BioBlobs: Unsupervised Discovery of Functional Substructures

arXiv:2510.01632 · q-bio.BM, cs.AI · Submitted 2025-10-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "BioBlobs: Unsupervised Discovery of Functional Substructures".

Marcus: Here's the gist of the paper. Protein function is driven by cohesive substructures, such as catalytic triads, binding pockets, and structural motifs, that occupy only a small fraction of a protein’s residues Across diverse protein function prediction tasks and multiple sequence- and structure-based encoders, BIOBLOBS matches or exceeds strong baselines while operating on only a small fraction of residues.

Ines: The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains.

Marcus: Pruning redundant reasoning tokens BIOBLOBS is an encoder-agnostic, end-to-end differentiable framework that compresses a protein into a small set of cohesive substructures (blobs) and predicts function from these blobs alone.

Ines: This approach addresses the central biological question of which substructure of a protein is responsible for its function.

Title and authors: Ines: So this paper introduces something called BioBlobs: Unsupervised Discovery of Functional Substructures. It basically suggests that we don't need to look at every single residue in a protein when figuring out what it does. Instead, function comes from these small, cohesive chunks, or blobs >

Marcus: Exactly. The core idea is compressing the whole protein into a small set of these blobs and then predicting the function based only on those blobs >

Ines: Right. They're saying that existing methods just give you per-residue embeddings, but those embeddings get diluted by all the surrounding non-functional stuff, so you lose sight of which specific part is actually doing the work >

Yuki: From a population genetics view, it makes sense because if we only look at one residue in isolation, that single measurement doesn't tell us much about the whole organism’s trait; we need to see how those functional units are organized across the structure >

Ines: The paper shows that these discovered blobs adapt their size based on what task you’re trying to do. They can be tiny local sites or they can span entire structural domains depending on what's needed >

Marcus: And that adaptation is key because it solves the problem of needing different sized regions for different biological functions, like a catalytic triad versus a whole binding pocket >

Yuki: That adaptability is interesting because it means this framework isn't locked into one type of biological mechanism; it can handle local things and big structural features depending on the prediction task >

Ines: They also show that you can use protein-level supervision—just knowing if the whole protein has a certain function—to find experimentally known catalytic sites without needing residue-level labels >

Marcus: That's a big deal for scaling up functional site discovery; it means we don't have to wait for expensive, residue-level data to even start looking for these important regions >

Yuki: If you can recover known catalytic sites just from the protein level label, that opens up a way to look at the entire unannotated proteome for functional clues >

Ines: The paper also discusses how they build their training signal, which involves two steps: seed selection and then expanding those seeds into soft, spatially local substructures using proximity bias >

Marcus: That expansion process uses both semantic compatibility and a distance-based proximity prior to decide if a residue belongs to a blob >

Yuki: I wonder how they manage the trade-off between finding enough blobs and not just making every small piece its own blob, given that the functional substructures are so small relative to the whole protein >

Ines: They use something called Hoyer’s scale-invariant sparseness measure to prevent the partitioner from absorbing too many candidates in each seed's neighborhood, which forces them to find a minimal set of residues >

Title and authors: Marcus: That regularization is what keeps the model from just being overly general and ensures it finds those actual functional spots rather than just random local clusters >

Yuki: So, when you look at the results on the M-CSA database, they scored catalytic residues using this method with a median accuracy of zero point nine two five against binary labels > <ref:2510.01632#pg1>

Ines: That comparison against a shuffled partition gives them zero point five zero one, and against an SASA baseline which assumes catalytic residues are buried gives them zero point six nine six > <ref:2510.01632#pg3>

Marcus: Those numbers suggest that this method is doing something substantial beyond just random chance or the existing simple baselines we're using right now >

Yuki: It’s interesting because they achieved this recovery using only protein-level supervision, which is a significant constraint for any biological discovery tool >

Ines: The paper also highlights that if you use attention-weighted alignment loss, you can push the partitioner and the prediction head to agree on which blobs are actually relevant for the final function prediction >

Marcus: That makes the blobs first-class units of computation, meaning their contribution is directly tied to whether or not they are used in the final result >

Yuki: So, if we think about this finding for biology, it means we can start mapping out functional relationships across massive datasets without needing perfect residue annotations upfront >

Ines: This whole framework is essentially reframing protein function prediction from analyzing a huge pool of residue embeddings to reasoning over a compact set of cohesive substructures >

Marcus: And the result is that they are matching or exceeding strong baselines on both ProteinShake and VenusX encoders while only looking at a small fraction of residues >

Yuki: It seems like this opens up a scalable path for discovering functional substructures across the entire unannotated proteome, which is huge for understanding evolution >

Ines: Overall, the main implication is that you can discover functionally relevant substructures without per-residue supervision, and that those blobs dynamically adjust their scale to match the specific biological task at hand >

Marcus: And they’ve shown that aggressive Hoyer regularization helps keep these blobs compact without losing accuracy, which makes them very efficient models computationally >

Yuki: It’s a solid result because it proves that functional information is indeed localized in these substructures, and this discovery process is robust across different types of protein encoders >

Ines: So to wrap up, the paper BioBlobs: Unsupervised Discovery of Functional Substructures gives us a way to predict function by reasoning over these small cohesive blobs that scale appropriately for the task >

Marcus: It’s powerful because it recovers known sites from just protein-level data and shows that these blobs are the true units of computation here >

Yuki: It really suggests that functional organization in proteins is hierarchical, with these smaller substructures building up to define a larger function across different scales >

The paper's summary: Ines: So we just talked about how these new blobs can adapt their size based on what you're trying to do, right? Now let's look at what this paper actually claims in its summary >

Marcus: Okay, so it boils down to this idea that protein function isn't driven by every single residue. instead, it’s driven by these small chunks called blobs that have a specific function >

Ines: Exactly. The paper says they compressed the whole protein into just a few of these blobs and predicted the function from those blobs alone >

Marcus: And this is important because it means you don't need to look at every single residue embedding anymore; you just need to focus on the functional regions >

Ines: They show that these discovered blobs adapt their spatial scale to the task, meaning they can be tiny local catalytic sites or they can be whole structural domains depending on what the prediction goal is >

Marcus: That adaptability is a huge deal because it solves a problem where you need different sized regions for different biological functions >

Ines: They also show that you can use only protein-level supervision—just knowing if the whole protein has a certain function—to find experimentally known catalytic sites without needing residue-level labels >

Marcus: That’s the big implication for scaling up discovery; it means we don't have to wait for expensive, residue-level data to even start looking for these important regions across the unannotated proteome >

Ines: The paper also details how they build their training signal, which involves two steps—seed selection and then expanding those seeds into soft, spatially local substructures using proximity bias >

Marcus: That expansion process uses both learned compatibility between the seed and a residue and a distance-based proximity prior to decide if that residue belongs to a blob >

Ines: They use something called Hoyer’s scale-invariant sparseness measure to stop the partitioner from absorbing too many candidates in each seed's neighborhood, forcing them to find just the minimal set of residues needed >

Marcus: That regularization is what keeps the model from being overly general and ensures it actually finds those specific functional spots rather than just random local clusters >

Ines: When they look at the results on the M-CSA database, they scored catalytic residues using this method with a median accuracy of zero point nine two five against binary labels >

Marcus: That comparison against a shuffled partition gives them zero point five zero one, and against an SASA baseline that assumes catalytic residues are buried gives them zero point six nine six >

Ines: Those numbers suggest the method is doing something substantial beyond just random chance or the existing simple baselines we're using right now >

Marcus: It’s interesting because they achieved this recovery using only protein-level supervision, which is a significant constraint for any biological discovery tool >

Ines: The paper also highlights that if you use attention-weighted alignment loss, you can push the partitioner and the prediction head to agree on which blobs are actually relevant for the final function prediction >

Marcus: That makes those blobs first-class units of computation, meaning their contribution is directly tied to whether or not they are used in the final result >

Ines: So, if we think about this finding for biology, it means we can start mapping out functional relationships across massive datasets without needing perfect residue annotations upfront >

Marcus: This whole framework is essentially reframing protein function prediction from analyzing a huge pool of residue embeddings to reasoning over a compact set of cohesive substructures >

Ines: And the result is that they are matching or exceeding strong baselines on both ProteinShake and VenusX encoders while only looking at a small fraction of residues >

Marcus: It seems like this opens up a scalable path for discovering functional substructures across the entire unannotated proteome, which is huge for understanding evolution >

Ines: Overall, the main implication is that you can discover functionally relevant substructures without per-residue supervision, and that those blobs dynamically adjust their scale to match the specific biological task at hand >

The paper's improvements: Tom: So we've seen how BioBlobs use protein blobs to predict function from small subsets of residues, and now let's talk about what the authors suggest they can do even better >

Ines: The paper points out that they can recover known experimental sites just from a protein-level label, which is a big deal for finding functional spots across the whole proteome >

Marcus: That recovery happens because the attention weight in their model is totally tied to the prediction, so catalytic residues naturally fall into those high-attention blobs >

Ines: They also showed that you can use attention weighted alignment loss, which makes the blobs first-class units of computation, so their contribution is directly tied to whether they are used in the final result >

Marcus: That means the partitioner and the prediction head are jointly pushed to agree on which blobs are fragment resident, which is a pretty tight coupling for learning >

Ines: They also found that aggressive Hoyer regularization compresses each blob size without sacrificing accuracy, leading to what they call near-free compactness >

Marcus: So they can keep the model accurate while making the discovered functional regions as small and efficient as possible computationally >

Ines: The authors note that they have shown this works across different types of protein encoders, which suggests this approach isn't tied to one specific way of representing a protein structure >

Marcus: That's key because if it works across different encoders, it means the functional substructures themselves are robust concepts that exist regardless of how you model them computationally >

Ines: But they also flag a limitation: the paper doesn't tell us exactly how to tune those scale parameters for every single protein type without some kind of task-specific guidance >

Marcus: That’s fair; they don't give a universal setting for the radius or the number of seeds, which means you still have to do some testing for each specific prediction task >

Ines: So what this changes for someone just listening is that we can start searching massive databases for functional clues using only protein-level info, even when we don't know where to start with the blob sizes >

Conclusion: Tom: So we’re wrapping up our take on BioBlobs: Unsupervised Discovery of Functional Substructures, and it really boils down to this idea that protein function is driven by these small, cohesive substructures called blobs >

Ines: Exactly. The paper reframes protein function prediction from looking at all the residue embeddings to reasoning over a compact set of functional units >

Marcus: And because these blobs adapt their size based on the task, they can find both local catalytic sites and entire structural domains, which is a big flexibility gain >

Yuki: From a population genetics view, this suggests that function isn't always defined by one single residue but by how those functional units are organized across the whole structure >

Ines: They proved that you can recover known experimental catalytic sites from just protein-level supervision, which is a significant step for finding functional sites in the unannotated proteome >

Marcus: That's powerful because it means we don't need to wait for residue-level data before we can even start looking for these important regions >

Yuki: It opens up a way to look at the entire unannotated proteome for functional clues, which is huge for understanding evolution and how functions might have been gained or lost over time >

Ines: The whole point is that these blobs are first-class units of computation because their attention weights are inseparable from the prediction, which makes them really useful tools >

Marcus: We saw the numbers on M-CSA were strong, median accuracy zero point nine two five against binary labels compared to a baseline of zero point five zero one >

Yuki: That level of recovery suggests that the biological signal is robust enough to be found even when you only have broad structural information >

Ines: The main implication is that you can discover functionally relevant substructures without per-residue supervision, and those blobs dynamically adjust their scale to match the specific biological task at hand >

Marcus: They also showed that aggressive Hoyer regularization helps keep these blobs compact without losing accuracy, which makes them very efficient models computationally >

Yuki: It’s a solid result because it proves that functional information is indeed localized in these substructures, and this discovery process is robust across different types of protein encoders >

Ines: So we have the BioBlobs framework, which uses compact substructures to predict function from protein-level data alone >

Marcus: It’s a method that moves us toward reasoning over functional units rather than just residue embeddings, and that's a shift in how we model these systems >

Yuki: This suggests that functional organization in proteins is hierarchical, with these smaller substructures building up to define a larger function across different scales >

Xin Wang, Kaiwen Shi, Carlos Oliver

Vanderbilt University · Yale University

q-bio.BM, cs.AI

Submitted: 2025-10-02

Updated: 2026-11-06

Comments: Accepted at the 40th Conference on Neural Information Processing Systems (NeurIPS 2026). 25 pages, 5 figures, 11 tables. Code: https://github.com/OliverLaboratory/BioBlobs

Code: https://github.com/OliverLaboratory/BioBlobs

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: Protein function is driven by cohesive substructures, such as catalytic triads, binding pockets, and structural motifs…

Key concepts

Blobs
These are small, cohesive substructures within a protein that are hypothesized to be responsible for its specific function. Instead of analyzing the entire protein sequence, BIOBLOBS focuses on these small regions, which can range in size from just a few residues to large structural domains.
Neural Blob Partitioner
This module proposes and defines the location and properties of these functional blobs. It works by selecting initial 'seed' residues and then growing them into soft regions using learned compatibility scores and distance measures, determining membership based on both semantic similarity and spatial proximity.
Hoyer’s Scale-Invariant Sparseness Measure
This is a regularization technique used during training to ensure the partitioner doesn't select too many candidates in any area. It forces the system to discover a minimal, sufficient set of residues for function prediction, promoting efficient and accurate functional substructure discovery.
MIL Decoder
This module takes the evidence from all discovered blobs and aggregates it to make a final prediction about the entire protein's function. It treats each blob as an instance of data, learning how to combine their individual contributions into a single, robust protein-level output.

Terminology

Summary

Protein function is driven by cohesive substructures, such as catalytic triads, binding pockets, and structural motifs, that occupy only a small fraction of a protein’s residues Across diverse protein function prediction tasks and multiple sequence- and structure-based encoders, BIOBLOBS matches or exceeds strong baselines while operating on only a small fraction of residues. p. 0

Where the paper says it

We introduce BIOBLOBS, an encoder-agnostic, end-to-end differentiable framework that compresses a protein into a small set of cohesive substructures (blobs) and predicts function from these blobs alone, so that each blob corresponds to a candidate functional region. Across diverse protein function prediction tasks and multiple sequence- and structure-based encoders, BIOBLOBS matches or exceeds strong baselines while operating on only a small fraction of residues.

The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains. p. 0
Where the paper says it

Across diverse protein function prediction tasks and multiple sequence- and structure-based encoders, BIOBLOBS matches or exceeds strong baselines while operating on only a small fraction of residues. The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains.

Pruning redundant reasoning tokens

BIOBLOBS is an encoder-agnostic, end-to-end differentiable framework that compresses a protein into a small set of cohesive substructures (blobs) and predicts function from these blobs alone. This approach addresses the central biological question of which substructure of a protein is responsible for its function. p. 0

Where the paper says it

BIOBLOBS: Unsupervised Discovery of Functional Substructures for Protein Function Prediction Xin Wang1,2 , Kaiwen Shi1 , Carlos Oliver1 1Vanderbilt University, Nashville, Tennessee, USA 2Yale University, New Haven, Connecticut, USA Abstract Protein function is driven by cohesive substructures, such as catalytic triads, binding pockets, and structural motifs, that occupy only a small fraction of a protein’s residues. Yet existing pipelines built on protein encoders do not model proteins at the substructure level, leaving the central biological question unanswered: which substructure of a protein is responsible for its function?

The framework consists of three modules: (1) a protein encoder that produces residue-level embeddings, (2) a neural blob partitioner that proposes blob regions and computes blob embeddings, and (3) a multiple instance learning (MIL) decoder that aggregates blob-level evidence into a protein-level prediction. p. 1
Where the paper says it

2 BIOBLOBS The BIOBLOBS framework dynamically partitions a protein into a set of cohesive substructures, termed blobs, and classifies the protein through attention-based aggregation over them. The pipeline consists of three modules: (1) a protein encoder that produces residue-level embeddings, (2) a neural blob partitioner that proposes blob regions and computes blob embeddings, and (3) a multipleinstance learning (MIL) decoder that aggregates blob-level evidence into a protein-level prediction.

How the training signal is built

The Neural Blob Partitioner discovers K local substructures (blobs) in two steps, involving Seed selection which identifies anchor residues through a learned, differentiable sampling procedure and Blob expansion which grows each seed into a soft, spatially local substructure by combining learned semantic compatibility with a distance-based proximity prior. Membership is determined by two complementary signals: a learned attention score aki that captures semantic compatibility between the seed and the residue, and a proximity bias ρki that decays linearly with Euclidean distance. p. 2

Where the paper says it

We form a candidate set from all residues within radius r of the seed’s Cα coordinate: C (k) = n i : ∥xi − x (k) seed∥2 ≤ r o . (2) For each candidate i ∈ C(k) , membership is determined by two complementary signals: a learned attention score aki that captures semantic compatibility between the seed and the residue, and a proximity bias ρki that decays linearly with Euclidean distance, favoring residues close to the seed: aki = (Wq z (k) seed) ⊤(Wkey zi) √ da , ρki = max 0, 1 − ∥xi − x (k) seed∥2 r !

The combined logit is passed through a sigmoid to yield a soft membership score mki = σ(lki/τmem), where lki = aki + λρ ρki. p. 2
Where the paper says it

, (3) where Wq,Wkey ∈ R da×D are learnable projections. The combined logit is passed through a sigmoid to yield a soft membership score: mki = σ(ℓki/τmem), ℓki = aki + λρ ρki, i ∈ C(k) , (4) where λρ weights the proximity bias and τmem is annealed on the same schedule as τseed.

Where the gains come from

The training objective includes both task loss and regularization, specifically Hoyer’s scale-invariant sparseness measure [Hoyer, 2004], which is defined as HSk = 1/nk mk1 / mk2 squared in h 1 nk. This regularization prevents the partitioner from absorbing all candidates in each seed’s spherical neighborhood, ensuring the discovery of a minimal set of residues sufficient for function prediction. p. 3

Where the paper says it

4 Training Objectives If we trained with only the task loss (e. , cross-entropy for multi-class classification), the partitioner tends to absorb all candidates in each seed’s spherical neighborhood, contradicting our goal of identifying a minimal set of residues sufficient for function prediction.

The attention-weighted alignment loss Lattn-align supervises the membership matrix M directly, weighting each blob’s contribution by its MIL attention weight αk, which jointly pushes the partitioner and the MIL attention head to agree on which blobs are fragment-resident. p. 5
Where the paper says it

BIOBLOBS (+attn align) instead supervises the membership matrix directly, weighting each blob’s contribution by its MIL attention weight. The loss therefore only constrains the blobs the classifier actually uses for prediction, and gradients flow into both the partitioner and the MIL attention head, jointly pushing them to agree on which blobs are fragmentresident.

Recovery of known functional sites without residue-level supervision

BIOBLOBS recovers experimentally annotated catalytic sites in the M-CSA database from protein-level supervision alone, demonstrating unsupervised functional substructure discovery. The MIL attention weight is inseparable from the prediction, and catalytic residues fall out as the residues inside high-attention blobs. p. 7

Where the paper says it

DeepFRI [Gligorijevic et al. To our knowledge, ´ BIOBLOBS is the first method to recover experimentally-curated catalytic sites by directly mining over learned substructures: blobs are first-class units of the computation, the MIL attention weight is inseparable from the prediction, and catalytic residues fall out as the residues inside high-attention blobs.

For example, scoring residues by their attention-weighted blob membership Pk αkmki gives a median per-protein AUROC of 0. p. 8
Where the paper says it

We then test the claim on 868 single-chain enzymes from the M-CSA database [Ribeiro et al. Scoring residues by their attention-weighted blob membership P k αkmki gives a median per-protein AUROC of 0.

925 against the binary catalytic label vs. p. 8
Where the paper says it

Scoring residues by their attention-weighted blob membership P k αkmki gives a median per-protein AUROC of 0. 925 against the binary catalytic label, vs.

0.501 under shuffled partition and 0. p. 8
Where the paper says it

925 against the binary catalytic label, vs. 501 under partition shuffling and 0.

696 for an SASA “catalytic residues are buried” baseline. p. 8
Where the paper says it

501 under partition shuffling and 0. 696 for an SASA “catalytic residues are buried” baseline.

Adaptation to natural granularity

The framework demonstrates that the optimal spatial scale of functional substructures varies by task, from local catalytic sites to entire structural domains, with BIOBLOBS adapting accordingly. Exhaustive grid search over K and r shows that for EC-L3, the best test metric is achieved at K = 20, r = 8 where blobs occupy only ∼ 40% of the protein. p. 6

Where the paper says it

Accordingly, BIOBLOBS reaches its best test metric of 0. 615 at K = 20, r = 8, where blobs occupy only ∼ 40% of the protein, confirming that a compact subset of residues suffices.

Furthermore, aggressive Hoyer regularization compresses the size of each blob without sacrificing accuracy, exhibiting near-free compactness. p. 6
Where the paper says it

02 above the unregularized regime. BIOBLOBS therefore exhibits near-free compactness: aggressive Hoyer regularization compresses the size of each blob without sacrificing accuracy.

The radius r is essentially free because it enters only as a sparsity mask in the attention. p. 6
Where the paper says it

46× its peak GPU memory. The blob radius r is essentially free: both time and memory remain flat across r ∈ {6, 8, 12, 16, 20} since r enters only as a sparsity mask in the attention.

Alignment with known functional units

BIOBLOBS (+attn align) achieves the highest Site soft recall (SR) in every cell, lifting SR by 1.6–2.7× over both BIOBLOBS and BIOBLOBS (+seed align), at no more than 0. p. 4

Where the paper says it

The VenusX benchmark provides per-residue functional-site annotations, which allow us to design two BIOBLOBS variants whose partitioner is trained with a fragment-alignment loss. BIOBLOBS (+seed align) adds a loss on the seed scorer’s logits: after softmaxing the logits over 5

05 absolute F1 cost. p. 5
Where the paper says it

7× over both BIOBLOBS and BIOBLOBS (+seed align), at no more than 0. 05 absolute F1 cost.

The loss only constrains the blobs the classifier actually uses for prediction, and gradients flow into both the partitioner and the MIL attention head, jointly pushing them to agree on which blobs are fragment-resident. p. 5
Where the paper says it

BIOBLOBS (+attn align) instead supervises the membership matrix directly, weighting each blob’s contribution by its MIL attention weight. The loss therefore only constrains the blobs the classifier actually uses for prediction, and gradients flow into both the partitioner and the MIL attention head, jointly pushing them to agree on which blobs are fragmentresident.

This makes BIOBLOBS first-class units of computation whose attention weights are inseparable from the prediction. p. 8
Where the paper says it

The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains, without any per-residue supervision. Because blobs are first-class units of the computation whose attention weights are inseparable from the prediction, BIOBLOBS recovers experimentally annotated M-CSA catalytic sites from protein-level labels alone, opening a scalable path to functional substructure discovery across the unannotated proteome.

Conclusion

BIOBLOBS reframes protein function prediction as reasoning over a compact set of cohesive substructures rather than pooled residue embeddings, matching or exceeding strong baselines across ProteinShake and VenusX on multiple sequence- and structure-based encoders while committing to only a small fraction of residues. The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains, without any per-residue supervision. p. 8

Where the paper says it

7 Conclusion BIOBLOBS reframes protein function prediction as reasoning over a compact set of cohesive substructures rather than pooled residue embeddings, matching or exceeding strong baselines across ProteinShake and VenusX on multiple sequence- and structure-based encoders while committing to only a small fraction of residues. The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains, without any per-residue supervision.

Because blobs are first-class units of the computation whose attention weights are inseparable from the prediction, BIOBLOBS recovers experimentally annotated M-CSA catalytic sites from protein-level labels alone, opening a scalable path to functional substructure discovery across the unannotated proteome. p. 8
Where the paper says it

The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains, without any per-residue supervision. Because blobs are first-class units of the computation whose attention weights are inseparable from the prediction, BIOBLOBS recovers experimentally annotated M-CSA catalytic sites from protein-level labels alone, opening a scalable path to functional substructure discovery across the unannotated proteome.

References

Emily Alsentzer, Samuel Finlayson, Michelle Li, and Marinka Zitnik. Subgraph neural networks. p. 9

Where the paper says it

References Emily Alsentzer, Samuel Finlayson, Michelle Li, and Marinka Zitnik. Subgraph neural networks.

Advances in Neural Information Processing Systems, 33:8017–8029, 2020

Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. p. 9

Where the paper says it

Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation.

arXiv preprint arXiv:1308. p. 9
Where the paper says it

Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.

3432, 2013

Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin. Peer Bork. Alan Bridge. p. 9

Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Lucy Colwell. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Julian Gough. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Daniel H Haft. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Ivica Letunic. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Felipe Llinares-López. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Aron Marchler-Bauer. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Laetitia Meng-Papaxanthos; Huaiyu Mi; Darren A Natale; Christine A Orengo; Arun P Pandurangan; Damiano Piovesan; Catherine Rivoire; Christian J A Sigrist. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Narmada Thanki. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Françoise Thibaud-Nissen. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Paul D Thomas. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Silvio C E Tosatto. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

Cathy H Wu and Alex Bateman. p. 9
Where the paper says it

3432, 2013. Matthias Blum, Antonina Andreeva, Laise Cavalcanti Florentino, Sara Rocio Chuguransky, Tiago Grego, Emma Hobbs, Beatriz Lazaro Pinto, Ailsa Orr, Typhaine Paysan-Lafosse, Irina Ponamareva, Gustavo A Salazar, Nicola Bordin, Peer Bork, Alan Bridge, Lucy Colwell, Julian Gough, Daniel H Haft, Ivica Letunic, Felipe Llinares-López, Aron Marchler-Bauer, Laetitia Meng-Papaxanthos, Huaiyu Mi, Darren A Natale, Christine A Orengo, Arun P Pandurangan, Damiano Piovesan, Catherine Rivoire, Christian J A Sigrist, Narmada Thanki, Françoise Thibaud-Nissen, Paul D Thomas, Silvio C E Tosatto, Cathy H Wu, and Alex Bateman.

InterPro: the protein sequence classification resource in 2025

Giorgos Bouritsas, Andreas Loukas, Nikolaos Karalias, and Michael Bronstein. Partition and code: learning how to compress graphs. p. 9

Where the paper says it

Giorgos Bouritsas, Andreas Loukas, Nikolaos Karalias, and Michael Bronstein. Partition and code: learning how to compress graphs.

Advances in Neural Information Processing Systems, 34:18603–18619, 2021

Dexiong Chen, Philip Hartout, Paolo Pellizzoni, Carlos Oliver, and Karsten Borgwardt. Endowing protein language models with structural knowledge. p. 9

Where the paper says it

Dexiong Chen, Philip Hartout, Paolo Pellizzoni, Carlos Oliver, and Karsten Borgwardt. Endowing protein language models with structural knowledge.

arXiv preprint arXiv:2401. p. 9
Where the paper says it

Endowing protein language models with structural knowledge. arXiv preprint arXiv:2401.

14819, 2024

Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. p. 9

Where the paper says it

Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples.

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

Justas Dauparas, Ivan Anishchenko, Nathaniel Bennett, Hua Bai, Robert J Ragotte, Lukas F Milles, Basile IM Wicky, Alexis Courbet, Rob J de Haas. Robust deep learning–based protein sequence design using proteinmpnn. p. 9

Where the paper says it

Justas Dauparas, Ivan Anishchenko, Nathaniel Bennett, Hua Bai, Robert J Ragotte, Lukas F Milles, Basile IM Wicky, Alexis Courbet, Rob J de Haas, Neville Bethel, et al. Robust deep learning–based protein sequence design using proteinmpnn.

Science, 378(6615):49–56

Saeedeh Davoudi, Christopher S Henry, Christopher S Miller, and Farnoush Banaei-Kashani. ECBench: a benchmark for enzyme commission number prediction. p. 9

Where the paper says it

Saeedeh Davoudi, Christopher S Henry, Christopher S Miller, and Farnoush Banaei-Kashani. ECBench: a benchmark for enzyme commission number prediction.

Bioinformatics Advances, 6(1):vbag004, January 2026

Alexander Derry, Alp Tartici, and Russ B Altman. Protein functional site annotation using local structure embeddings. p. 9

Where the paper says it

Alexander Derry, Alp Tartici, and Russ B Altman. Protein functional site annotation using local structure embeddings.

Proceedings of the National Academy of Sciences, 122(34):e2513219122, 2025

G. Dodson and A. p. 1

Where the paper says it

NTPase sites are paradigmatic examples [Dodson and Wlodawer, 1998, Rossmann et al. , 1974, Leipe et al.

Wlodawer. Catalytic triads and their relatives. p. 9
Where the paper says it

Dodson and A. Catalytic triads and their relatives.

Trends in Biochemical Sciences, 23(9):347–352, 1998

Tomas Feder, Pavol Hell, Sulamita Klein, and Rajeev Motwani. Complexity of graph partition problems. p. 9

Where the paper says it

Tomas Feder, Pavol Hell, Sulamita Klein, and Rajeev Motwani. Complexity of graph partition problems.

In Proceedings of the thirty-first annual ACM symposium on Theory of computing, pages 464–472, 1999

Vladimir Gligorijevic, P Douglas Renfrew, Tomasz Kosciolek, Julia Koehler Leman, Daniel Beren-berg. Tommi Vatanen. p. 9

Where the paper says it

In Proceedings of the thirty-first annual ACM symposium on Theory of computing, pages 464–472, 1999. Vladimir Gligorijevic, P Douglas Renfrew, Tomasz Kosciolek, Julia Koehler Leman, Daniel Beren- ´ berg, Tommi Vatanen, Chris Chandler, Bryn C Taylor, Ian M Fisk, Hera Vlamakis, et al.

Chris Chandler. p. 9
Where the paper says it

In Proceedings of the thirty-first annual ACM symposium on Theory of computing, pages 464–472, 1999. Vladimir Gligorijevic, P Douglas Renfrew, Tomasz Kosciolek, Julia Koehler Leman, Daniel Beren- ´ berg, Tommi Vatanen, Chris Chandler, Bryn C Taylor, Ian M Fisk, Hera Vlamakis, et al.

Bryn C Taylor. p. 9
Where the paper says it

In Proceedings of the thirty-first annual ACM symposium on Theory of computing, pages 464–472, 1999. Vladimir Gligorijevic, P Douglas Renfrew, Tomasz Kosciolek, Julia Koehler Leman, Daniel Beren- ´ berg, Tommi Vatanen, Chris Chandler, Bryn C Taylor, Ian M Fisk, Hera Vlamakis, et al.

Ian M Fisk. p. 9
Where the paper says it

In Proceedings of the thirty-first annual ACM symposium on Theory of computing, pages 464–472, 1999. Vladimir Gligorijevic, P Douglas Renfrew, Tomasz Kosciolek, Julia Koehler Leman, Daniel Beren- ´ berg, Tommi Vatanen, Chris Chandler, Bryn C Taylor, Ian M Fisk, Hera Vlamakis, et al.

Hera Vlamakis; et al. p. 9
Where the paper says it

In Proceedings of the thirty-first annual ACM symposium on Theory of computing, pages 464–472, 1999. Vladimir Gligorijevic, P Douglas Renfrew, Tomasz Kosciolek, Julia Koehler Leman, Daniel Beren- ´ berg, Tommi Vatanen, Chris Chandler, Bryn C Taylor, Ian M Fisk, Hera Vlamakis, et al.

Structurebased protein function prediction using graph convolutional networks. p. 9
Where the paper says it

Vladimir Gligorijevic, P Douglas Renfrew, Tomasz Kosciolek, Julia Koehler Leman, Daniel Beren- ´ berg, Tommi Vatanen, Chris Chandler, Bryn C Taylor, Ian M Fisk, Hera Vlamakis, et al. Structurebased protein function prediction using graph convolutional networks.

Nature communications, 12(1):3168, 2021

Patrik O Hoyer. Non-negative matrix factorization with sparseness constraints. p. 9

Where the paper says it

Patrik O Hoyer. Non-negative matrix factorization with sparseness constraints.

Journal of machine learning research, 5(Nov):1457–1469, 2004

Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. p. 9

Where the paper says it

Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning.

In International conference on machine learning, pages 2127–2136. p. 9
Where the paper says it

Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127–2136.

PMLR, 2018

Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. p. 9

Where the paper says it

Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax.

arXiv preprint arXiv:1611. p. 9
Where the paper says it

Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.

01144, 2016

John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek. Highly accurate protein structure prediction with alphafold. p. 0

Where the paper says it

For a long time, a key obstacle to assigning function was the difficulty of determining the structure for a given sequence. This bottleneck has been alleviated by accurate structure predictors such as the AlphaFold [Jumper et al.

nature, 596(7873):583–589

Nikolaos Karalias and Andreas Loukas. Erdos goes neural: an unsupervised learning framework for combinatorial optimization on graphs. p. 9

Where the paper says it

Partition and code: learning how to compress graphs. Advances in Neural Information Processing Systems, 34: 18603–18619, 2021.

Advances in Neural Information Processing Systems, 33:6659–6672, 2020

Tim Kucera, Carlos Oliver, Dexiong Chen, and Karsten Borgwardt. Proteinshake: Building datasets and benchmarks for deep learning on protein structures. p. 0

Where the paper says it

, 2023] families, yielding databases with hundreds of millions of predicted structures. With structures now broadly available, the central challenge shifts to learning representations that support downstream tasks like function prediction and protein design.

In Advances in Neural Information Processing Systems 36 (NeurIPS 2023) — Datasets and Benchmarks Track, pages 58277–58289

Dmitry D. Leipe, Eugene V. p. 1

Where the paper says it

NTPase sites are paradigmatic examples [Dodson and Wlodawer, 1998, Rossmann et al. , 1974, Leipe et al.

Koonin, and L. Aravind. Evolution and classification of p-loop kinases and NTPases. p. 0
Where the paper says it

A central goal in studying protein function is to identify which substructures of a protein are responsible for its function. These functional substructures are typically cohesive, modular units: catalytic triads (Ser–His–Asp), Rossmann-like nucleotide-binding cores, and P-loop (Walker A/B) Preprint.

Journal of Molecular Biology, 333(4):781–815, 2003

Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin. Evolutionary-scale prediction of atomic-level protein structure with a language model. p. 8

Where the paper says it

, 2021] through gradient-saliency on ´ a trained function classifier, and sparse autoencoders applied to protein language models through feature decomposition [Simon and Zou, 2025]. BIOBLOBS instead reasons over learned substructures as first-class units of the prediction, recovering experimentally annotated catalytic sites under protein-level supervision alone.

Science, 379(6637):1123–1130

Chuang Liu, Yibing Zhan, Jia Wu, Chang Li, Bo Du, Wenbin Hu, Tongliang Liu. Graph pooling for graph neural networks: Progress, challenges, and opportunities. p. 9

Where the paper says it

Vladimir Gligorijevic, P Douglas Renfrew, Tomasz Kosciolek, Julia Koehler Leman, Daniel Beren- ´ berg, Tommi Vatanen, Chris Chandler, Bryn C Taylor, Ian M Fisk, Hera Vlamakis, et al. Structurebased protein function prediction using graph convolutional networks.

arXiv preprint arXiv:2204. p. 0
Where the paper says it

These functional substructures are typically cohesive, modular units: catalytic triads (Ser–His–Asp), Rossmann-like nucleotide-binding cores, and P-loop (Walker A/B) Preprint. arXiv:2510.

07321

António J M Ribeiro, Gemma L Holliday, Nicholas Furnham, Jonathan D Tyzack, Katherine Ferris; Janet M Thornton. Mechanism and Catalytic Site Atlas (M-CSA): a database of enzyme reaction mechanisms and active sites. p. 0

Where the paper says it

The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains. Trained only on protein-level labels, BIOBLOBS recovers experimentally annotated catalytic sites in the M-CSA database, demonstrating unsupervised functional substructure discovery and opening a path to large-scale functional site discovery across the unannotated proteome.

Nucleic Acids Research, 46(D1):D618–D623

Michael G. Rossmann, Dino Moras, and Kenneth W. p. 0

Where the paper says it

A central goal in studying protein function is to identify which substructures of a protein are responsible for its function. These functional substructures are typically cohesive, modular units: catalytic triads (Ser–His–Asp), Rossmann-like nucleotide-binding cores, and P-loop (Walker A/B) Preprint.

Olsen. Chemical and biological evolution of a nucleotide-binding protein. p. 0
Where the paper says it

A central goal in studying protein function is to identify which substructures of a protein are responsible for its function. These functional substructures are typically cohesive, modular units: catalytic triads (Ser–His–Asp), Rossmann-like nucleotide-binding cores, and P-loop (Walker A/B) Preprint.

Nature, 250:194–199, 1974

Elana Simon and James Zou. InterPLM: discovering interpretable features in protein language models via sparse autoencoders. p. 8

Where the paper says it

Post-hoc methods surface sites without residue-level supervision: DeepFRI [Gligorijevic et al. , 2021] through gradient-saliency on ´ a trained function classifier, and sparse autoencoders applied to protein language models through feature decomposition [Simon and Zou, 2025].

Nature Methods, 22(10):2107–2117, October 2025

Martin Steinegger and Johannes Söding. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. p. 4

Where the paper says it

For every ProteinShake dataset, we evaluate under two splitting strategies designed to reduce train/test leakage and probe generalization: a sequence split, which clusters proteins by sequence identity, and a structure split, which clusters by structural similarity. In both cases, clusters are assigned to disjoint train, validation, and test sets at a 0.

Nature Biotechnology, 35(11):1026–1028, November 2017

Jin Su, Chenchen Han, Yuyang Zhou, Junjie Shan, Xibin Zhou, and Fajie Yuan. SaProt: Protein Language Modeling with Structure-aware Vocabulary. p. 4

Where the paper says it

, 2023], a sequence-only transformer trained with masked language modeling; and SaProt-650M [Su et al. , 2023], a sequence-structure transformer trained on residues augmented with Foldseek structural tokens.

October 2023

Freyr Sverrisson, Jean Feydy, Bruno E Correia, and Michael M Bronstein. Fast end-to-end learning on protein surfaces. p. 0

Where the paper says it

Yet existing pipelines built on protein encoders do not model proteins at the substructure level, leaving the central biological question unanswered: which substructure of a protein is responsible for its function? We introduce BIOBLOBS, an encoder-agnostic, end-to-end differentiable framework that compresses a protein into a small set of cohesive substructures (blobs) and predicts function from these blobs alone, so that each blob corresponds to a candidate functional region.

In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15272–15281, 2021

Yang Tan, Wenrui Gou, Bozitao Zhong, Liang Hong, Huiqun Yu; Bingxin Zhou. Venusx: unlocking fine-grained functional understanding of proteins. p. 4

Where the paper says it

VenusX [Tan et al. , 2025] provides four fragment-level functional-site tasks derived from InterPro [Blum et al.

arXiv preprint arXiv:2505. p. 0
Where the paper says it

These functional substructures are typically cohesive, modular units: catalytic triads (Ser–His–Asp), Rossmann-like nucleotide-binding cores, and P-loop (Walker A/B) Preprint. arXiv:2510.

11812, 2025

Jérôme Tubiana, Dina Schneidman-Duhovny, and Haim J Wolfson. Scannet: an interpretable geometric deep learning model for structure-based protein binding site prediction. p. 8

Where the paper says it

Our biological validation also focuses on catalytic-site recovery, leaving systematic inspection of the substructures learned for binding, regulatory, and structural aspects to future work. 7 Conclusion BIOBLOBS reframes protein function prediction as reasoning over a compact set of cohesive substructures rather than pooled residue embeddings, matching or exceeding strong baselines across ProteinShake and VenusX on multiple sequence- and structure-based encoders while committing to only a small fraction of residues.

Nature methods, 19(6):730–739, 2022

Joseph L Watson, David Juergens, Nathaniel R Bennett, Brian L Trippe, Jason Yim, Helen E Eisenach. De novo design of protein structure and function with rfdiffusion. p. 0

Where the paper says it

, 2023] families, yielding databases with hundreds of millions of predicted structures. With structures now broadly available, the central challenge shifts to learning representations that support downstream tasks like function prediction and protein design.

Nature, 620(7976):1089–1100

Lirong Wu, Yufei Huang, Haitao Lin, and Stan Z Li. A survey on protein representation learning: Retrospect and prospect. p. 0

Where the paper says it

With structures now broadly available, the central challenge shifts to learning representations that support downstream tasks like function prediction and protein design. Protein representation learning (PRL) [Wu et al.

arXiv preprint arXiv:2301. p. 0
Where the paper says it

These functional substructures are typically cohesive, modular units: catalytic triads (Ser–His–Asp), Rossmann-like nucleotide-binding cores, and P-loop (Walker A/B) Preprint. arXiv:2510.

00813, 2022

Huanrui Yang, Wei Wen, and Hai Li. Deephoyer: Learning sparser neural network with differentiable scale-invariant sparsity measures. p. 3

Where the paper says it

We therefore regularize the soft membership matrix M with Hoyer’s scale-invariant sparseness measure [Hoyer, 2004], originally proposed for non-negative matrix factorization with sparse, localized basis columns and later adapted by DeepHoyer [Yang et al. , 2019] for sparse neural networks.

arXiv preprint arXiv:1908. p. 0
Where the paper says it

These functional substructures are typically cohesive, modular units: catalytic triads (Ser–His–Asp), Rossmann-like nucleotide-binding cores, and P-loop (Walker A/B) Preprint. arXiv:2510.

09979, 2019

Zuobai Zhang, Minghao Xu, Arian Jamasb, Vijil Chenthamarakshan, Aurelie Lozano, Payel Das; Jian Tang. Protein representation learning by geometric structure pretraining. p. 0

Where the paper says it

Protein representation learning (PRL) [Wu et al. , 2022] addresses this by training neural encoders that map a protein’s sequence and/or structure into vector representations of its structural and biophysical properties.

arXiv preprint arXiv:2203. p. 0
Where the paper says it

These functional substructures are typically cohesive, modular units: catalytic triads (Ser–His–Asp), Rossmann-like nucleotide-binding cores, and P-loop (Walker A/B) Preprint. arXiv:2510.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Protein function prediction can be performed by compressing a protein into a small set of cohesive substructures called blobs and predicting function from these blobs alone, so that each blob corresponds to a candidate functional region.

  2. The improved system can adapt its spatial scale to the task, ranging from local catalytic sites to entire structural domains, as the discovered blobs adapt their spatial scale based on the downstream task.

  3. The system can recover experimentally annotated catalytic sites in databases like M-CSA from protein-level supervision alone by mining over learned substructures, opening a path to large-scale functional site discovery across the unannotated proteome.

Abstract

Protein function is often localized to cohesive substructures, such as catalytic triads, binding pockets, and structural motifs, that occupy only a small fraction of a protein's residues. Yet existing predictive pipelines typically aggregate residue-level embeddings and do not explicitly model or infer these functional substructures, leaving the central biological question unanswered: which substructure of a protein is responsible for its function? We introduce BioBlobs, an encoder-agnostic, end-to-end differentiable framework that compresses a protein into a small set of cohesive substructures (blobs) and predicts function from these blobs alone, so that each blob corresponds to a candidate functional region. Across diverse protein function prediction tasks and both sequence-only and structure-aware protein language models, BioBlobs is competitive with strong baselines while operating on only a small fraction of residues. Trained only on protein function labels, BioBlobs recovers expert-curated functional sites in InterPro and experimentally annotated catalytic sites in the Mechanism and Catalytic Site Atlas, demonstrating unsupervised functional substructure discovery and opening a path to large-scale functional site discovery in unannotated proteins. The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains. Code is available at https://github.com/OliverLaboratory/BioBlobs.

Sources

Related papers