sequence embeddings to construct a region-level representation. We then compute cosine similarity between this representation and the functional phrase embedding to rank the regions. For ProTrek, we additionally incorporate its structural embedding into the region representation. In contrast, ProtGround-Hier directly encodes each local 3D region using its sequence and geometric information.

At Level 2, all methods evaluate the same predefined candidate regions. At Level 3, all methods use the same 10 Å neighborhoods centered at every canonical residue. The resulting region scores are assigned to their center residues, and the same top-5 selection and site-matching rules are applied to all methods.

Therefore, all methods share the same candidate regions, functional queries, and evaluation metrics. Their differences lie in how local regions are represented and aligned with functional phrases.

AuthorsOpenReview ↗

Response to Reviewer Comments

4. Training Supervision. The baselines are evaluated using their original pretrained representations. To further investigate the contribution of our model architecture, we conducted an additional structural encoder ablation by replacing our geometry-aware region encoder with ProTrek's structure encoder, while keeping the training data, objectives, and remaining model components unchanged. As shown in the table below, this replacement consistently reduces performance across all grounding and downstream metrics. In particular, Level-3 Core mAP decreases from 0.537 to 0.450, and SiteRecall@5 decreases from 0.865 to 0.786. These results demonstrate that our task-specific local 3D structural encoder contributes to the performance improvements beyond the fine-grained training data itself.

Ablation study of pre-training objectives and structural encoding in ProtGround-Hier.

Config.Grounding Level1 mAPGrounding Level2 mAPGrounding Level3 Core mAPGrounding SiteRecall@5Loc. Fixed (Acc.%)Loc. Full (Acc.%)Fit. Fixed (ρ)Fit. Full (ρ)Func. Fixed (Fmax)Func. Full (Fmax)
Full objective0.7220.9140.5370.86589.9089.750.6730.8130.6330.696
w/o LGCL_{GC}0.6940.8680.4870.81987.4488.010.6470.7900.5920.658
w/o LOTL_{OT}0.6820.8410.4620.79787.8688.360.6460.7870.5890.654
w/o LMMPL_{MMP}0.7100.8940.5160.84988.8788.920.6590.7910.6100.666
w/o LMPML_{MPM}0.7030.8850.5080.84088.7288.740.6620.7930.6070.664
w/ Alt. 3D Encoder0.6770.8350.4500.78687.9388.450.6530.7950.5980.657

We have incorporated this additional structural encoder ablation into Table 4 of the revised manuscript.

Manuscript Revision. We have explicitly clarified the Level 3 inference procedure, residue-level scoring, top-5 region selection, and exact definitions of Core mAP and SiteRecall@5 in Lines 359-425 of the revised manuscript. We have also expanded the baseline descriptions in Lines 418-425, specifying the region-level representations, shared candidate-generation procedure, and similarity-based ranking used for local grounding.

We hope our detailed responses and manuscript revisions have addressed your questions and concerns about novelty. If so, we would sincerely appreciate your consideration of a higher rating. Thank you very much for your valuable feedback.

Reviewer_VxxtOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of grounding functional descriptions in local regions of protein 3D structures. The work consists of two main components: dataset construction and model evaluation. The dataset construction appears to involve substantial effort. I find the model evaluation somewhat incomplete, but overall I acknowledge the substantial work involved. In dataset construction, I’m not clear on how the 10 Å cutoff was chosen for defining local structural environments during dataset construction. About the comparison of baselines, the proposed model requires a specified region as input, but the paper does not explain how Level 3 blind grounding is implemented without predefined regions, how are the Level 3 evaluation metrics calculated, and how are the sequence–text baselines adapted to perform the Level 2 and Level 3 grounding tasks?

Critical strengths
  1. The paper addresses a meaningful problem: grounding functional descriptions in local protein 3D regions, moving beyond whole-protein–text matching toward more spatially specific protein understanding.

  2. The work combines a large-scale region–function dataset with a dedicated modeling framework. The reported improvements across retrieval, candidate grounding, and blind grounding, together with downstream evaluations, suggest the potential value of learning local structure–function correspondences.

Critical weaknesses
  1. The implementation of Level 3 blind grounding is insufficiently described. The method is primarily formulated for predefined local regions, but it is unclear how the model processes an entire protein without candidate regions, produces query-dependent residue scores, and converts these scores into the top five predicted regions. The definitions of Core mAP and SiteRecall@5 also need clarification, particularly the criterion for a predicted region to recover a true site. These omissions make the central grounding results difficult to assess and reproduce.

  2. The adaptation of the baselines to local grounding is unclear. For sequence–text models such as ProteinCLAP and ProtCLIP, the paper should explain how candidate-region representations and residue-level predictions are obtained for Levels 2 and 3. It should also clarify whether the baselines are trained or fine-tuned on the same region–function pairs and whether they share the same region-generation procedure. Without these details, it is difficult to separate the proposed method's contribution from differences in supervision or task adaptation.

Questions for authors
  1. What motivates the choice of a 10 Å neighborhood for constructing local regions? Have you examined how alternative neighborhood sizes affect grounding performance?

  2. How are ProteinCLAP and ProtCLIP adapted to Levels 2 and 3? Are the baselines trained or fine-tuned on the same region–function pairs, and do all methods use the same region-generation procedure?

  3. How is Level 3 blind grounding implemented without predefined candidate regions? Please clarify how query-dependent residue scores and the top five predicted regions are obtained, and specify the definitions of Core mAP and SiteRecall@5, including the site-matching criterion.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Comments

We sincerely thank the reviewer for their positive assessment of our work and valuable feedback.

For W1 and Q3. Level 3 Blind Grounding Implementation and Evaluation Metrics. We thank the reviewer for highlighting the need for clearer implementation and evaluation details. We have clarified the complete Level 3 Blind Grounding procedure and the definitions of Core mAP and SiteRecall@5 in the revised manuscript.

1. Blind Grounding and Query-dependent Residue Scoring.​ Unlike Level 2, which ranks predefined functional candidate regions, Level 3 performs localization across the entire protein without requiring any predefined functional sites or candidate regions.​ Given a complete protein sequence, its 3D structure, and a functional phrase query, we construct a local 10 Å spatial neighborhood around every canonical residue. Each neighborhood contains the residues within a minimum interatomic distance of 10 Å from its center, including residues that may be distant in sequence but adjacent in 3D space.​

ProtGround-Hier then encodes each neighborhood using its local sequence and 3D structural information and computes the cosine similarity between the resulting region embedding and the functional phrase embedding. The similarity score of each residue-centered neighborhood is assigned to its center residue, yielding a query-dependent score for every residue in the protein.​ We rank all residue centers by their similarity scores and select the five highest-scoring centers. Their corresponding 10 Å neighborhoods constitute the top-5 predicted functional regions. Importantly, this procedure requires no ground-truth functional residues or predefined functional regions during inference.

  1. Core mAP.​

Core mAP evaluates whether annotated functional core residues receive high query-dependent ranking scores.​ For a functional phrase query q, let CqC_q denote the set of annotated core residues, LqL_q the total number of canonical residues, and rkr_k the residue ranked at position k. Average precision (AP) averages precision values at the ranks of all annotated core residues, where precision at each rank is the fraction of true core residues among the residues ranked up to that position. Average precision is defined as:​ AP(q)=(1/∣Cq∣)Σk=1LqPq(k)⋅I[rk∈Cq]AP(q) = (1/|C_q|) Σ_{k=1}^{L_q} P_q(k) · I[r_k ∈ C_q]​ where:​ Pq(k)=(1/k)Σj=1kI[rj∈Cq]P_q(k) = (1/k) Σ_{j=1}^{k} I[r_j ∈ C_q]​ Here, Pq(k)P_q(k) is the precision among the top-k ranked residues, and I[·] is the indicator function, equal to 1 when the condition is satisfied and 0 otherwise.​ Core mAP is the mean AP across all functional phrase queries:​ CoremAP=(1/N)Σq=1NAP(q)Core mAP = (1/N) Σ_{q=1}^{N} AP(q)​ Thus, Core mAP measures the quality of residue-center ranking over the entire protein, rather than evaluating only the five selected regions.

  1. SiteRecall@5 and Site-matching Criterion.​

SiteRecall@5 evaluates whether the five highest-scoring predicted neighborhoods recover the annotated functional sites.​ For each query q, let SqS_q be the set of ground-truth functional sites and Cq,sC_{q,s} the annotated core residues belonging to site s. Let Rq,kpredR_{q,k}^{pred} denote the k-th highest-scoring predicted 10 Å neighborhood. We define the union of the top five predicted regions as: Uq5=Rq,1pred∪Rq,2pred∪Rq,3pred∪Rq,4pred∪Rq,5predU_q^5 = R_{q,1}^{pred} ∪ R_{q,2}^{pred} ∪ R_{q,3}^{pred} ∪ R_{q,4}^{pred} ∪ R_{q,5}^{pred} SiteRecall@5 is then calculated as:​ SiteRecall@5=(1/N)Σq=1N(1/∣Sq∣)Σs∈SqI[Cq,s∩Uq5≠∅]SiteRecall@5 = (1/N) Σ_{q=1}^{N} (1/|S_q|) Σ_{s ∈ S_q} I[C_{q,s} ∩ U_q^5 ≠ ∅]​ A ground-truth site is considered recovered if at least one of its annotated core residues is contained in the union of the top five predicted regions. For each functional phrase query, we calculate the fraction of ground-truth sites recovered by the top five predicted regions. SiteRecall@5 is the mean of these fractions across all queries.​ Together, Core mAP evaluates the ranking quality of all residue centers, while SiteRecall@5 measures whether the highest-scoring local 3D regions successfully cover the annotated functional sites. These complementary metrics assess both residue-level ranking and functional-site recovery under blind grounding.

We have explicitly clarified the Level 3 candidate construction, query-dependent residue scoring, top-5 region selection, and the mathematical definitions of Core mAP and SiteRecall@5 in Lines 359-425 of the revised manuscript, including the precise criterion for recovering a ground-truth functional site.

AuthorsOpenReview ↗

Response to Reviewer Comments

For W2 and Q2. Baseline Adaptation and Evaluation Fairness. We thank the reviewer for raising this question. We have clarified how the baseline models are adapted to local grounding and how their evaluation settings differ from ProtGround-Hier.

1. Baseline Adaptation for Levels 2 and 3. For sequence–text baselines, including ProteinCLAP, ProtCLIP, and Prot2Text-V2, we first encode the complete protein sequence to obtain contextualized residue representations. For each candidate 3D region, we identify its constituent residues in the canonical sequence and pool their corresponding residue embeddings to obtain a region-level representation. We then compute cosine similarity between this representation and the functional phrase embedding to rank candidate regions. For ProTrek, which incorporates protein structural information, we additionally use its structural embedding in constructing the region representation. In contrast, ProtGround-Hier directly encodes each local 3D region using its sequence and geometric information.

At Level 2, all models rank the same predefined candidate regions within each protein. At Level 3, we construct a 10 Å spatial neighborhood around every canonical residue and apply the same candidate-generation procedure to all methods. Each neighborhood receives a phrase-dependent similarity score, which is assigned to its center residue for Core mAP evaluation. The five highest-scoring neighborhoods are selected for SiteRecall@5 evaluation.

Thus, all methods use the same candidate regions, functional queries, and evaluation metrics. The difference lies in how each model represents and scores the local structural regions.

2. Training Supervision. The baseline models are evaluated using their original pretrained representations without additional training or fine-tuning on ProtGround-Align region–phrase pairs. In contrast, ProtGround-Hier is pretrained on ProtGround-Align with explicit fine-grained region–phrase supervision. We therefore acknowledge that the baseline comparisons evaluate the overall effectiveness of our fine-grained pretraining framework, including both the dataset and model design. To further assess the contribution of our model architecture under controlled conditions, we conducted an additional structural encoder ablation, replacing our geometry-aware region encoder with ProTrek's structure encoder while keeping the training data, objectives, and remaining model components unchanged. As reported in Table 4 of the revised manuscript, this replacement reduces Level-3 Core mAP from 0.537 to 0.450 and SiteRecall@5 from 0.865 to 0.786, providing additional evidence for the effectiveness of our task-specific structural encoding.

We have explicitly clarified the baseline adaptation procedures, shared region-generation protocol, and region-level similarity scoring in Lines 418–425 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Comments

For Q1. Motivation and Sensitivity Analysis of the 10 Å Neighborhood. We thank the reviewer for this question. The 10 Å neighborhood is motivated by the importance of local 3D microenvironments, where spatially adjacent but sequence-distant residues can contribute to protein function. Previous studies, including Bagley and Altman (1995) and MicroMiner (Sieg et al., 2023), support spatial microenvironment analysis, with 10 Å neighborhoods used in earlier structural-site studies.

To examine this choice, we conducted a geometric sensitivity analysis across seven radii (4–16 Å) on ProtGround-Unseen, together with analyses of sequence-discontinuous structural context and spatial-shell composition.

1. Neighborhood Size Sensitivity Across Seven Radii. We evaluated the same 737,481 U50 and 370,661 U30-hard residue centers at all seven radii, using minimum interatomic distances.

Neighborhood-size sensitivity across seven spatial radii.

Radius (Å)U50 MeanU50 MedianU50 P95U30-hard MeanU30-hard MedianU30-hard P95
48.889138.61913
613.73142113.291321
824.21243923.262339
1035.94366034.393460
1251.12518748.704986
1468.676811965.2465117
1687.778715283.2083149

At 10 Å, the mean neighborhood contains 35.94 residues in U50 and 34.39 in U30-hard, with a 95th percentile of 60 residues in both subsets. In U50, reducing the radius to 8 Å decreases the mean size by 32.6%, while increasing it to 12 or 16 Å enlarges the mean size by 42.2% or 144.2%, respectively. Thus, 10 Å provides an intermediate scale balancing structural coverage and compactness.

2. Sequence-discontinuous Structural Context at 10 Å. We quantified the number of distinct sequence segments within each neighborhood and the proportion of spatial neighbors separated by more than 20 sequence positions.

Sequence nonlocality of 10 Å neighborhoods.

SubsetCentersMean SizeMean SegmentsMulti-segment (%)Distant Neighbors (%)Best Sequence-window Capture (%)
U50All35.948.5792.5048.0460.24
U50Functional Core45.7412.1699.4559.2954.00
U30-hardAll34.398.1289.6346.3761.47
U30-hardFunctional Core44.3511.6599.0557.1755.13

"Distant Neighbors" is the mean per-neighborhood percentage of non-center residues more than 20 sequence positions away. "Best Sequence-window Capture" measures the mean fraction of neighborhood residues covered by the best-matching contiguous sequence window of equal size. In U50, 92.50% of neighborhoods contain multiple sequence segments, and 48.04% of non-center neighbors are sequence-distant on average. These proportions increase to 99.45% and 59.29% around annotated functional core centers. Similar patterns occur in U30-hard, confirming that 10 Å neighborhoods capture substantial sequence-discontinuous structural context.

AuthorsOpenReview ↗

Response to Q1

3. Equal-size Sequence-window Comparison. We compared each 10 Å neighborhood with a centered contiguous sequence window and the best-matching contiguous window, both containing the same number of residues.

Equal-cardinality sequence-window controls at 10 Å.

SubsetCentersCentered Capture (%)Best Capture (%)Outside Best, Pooled (%)
U50All52.9860.2443.96
U50Functional Core44.3054.0047.83
U30-hardAll54.3661.4743.70
U30-hardFunctional Core45.6055.1347.18

"Centered Capture" and "Best Capture" are mean per-neighborhood percentages; "Outside Best, Pooled" measures the fraction of pooled neighborhood–residue memberships not recovered by the best-matching windows. Even the best contiguous window captures only 60.24% of U50 neighborhood residues on average, decreasing to 54.00% around functional core centers. These results demonstrate that equal-size contiguous sequence windows cannot fully reproduce local 3D neighborhood membership.

4. Incremental Spatial-shell Analysis. We measured the residues newly included in successive spatial shells and their sequence separation from the center.

Incremental spatial-shell statistics.

SubsetCentersShell (Å)Mean Added>5 (%)>10 (%)>20 (%)
U50All8–1011.7393.7382.0171.96
U50All10–1215.1896.0486.1975.79
U50Functional Core8–1016.1595.0586.6777.19
U50Functional Core10–1220.7497.0290.0980.76
U30-hardAll8–1011.1293.0481.5871.70
U30-hardAll10–1214.3295.3485.5575.39
U30-hardFunctional Core8–1015.5094.5486.2076.17
U30-hardFunctional Core10–1219.3996.6389.8180.20

"Mean Added" denotes newly included residues per center; the remaining columns give pooled percentages exceeding the indicated sequence-separation thresholds. Expanding from 8 to 10 Å adds 11.73 residues per U50 center, of which 71.96% are separated by more than 20 sequence positions. Around functional core centers, these values increase to 16.15 residues and 77.19%. U30-hard shows similar results, indicating that the 8–10 Å shell contributes substantial sequence-discontinuous context.

5. Scope of the Analysis and Manuscript Revision. The results explain why 10 Å is a suitable intermediate scale: compared with 8 Å, it captures substantially more local structural context (35.94 vs. 24.21 residues in U50), including additional sequence-discontinuous neighbors. In contrast, larger radii (12–16 Å) substantially expand the neighborhoods (51.12–87.77 residues in U50), potentially introducing excessive structural context. Thus, 10 Å provides a practical balance between capturing spatially adjacent, sequence-discontinuous residues and maintaining compact local regions. We have incorporated these analyses into Appendix A.4 (ProtGround-Unseen: Benchmark Composition and Structural Analysis) of the revised manuscript.

Reviewer_QNWHOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, authors propose ProtGround to address natural-langauge functiona phrases in local 3D structural regions rather than whole protein. They contribute ~3M region-phrase pairs over 376K+ proteins by direct route mapping with residue level features onto AlphaFold or RCSB structures as 10 angstrom micro environments. ProtGround-Hier couples ESM-2 and SapBERT to a geometry-aware graph encoder called GeoGAT. Authors trained GeoGAT with region-phrase contractive loss, Sinkhorn resiude-level optimal transport and multi-level masked modeling.

Critical strengths
  1. ProtGround-Align is the strongest contribution where authors provide ~3M pairs over 376,866 proteins. There are +45% recurring phrases across regions which creates a many to many structure that is valuable.
  2. Table 4 demonstrates the effectiveness of the mechanism proposed in the paper via ablations. Dropping L$_{OT}$ costs more Level-3 Core mAP (0.537→0.4620.537 \to 0.462) than dropping L$_{GC}$ (0.537→0.4870.537 \to 0.487), while L$_{GC}$ matters more for fixed-encoder localization.
Critical weaknesses
  1. Neither the grounding gain nor its attribution to 3D geometry is well supported. Removing L$_{GC}$ results at Level-1 mAP staying at 0.6940.694 against ProTrek 0.4670.467 which suggests in-domain data rather than the method.
  2. The claim where protein-text pretraining stays at whole-protein level is contradicted by the prior work. The ProtCLIP (AAAI 2025) utilizes property spesific text with dynamically selected residue segments. Another work MMSite (NeurIPS 2024) predicts residue-level active sites from sequence and text.
Questions for authors
  1. The results presented in Table 3 is identical to published results of ProtST and ProtCLIP published results while authors claiming PLMs were evalauted, are they re-run or just adopted the results?
  2. What fraction of the DeepLoc, DeepFRI and fitness test proteins occur in ProtGround-Align?
  3. Which ESM-2 and SapBERT checkpoints were used?
  4. What are the candidate pool sizes at Levels 2 and 3 and the mean residue count of the regions defined by 10 angstrom?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Comments

We sincerely thank the reviewer for their valuable and constructive feedback, which we greatly appreciate.

For W1. Grounding Gains and Contributions of 3D Geometry. We believe that our fine-grained training data, together with the model architecture specifically designed to leverage it, jointly contributes to the performance gains and enables state-of-the-art results. Our main contribution is moving beyond whole-protein--text alignment toward spatially localized protein function understanding. In fact, introducing ProtGround-Align, which provides large-scale supervision linking functional phrases to local 3D structural regions, is a central contribution of our work. While existing baselines such as ProTrek are primarily pretrained for global protein--text alignment, ProtGround-Hier is specifically designed to leverage this fine-grained supervision through geometry-aware region encoding, multi-positive region--phrase alignment, and residue--token optimal transport. Thus, we view the dataset and the model as complementary contributions. The ablation results further demonstrate the effectiveness of our pre-training objectives. Removing LGCL_{GC} reduces Level-1 mAP from 0.722 to 0.694 and Level-3 Core mAP from 0.537 to 0.487, while removing LOTL_{OT} lowers these scores to 0.682 and 0.462, respectively. These results highlight the importance of region–phrase alignment and residue–token correspondence for fine-grained grounding.

To directly investigate the contribution of structural encoding, we conducted an additional ablation by replacing our Local 3D Geometry module and Local Region Graph Encoder with the ProTrek structure encoder, while keeping the training data, objectives, and remaining model components unchanged. As shown in Table 1, this replacement reduces Level-1 mAP from 0.722 to 0.677, Level-2 mAP from 0.914 to 0.835, Level-3 Core mAP from 0.537 to 0.450, and SiteRecall@5 from 0.865 to 0.786.

Table 1: Ablation study of pre-training objectives and structural encoding in ProtGround-Hier.

Config.Grounding Level1 mAPGrounding Level2 mAPGrounding Level3 Core mAPGrounding SiteRecall@5Loc. Fixed (Acc.%)Loc. Full (Acc.%)Fit. Fixed (ρ)Fit. Full (ρ)Func. Fixed (Fmax)Func. Full (Fmax)
Full objective0.7220.9140.5370.86589.9089.750.6730.8130.6330.696
w/o LGCL_{GC}0.6940.8680.4870.81987.4488.010.6470.7900.5920.658
w/o LOTL_{OT}0.6820.8410.4620.79787.8688.360.6460.7870.5890.654
w/o LMMPL_{MMP}0.7100.8940.5160.84988.8788.920.6590.7910.6100.666
w/o LMPML_{MPM}0.7030.8850.5080.84088.7288.740.6620.7930.6070.664
w/ Alt. 3D Encoder0.6770.8350.4500.78687.9388.450.6530.7950.5980.657

The difference is closely related to the requirements of fine-grained grounding. Our geometry-aware encoder explicitly models inter-residue distances and spatial connectivity within local 3D regions, allowing the model to represent the geometric relationships surrounding functional residues. In contrast, ProTrek encodes structural information through discrete Foldseek 3Di tokens, which retain local geometric cues but do not explicitly preserve the original continuous inter-residue distances and region-specific graph connectivity. Its trimodal representation is primarily designed for protein-level alignment and retrieval. The larger performance degradation at Levels 2 and 3 suggests that explicit local geometric relationships are particularly beneficial for precise functional region localization.

These results show that, although fine-grained training data is important, the grounding gains cannot be attributed to the data alone. The proposed pre-training objectives and geometry-aware local structural encoder provide additional contributions, supporting the effectiveness of our model design for fine-grained protein function grounding. We have incorporated these new results into the ablation study in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Comments

For W2. Distinction from Prior Fine-Grained Protein Methods.​ We thank the reviewer for highlighting ProtCLIP and MMSite. Although both methods build on whole-protein sequences and protein-associated textual descriptions, neither explicitly supervises the grounding of functional phrases in local 3D structural regions.

Specifically, ProtCLIP combines global protein–text alignment with text-guided segment reconstruction and property-guided dynamic residue selection. While these objectives capture fine-grained sequence information, they do not establish explicitly annotated correspondences between functional phrases and local 3D regions. MMSite performs residue-level active-site prediction using protein sequences and multi-attribute textual descriptions. However, it identifies active-site residues rather than localizing different free-form functional phrases to their corresponding 3D structural regions.

In contrast, ProtGround introduces large-scale, explicitly annotated local 3D region–phrase pairs with many-to-many functional relationships. ProtGround-Hier further learns these relationships through geometry-aware region encoding, multi-positive alignment, and residue–token optimal transport, enabling phrase-conditioned localization of functional regions and their core residues.

Due to space limitations, our original Related Work discussion did not make these distinctions sufficiently clear. We have now revised the Related Work section to explicitly acknowledge the modeling contributions of ProtCLIP and MMSite and clarify their differences from our local 3D structural grounding framework.

For Q1. Source of Baseline Results in Table 3. We thank the reviewer for raising this question. The baseline results in Table 3 of the revised manuscript are adopted from the published results of ProtST and ProtCLIP. These results serve as reference comparisons on established downstream benchmarks. In contrast, all ProtGround-Hier results were obtained from our own experiments under both fixed-encoder learning and full-model tuning. We have now explicitly clarified the sources of the baseline results in Table 3 of the revised manuscript. Furthermore, we conducted five independent runs with different random seeds for ProtGround-Hier across all downstream tasks. Table 3 of the revised manuscript now reports mean ± standard deviation over five seeds, with detailed results and 95% confidence intervals provided in Appendix A.6. The detailed experimental results are presented in the table below.

SettingTask / MetricMean ± SD95% CI
Fixed-encoderBinary localization (Acc.%)94.65 ± 0.09[94.54, 94.76]
Fixed-encoderSubcellular localization (Acc.%)85.15 ± 0.14[84.97, 85.33]
Fixed-encoderFitness (mean ρ)0.6732 ± 0.0033[0.6691, 0.6774]
Fixed-encoderFunction (mean AUPR)0.6292 ± 0.0015[0.6274, 0.6310]
Fixed-encoderFunction (mean Fmax)0.6330 ± 0.0028[0.6296, 0.6364]
Full-modelBinary localization (Acc.%)95.24 ± 0.05[95.18, 95.31]
Full-modelSubcellular localization (Acc.%)84.26 ± 0.39[83.78, 84.74]
Full-modelFitness (mean ρ)0.8131 ± 0.0027[0.8098, 0.8164]
Full-modelFunction (mean AUPR)0.6984 ± 0.0021[0.6957, 0.7010]
Full-modelFunction (mean Fmax)0.6967 ± 0.0044[0.6912, 0.7021]

For Q2. Overlap with Downstream Test Proteins. We explicitly excluded all test proteins from the DeepLoc, DeepFRI, and five fitness prediction benchmarks during the construction of ProtGround-Align, before pretraining. We further conducted a protein-level overlap audit between the final ProtGround-Align dataset and each downstream test set, confirming 0% overlap across all benchmarks. Therefore, none of the downstream test proteins appears in our pretraining dataset, preventing direct test-protein overlap between pretraining and downstream evaluation. We have explicitly clarified this in Line 206 of the revised manuscript: "To prevent test-set leakage, we excluded all downstream test proteins during construction and verified zero protein-level overlap."

For Q3. ESM-2 and SapBERT Checkpoints. We use facebook/esm2_t33_650M_UR50D (650M parameters, 33 Transformer layers, 1,280-dimensional representations) as the protein sequence encoder and cambridgeltl/SapBERT-from-PubMedBERT-fulltext (approximately 110M parameters, 12 Transformer layers, 768-dimensional representations) as the biomedical text encoder. Both encoders are initialized from their publicly available pretrained checkpoints and jointly fine-tuned during ProtGround-Hier pretraining. We have specified these checkpoint identifiers in the revised manuscript to ensure reproducibility.

AuthorsOpenReview ↗

Response to Reviewer Comments

For Q4. Candidate Pool Sizes and 10 Å Neighborhood Statistics.

We thank the reviewer for this question. We have conducted a detailed audit of the candidate pools and local 3D structural regions in ProtGround-Unseen, covering both U50 and U30-hard. The reported statistics are based on the corresponding evaluation subsets after applying sequence-length and structural quality filters.

At Level 2 (Candidate Grounding), each protein is evaluated against its predefined local structural regions. The U50 subset contains 11,494 candidate regions across 1,769 proteins, corresponding to an average of 6.50 candidates per protein. The U30-hard subset contains 5,832 candidate regions across 935 proteins, averaging 6.24 candidates per protein. The mean region sizes are 51.23 and 49.77 residues, respectively, calculated across all predefined candidate regions.

At Level 3 (Blind Grounding), no predefined functional regions are provided. Instead, we construct a 10 Å 3D neighborhood around each canonical residue and rank these residue-centered candidates according to their similarity to the queried functional phrase. This produces 737,481 candidates across 1,770 U50 proteins (416.66 per protein) and 370,661 candidates across 933 U30-hard proteins (397.28 per protein). The mean 10 Å neighborhood sizes are 35.94 and 34.39 residues, respectively.

Candidate-pool composition of ProtGround-Unseen.

StatisticU50U30-hard
Level 2: Predefined Local Regions
Evaluated proteins1,769935
Total candidate regions11,4945,832
Mean candidates per protein6.506.24
Mean residues per region51.2349.77
Level 3: Residue-centered 10 Å Neighborhoods
Evaluated proteins1,770933
Total candidate neighborhoods737,481370,661
Mean candidates per protein416.66397.28
Mean residues per neighborhood35.9434.39

The two levels therefore involve substantially different candidate-pool sizes: approximately 6 predefined regions per protein at Level 2, compared with approximately 400 residue-centered neighborhoods at Level 3. Their evaluated protein counts differ slightly because of their distinct structural eligibility requirements. Level 2 requires valid predefined local structural regions, whereas Level 3 requires complete 3D structures with reliable canonical residue mapping. Moreover, Level 2 regions may encompass multiple functionally related core residues, whereas Level 3 neighborhoods are centered on individual residues. These differences in candidate construction also contribute to the differences in their mean region sizes.

The substantially larger candidate space at Level 3 highlights the increased localization demands of blind grounding.

We have included these updated candidate-pool statistics and detailed structural analyses in Appendix A.4 (ProtGround-Unseen: Benchmark Composition and Structural Analysis) of the revised manuscript.

This concludes our responses to all your comments and questions. We sincerely hope that our clarifications, additional experiments, and manuscript revisions have adequately addressed your concerns. If you find our responses satisfactory and recognize the contributions of our work, we would be grateful if you could consider increasing your rating. Thank you very much for your valuable feedback and thoughtful consideration.

Reviewer_RKxfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces the ProtGround-Align dataset and the ProtGround-Hier model for protein multimodal representation learning. The ProtGround-Align dataset is annotated to establish connections between local protein 3D structural regions and natural-language function descriptions. The ProtGround-Hier incorporates contrastive learning, residue-token optimal transport and multi-level masked modeling. The paper demonstrates strong performance improvement on the ProtGround-Unseen benchmark.

Critical strengths
  1. The paper focuses on biologically meaningful research questions, and pushes the boundary of protein function prediction from “what functions a protein performs” to “which regions of the protein 3D structure are responsible for certain features”.

  2. The produced dataset with millions of local structure function annotations is a valuable contribution both to this paper and to the entire research community.

  3. The paper demonstrates strong performance improvements compared to baselines, without the burden of selecting the candidate functional regions in advance.

Critical weaknesses
  1. The novelty of the paper mainly comes from the new task and data design, and the model components are largely built upon established techniques. The methodological contributions appears relatively incremental.

  2. The comparison on the benchmark is not entirely fair, as ProtGround-Hier uses the newly annotated fine-grained region-function annotation, while baselines use the ordinary global protein-function annotation. This raises concerns on whether the better performance comes from the new training data or the new model architecture.

  3. The semantic-route annotation of ProtGround-Align with SapBERT and GLM-5.3 lacks independent biological quality control. Both the semantic matching and the LLM validator relies on the same original function annotation. If the original annotation is vague or inaccurate in biological details, the second LLM may not be able to identify the problem.

Questions for authors
  1. In section 3 the paper states that the trained model takes a local 3D structure of a protein as input to predict the function. However, in section 4.2.1, the paper also introduces blind grounding. How does the model locate the target local 3D structures?

  2. Some of the performance improvements in Table 3 are marginal, could the authors provide statistical significance tests or confidence intervals?

  3. The three baselines in Table 2 primarily focuses on global protein-text alignment and do not provide sufficiently comprehensive comparisons. Could the authors compare with newly emerged local protein function alignment methods (such as ProLoc [1]) or more recent alternative contrastive learning techniques (such as H-SCALE [2])?

Overall, I find the paper sound and meaningful, with a few doubts that I’ve stated in weaknesses and questions. I look forward to the authors' responses and am open to revisiting my rating.

[1] Liu, P., Fan, J., Pan, M., & Zhang, J. (2026). ProLoc: Text-guided Localization of Protein Functional Regions. bioRxiv, 2026-06.

[2] Fei, X., Chatzianastasis, M., Carneiro, S., Abdine, H., Petalidis, L., & Vazirgiannis, M. (2026). Prot2text-v2: Protein function prediction with multimodal contrastive alignment. Advances in Neural Information Processing Systems, 38, 26205-26232.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Comments

We thank you for recognizing the importance of our research question and dataset. We also appreciate your willingness to reconsider the rating based on our responses.

For W1. Methodological Novelty. Our main contribution is moving beyond whole-protein--text alignment toward spatially localized protein function understanding. We introduce ProtGround-Align, which captures hierarchical and many-to-many protein--region--phrase relationships, and design ProtGround-Hier specifically to model these relationships. First, our geometry-aware region encoder explicitly models inter-residue distances and local 3D spatial connectivity, capturing the structural microenvironments surrounding functional residues. Second, unlike conventional single-positive contrastive learning, our multi-positive region--phrase alignment handles multiple functional phrases per region without treating valid positives as negatives. Third, to our knowledge, ProtGround is the first to use optimal transport to align geometry-aware residues within local 3D regions with functional phrase tokens, enabling fine-grained structural grounding. Together, these designs enable both phrase-conditioned functional region localization and core-residue identification, beyond global protein--text matching.

To further demonstrate the importance of our model design, we conducted an additional structural encoder ablation, replacing our local 3D geometry features and region graph encoder with the ProTrek structure encoder, while keeping the training data, objectives, and remaining model components unchanged. As shown in Table 1, this replacement causes the largest drop across all grounding metrics, with Level-3 Core mAP decreasing from 0.537 to 0.450 and SiteRecall@5 from 0.865 to 0.786. While ProTrek represents structural geometry through discrete 3Di tokens, our encoder explicitly models inter-residue distances and local spatial connectivity, which are particularly important for precise functional grounding.

Table 1: Ablation study of pre-training objectives and structural encoding in ProtGround-Hier.

Config.Level1 mAPLevel2 mAPLevel3 Core mAPSiteRecall@5Loc. Fixed (Acc.%)Loc. Full (Acc.%)Fit. Fixed (ρ\rho)Fit. Full (ρ\rho)Func. Fixed (Fmax)Func. Full (Fmax)
Full objective0.7220.9140.5370.86589.9089.750.6730.8130.6330.696
w/o LGC\mathcal{L}_{\mathrm{GC}}0.6940.8680.4870.81987.4488.010.6470.7900.5920.658
w/o LOT\mathcal{L}_{\mathrm{OT}}0.6820.8410.4620.79787.8688.360.6460.7870.5890.654
w/o LMMP\mathcal{L}_{\mathrm{MMP}}0.7100.8940.5160.84988.8788.920.6590.7910.6100.666
w/o LMPM\mathcal{L}_{\mathrm{MPM}}0.7030.8850.5080.84088.7288.740.6620.7930.6070.664
w/ Alt. 3D Encoder0.6770.8350.4500.78687.9388.450.6530.7950.5980.657

The objective ablations further confirm the contributions of our training design. Removing LGCL_{GC} reduces Level-3 Core mAP from 0.537 to 0.487, while removing LOTL_{OT} lowers it to 0.462, demonstrating the importance of region--phrase alignment and residue--token correspondence. Together with the structural encoder ablation, these results show that ProtGround-Hier's task-specific architecture and training objectives provide substantial benefits beyond the fine-grained dataset itself. We have incorporated the new structural encoder results into the ablation study (Table 4) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Comments

For W2. Contributions of Training Data and Model Architecture. The performance gains arise from the combination of our fine-grained training data and task-specific model design. ProtGround-Align provides explicitly annotated local 3D region--phrase supervision, while ProtGround-Hier is specifically designed to learn these fine-grained relationships. Existing baselines, including ProteinCLAP, ProtCLIP, and ProTrek are designed for global protein-text alignment and cannot directly utilize our fine-grained region-function annotations without modifying their training objectives.

Importantly, we further investigated the contribution of model architecture under the same fine-grained training data. As detailed in our response to W1, the new structural encoder ablation in Table 4 of the revised manuscript replaces our geometry-aware region encoder with ProTrek's structure encoder while keeping the training data, objectives, and remaining components unchanged. This replacement reduces performance across all grounding metrics, demonstrating that our task-specific local 3D structural encoding provides additional benefits beyond the training data itself. The objective ablations further confirm the contributions of region--phrase alignment and residue--token correspondence.

Thus, Tables 2 and 3 of the revised manuscript demonstrate the overall benefits of our fine-grained pretraining framework for grounding and downstream protein understanding, respectively, while Table 4 provides controlled evidence for the effectiveness of our model design.

For W3. Independent Biological Validation. We clarify that semantic matching and LLM validation do not rely on the same evidence alone. SapBERT matches candidate functional phrases to local regions based on their original functional annotations. In contrast, our GLM-5.3 validator operates independently of the phrase-generation process and additionally cross-checks each matched pair against protein-specific biological evidence from UniProt, M-CSA, Q-BioLiP, LIGYSIS, and relevant scientific literature. This additional evidence helps identify potential inconsistencies or unsupported functional claims beyond the original annotations. Ambiguous, unsupported, or functionally inconsistent pairs are rejected, providing a second level of quality control. We have clarified this independent validation process in Lines 196--202 of the revised manuscript.

For Q1. Blind Grounding. We clarify that ProtGround-Hier can localize functional regions without predefined candidate regions. Although the model encodes local 3D regions, their functional locations need not be known in advance. Given a protein sequence, its full 3D structure, and a functional phrase, we treat every residue as a potential center and construct its 10 Å spatial neighborhood using minimum interatomic distances. Each local region is then encoded using sequence and geometric features and ranked by cosine similarity to the functional phrase embedding. Core mAP evaluates the ranking of residue centers, while SiteRecall@5 measures whether the top five predicted regions recover the true functional sites. Importantly, no ground-truth functional annotations or site coordinates are used during candidate generation or inference; they are used only for evaluation. We have clarified this process in Lines 359--370 and 418--425 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Comments

For Q2. Statistical Significance and Confidence Intervals. We conducted five independent runs with different random seeds across all downstream tasks under both fixed-encoder learning and full-model tuning. Table 3 of the revised manuscript reports mean ± standard deviation, while Appendix A.6 provides detailed results and 95% confidence intervals for each task. The key statistical results are summarized below.

Table 2: Summary of five-run downstream results and 95% confidence intervals for ProtGround-Hier.

SettingTask / MetricMean ± SD95% CI
Fixed-encoderBinary localization (Acc.%)94.65 ± 0.09[94.54, 94.76]
Fixed-encoderSubcellular localization (Acc.%)85.15 ± 0.14[84.97, 85.33]
Fixed-encoderFitness (mean ρ)0.6732 ± 0.0033[0.6691, 0.6774]
Fixed-encoderFunction (mean AUPR)0.6292 ± 0.0015[0.6274, 0.6310]
Fixed-encoderFunction (mean Fmax)0.6330 ± 0.0028[0.6296, 0.6364]
Full-modelBinary localization (Acc.%)95.24 ± 0.05[95.18, 95.31]
Full-modelSubcellular localization (Acc.%)84.26 ± 0.39[83.78, 84.74]
Full-modelFitness (mean ρ)0.8131 ± 0.0027[0.8098, 0.8164]
Full-modelFunction (mean AUPR)0.6984 ± 0.0021[0.6957, 0.7010]
Full-modelFunction (mean Fmax)0.6967 ± 0.0044[0.6912, 0.7021]

Under fixed-encoder learning, ProtGround-Hier achieves higher mean performance than the published ProtCLIP results across all 15 metrics. In particular, mean fitness Spearman's ρ improves from 0.6154 to 0.6732, while mean function annotation Fmax increases from 0.6093 to 0.6330. Under full-model tuning, ProtGround-Hier outperforms the reported ProtCLIP results on 11 of 15 metrics. Mean fitness Spearman's ρ improves from 0.7932 to 0.8131, and mean function annotation Fmax increases from 0.6785 to 0.6967.

The narrow confidence intervals and low standard deviations indicate stable performance across random seeds for most downstream metrics. While some individual improvements are modest, the results demonstrate consistent overall performance gains across multiple downstream tasks and evaluation settings. We have updated Table 3 of the revised manuscript to report mean ± standard deviation over five seeds and added detailed results, 95% confidence intervals, and variability analysis in Appendix A.6 (Downstream Training and Evaluation Result).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.