PIT-GCL: Protein Interaction using Topological Graph Contrastive Learning

Jae Won Choi; Ryoonki Hong; Alan Liang; Manjula Adiveppa Wader; Bingsong Zeng; Peiyang Tang; Longwei Liu; Ruishan Liu

cs.LGOct 4, 2026

PIT-GCL: Protein Interaction using Topological Graph Contrastive Learning

Authors: Jae Won Choi, Ryoonki Hong, Alan Liang, Manjula Adiveppa Wader, Bingsong Zeng, Peiyang Tang, Longwei Liu, Ruishan Liu

Organizations: Department of Computer Science, University of Southern California

Abstract

Protein binding prediction is central to target identification, therapeutic binder design, and large scale screening, yet remains challenging because binding depends on sequence, three dimensional geometry, and global structural organization. Recent folding models such as AlphaFold3 and Boltz-2 have substantially improved structure prediction, but their confidence outputs (pLDDT, pTM, ipTM) are not specifically designed for binary binding prediction, and dedicated structure aware predictors often require bound complex structures that are unavailable at screening scale. We introduce PIT-GCL, a dual tower structure aware framework that encodes each protein independently from its amino acid sequence, Cα point cloud, and a global persistent homology descriptor. Each tower combines residue ESM-2 embeddings with a topological summary computed from the H0 and H1 persistence landscapes of a Vietoris-Rips filtration, and processes the resulting tokens with a structure aware Transformer in which pairwise Cα distances enter as a learned attention bias. A bidirectional cross attention module then performs latent space soft docking between the two per-protein representations, and the model is trained with a combined binary cross entropy and NT-Xent contrastive objective. On three binary interaction prediction benchmarks, general PPI on PPIRef, TCRpMHC binding on STAG, and whole chain pairs on PPB-Affinity, PIT-GCL outperforms representative sequence based, structure aware, and task specific baselines on general PPI under our evaluation, and is the only method above chance on PPB-Affinity; on TCR-pMHC it leads at a fixed decision threshold but is outranked by a task specific sequence model. Because each protein is encoded independently in the first phase, its representation can be precomputed and reused across candidate pairs, which is convenient for large scale screening.

Figures & tables

Figure 1 : The per-protein encoder. A frozen ESM-2 embedding is projected to the model width, a two-layer MLP maps the persistence landscape to the same width, and the two are added so every residue token carries the protein’s global topology; four structure-aware layers then attend over the residues with a bias built from the C

\alpha

coordinates (Figure 2 ). How the landscape itself is built is shown in Figure 3 .

Model Name	AUC $\uparrow$	ACC $\uparrow$	F1 $\uparrow$	PR-AUC $\uparrow$	RMSE $\downarrow$
DeepDTA	$0.8691_{\scriptscriptstyle\pm 0.0059}$	$0.7785_{\scriptscriptstyle\pm 0.0070}$	$0.7791_{\scriptscriptstyle\pm 0.0042}$	$0.8765_{\scriptscriptstyle\pm 0.0038}$	$0.3865_{\scriptscriptstyle\pm 0.0055}$
PIPR	$0.8840_{\scriptscriptstyle\pm 0.0033}$	$0.7944_{\scriptscriptstyle\pm 0.0024}$	$0.7825_{\scriptscriptstyle\pm 0.0042}$	$0.8948_{\scriptscriptstyle\pm 0.0026}$	$0.3723_{\scriptscriptstyle\pm 0.0030}$
D-SCRIPT	$0.6517_{\scriptscriptstyle\pm 0.0584}$	$0.5409_{\scriptscriptstyle\pm 0.0243}$	$0.2060_{\scriptscriptstyle\pm 0.1802}$	$0.6760_{\scriptscriptstyle\pm 0.0553}$	$0.5273_{\scriptscriptstyle\pm 0.0222}$
PPIFormer	$0.8334_{\scriptscriptstyle\pm 0.0058}$	$0.6546_{\scriptscriptstyle\pm 0.0608}$	$0.4692_{\scriptscriptstyle\pm 0.1417}$	$0.8692_{\scriptscriptstyle\pm 0.0051}$	$0.4636_{\scriptscriptstyle\pm 0.0085}$
GIGN †	$0.9986_{\scriptscriptstyle\pm 0.0009}$	$0.7781_{\scriptscriptstyle\pm 0.1596}$	$0.6754_{\scriptscriptstyle\pm 0.2718}$	$0.9954_{\scriptscriptstyle\pm 0.0036}$	$0.3426_{\scriptscriptstyle\pm 0.1300}$
PIT-GCL (ours)	$\mathbf{0.9302}_{\scriptscriptstyle\pm\mathbf{0.0037}}$	$\mathbf{0.8410}_{\scriptscriptstyle\pm\mathbf{0.0057}}$	$\mathbf{0.8413}_{\scriptscriptstyle\pm\mathbf{0.0052}}$	$\mathbf{0.9365}_{\scriptscriptstyle\pm\mathbf{0.0035}}$	$\mathbf{0.3314}_{\scriptscriptstyle\pm\mathbf{0.0057}}$

Table 1: Performance on the PPIRef test set. A logistic regression on five chain-length features alone reaches AUC

0.8138

on this split, so the absolute numbers include a benchmark cue no method has to learn; Appendix B measures it and Section 4.3 reports what survives when it is removed. † GIGN is reported for completeness but is not a binding-prediction result and is excluded from the ranking (Appendix B ). D-SCRIPT is the eight-member ensemble its authors’ pipeline produces (Appendix C ).

Model Name	AUC $\uparrow$	ACC $\uparrow$	F1 $\uparrow$	PR-AUC $\uparrow$	RMSE $\downarrow$
ConTCR	$0.5001_{\scriptscriptstyle\pm 0.0017}$	$0.5004_{\scriptscriptstyle\pm 0.0016}$	$0.3104_{\scriptscriptstyle\pm 0.2898}$	$0.5000_{\scriptscriptstyle\pm 0.0026}$	$0.5000_{\scriptscriptstyle\pm 0.0000}$
epiTCR	$0.5314_{\scriptscriptstyle\pm 0.0188}$	$0.5080_{\scriptscriptstyle\pm 0.0016}$	$0.0500_{\scriptscriptstyle\pm 0.0075}$	$0.5348_{\scriptscriptstyle\pm 0.0149}$	$0.5121_{\scriptscriptstyle\pm 0.0018}$
epiTCR-MHC	$0.5342_{\scriptscriptstyle\pm 0.0107}$	$0.5097_{\scriptscriptstyle\pm 0.0023}$	$0.0920_{\scriptscriptstyle\pm 0.0058}$	$0.5434_{\scriptscriptstyle\pm 0.0088}$	$0.5051_{\scriptscriptstyle\pm 0.0009}$
NetTCR-2.0	$0.5048_{\scriptscriptstyle\pm 0.0278}$	$0.5028_{\scriptscriptstyle\pm 0.0225}$	$0.3341_{\scriptscriptstyle\pm 0.1814}$	$0.5024_{\scriptscriptstyle\pm 0.0199}$	$0.5003_{\scriptscriptstyle\pm 0.0016}$
GraphDTA	$0.4972_{\scriptscriptstyle\pm 0.0165}$	$0.4983_{\scriptscriptstyle\pm 0.0057}$	$0.4642_{\scriptscriptstyle\pm 0.1988}$	$0.5030_{\scriptscriptstyle\pm 0.0129}$	$0.5118_{\scriptscriptstyle\pm 0.0241}$
DeepDTA ∗	$0.5054_{\scriptscriptstyle\pm 0.0098}$	$0.5050_{\scriptscriptstyle\pm 0.0067}$	$0.3005_{\scriptscriptstyle\pm 0.2644}$	$0.5091_{\scriptscriptstyle\pm 0.0059}$	$0.5816_{\scriptscriptstyle\pm 0.0784}$

Table 2: Performance on PPB-Affinity. Negatives are re-paired chains from different complexes, as on PPIRef; the measured affinities are not thresholded and do not enter the label. ∗ marks the general-purpose PPI baselines of Table 1 , retrained on this benchmark’s training split; the unmarked rows are models built for other interaction settings. Every baseline sits at chance either way. No RMSE is bolded: the column’s lowest values belong to arms that emit a constant

0.5

, which scores

0.5000

on a balanced split without ordering anything.

Arm	AUC $\uparrow$	ACC $\uparrow$	F1 $\uparrow$	PR-AUC $\uparrow$	RMSE $\downarrow$
main full model	$\mathbf{0.9305}_{\scriptscriptstyle\pm\mathbf{0.0048}}$	$\mathbf{0.8414}_{\scriptscriptstyle\pm\mathbf{0.0077}}$	$\mathbf{0.8422}_{\scriptscriptstyle\pm\mathbf{0.0071}}$	$\mathbf{0.9370}_{\scriptscriptstyle\pm\mathbf{0.0042}}$	$\mathbf{0.3295}_{\scriptscriptstyle\pm\mathbf{0.0067}}$
A1 $-$ topology	$0.9149_{\scriptscriptstyle\pm 0.0039}$	$0.8246_{\scriptscriptstyle\pm 0.0061}$	$0.8208_{\scriptscriptstyle\pm 0.0025}$	$0.9248_{\scriptscriptstyle\pm 0.0041}$	$0.3516_{\scriptscriptstyle\pm 0.0035}$
A1z topo. zeroed	$0.9165_{\scriptscriptstyle\pm 0.0046}$	$0.8276_{\scriptscriptstyle\pm 0.0049}$	$0.8263_{\scriptscriptstyle\pm 0.0046}$	$0.9259_{\scriptscriptstyle\pm 0.0037}$	$0.3460_{\scriptscriptstyle\pm 0.0038}$
A1s topo. shuffled	$0.9157_{\scriptscriptstyle\pm 0.0043}$	$0.8266_{\scriptscriptstyle\pm 0.0056}$	$0.8240_{\scriptscriptstyle\pm 0.0065}$	$0.9255_{\scriptscriptstyle\pm 0.0040}$	$0.3473_{\scriptscriptstyle\pm 0.0029}$
A6 $-$ bias	$0.9302_{\scriptscriptstyle\pm 0.0040}$	$0.8430_{\scriptscriptstyle\pm 0.0046}$	$0.8427_{\scriptscriptstyle\pm 0.0039}$	$0.9360_{\scriptscriptstyle\pm 0.0032}$	$0.3309_{\scriptscriptstyle\pm 0.0034}$
A6z bias zeroed	$0.9305_{\scriptscriptstyle\pm 0.0060}$	$0.8427_{\scriptscriptstyle\pm 0.0072}$	$0.8392_{\scriptscriptstyle\pm 0.0068}$	$0.9364_{\scriptscriptstyle\pm 0.0050}$	$0.3322_{\scriptscriptstyle\pm 0.0076}$

Table 3 : Ablation on the PPIRef test set, seeds

42

–

44

, on the

H_{0}{+}H_{1}

configuration of Table 1 . Bold marks the full model as the reference, not a winner. Each A arm is one flag on the released code path. † C1 is not: the same cached embeddings mean-pooled into an MLP, untuned, over five seeds. ‡ A2 and § A5 are not read at face value. Appendix E explains all three.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Train		Val		Test
Dataset	Pos	Neg	Pos	Neg	Pos	Neg
STAG [ 34 ]	5,906	19,497	-	-	1,476	4,874
PPIRef [ 4 ]	40,868	40,868	5,099	5,099	5,121	5,121
PPB-Affinity [ 23 ]	4,207	4,207	518	518	517	517

Appendix

Table 4 : Dataset statistics for each task. STAG is released as a two-way split, so its validation fold is the

10\%

of the train column that each seed holds out rather than a separate fold; PPIRef and PPB-Affinity both use an

80/10/10

partition with held-out validation and test folds. The PPIRef counts are the interfaces actually used: the

80/10/10

partition of the

51{,}755

deduplicated keys yields

41{,}404/5{,}175/5{,}176

entries, of which

536/76/55

are dropped because the corresponding structure file is absent from the released ppi_6A archive. The PPIRef test set therefore contains

10{,}242

pairs. PPB-Affinity is partitioned

80/10/10

over

10{,}484

balanced pairs and, like PPIRef, carries an equal number of positives and negatives in every fold.

Config.	AUC $\uparrow$	ACC $\uparrow$	F1 $\uparrow$	PR-AUC $\uparrow$	RMSE $\downarrow$
Ensemble of 8	$0.6517_{\scriptscriptstyle\pm 0.0584}$	$0.5409_{\scriptscriptstyle\pm 0.0243}$	$0.2060_{\scriptscriptstyle\pm 0.1802}$	$0.6760_{\scriptscriptstyle\pm 0.0553}$	$0.5273_{\scriptscriptstyle\pm 0.0222}$
Single model	$0.5158_{\scriptscriptstyle\pm 0.0276}$	$0.5130_{\scriptscriptstyle\pm 0.0223}$	$0.3345_{\scriptscriptstyle\pm 0.2577}$	$0.5571_{\scriptscriptstyle\pm 0.0374}$	$0.6125_{\scriptscriptstyle\pm 0.0491}$

Appendix

Table 5: D-SCRIPT on the PPIRef test set, means over five seeds. Table 1 carries the ensemble, the stronger of the two configurations.

Model	Dataset	Threshold	Acc	F1	Predicted 1
PIT-GCL	PPIRef	0.1	0.7598	0.8026	7343
PIT-GCL	PPIRef	0.3	0.8335	0.8466	5994
PIT-GCL	PPIRef	0.5	0.8501	0.8500	5114
PIT-GCL	PPIRef	0.7	0.8409	0.8268	4289
PIT-GCL	PPIRef	0.9	0.7976	0.7520	3238
PIT-GCL	STAG	0.1	0.8619	0.7203	1659

Appendix

Table 6: PIT-GCL under different classification thresholds on PPIRef and STAG, on the seed-

42

partition of each; the single-seed accuracy at

0.5

is therefore

0.8501

rather than the five-seed

0.8410

of Table 1 . Best value per column block in bold .

Method	AUC $\uparrow$	F1 @ $0.5$	F1 @ val-thr.	F1 @ oracle
DeepDTA	$0.8691_{\scriptscriptstyle\pm 0.0059}$	$0.7791_{\scriptscriptstyle\pm 0.0042}$	$0.7849_{\scriptscriptstyle\pm 0.0041}$	$0.7865_{\scriptscriptstyle\pm 0.0044}$
PIPR	$0.8840_{\scriptscriptstyle\pm 0.0033}$	$0.7825_{\scriptscriptstyle\pm 0.0042}$	$0.7959_{\scriptscriptstyle\pm 0.0040}$	$0.7980_{\scriptscriptstyle\pm 0.0042}$
PPIFormer	$0.8334_{\scriptscriptstyle\pm 0.0058}$	$0.4692_{\scriptscriptstyle\pm 0.1417}$	$0.6676_{\scriptscriptstyle\pm 0.0411}$	$0.7613_{\scriptscriptstyle\pm 0.0067}$
PIT-GCL (ours)	$\mathbf{0.9302}_{\scriptscriptstyle\pm\mathbf{0.0037}}$	$\mathbf{0.8413}_{\scriptscriptstyle\pm\mathbf{0.0052}}$	$\mathbf{0.8439}_{\scriptscriptstyle\pm\mathbf{0.0051}}$	$\mathbf{0.8458}_{\scriptscriptstyle\pm\mathbf{0.0051}}$

Appendix

Table 7: Threshold transfer on PPIRef, five seeds. “val-thr.” selects the threshold maximising F1 on that seed’s validation predictions and applies it unchanged to test; “oracle” is the best test threshold, an upper bound no method can reach in practice. The gap between the last two columns is how much a method loses to miscalibration:

0.0937

for PPIFormer against

0.0019

for PIT-GCL. GIGN and D-SCRIPT are omitted because they do not expose per-seed validation predictions in the form this comparison needs.

Figure 2 : Construction of the structural attention bias

B_{\mathrm{struct}}

. Pairwise C

\alpha

distances are expanded by

16

radial basis functions whose centres are fixed on

[0,20]\,\text{\AA}

and whose widths are learned, then projected to one bias matrix per attention head. The bias is added to the attention logits inside the softmax, so geometry reweights attention rather than post-processing it, and it is dense over all

N^{2}

residue pairs rather than restricted to a neighbourhood graph. The kernel widths and the

16\!\to\!8

projection are the only learned parameters of this branch (

152

per layer,

608

across the four layers), which is the component removed in arm A6 of Section 4.3 .

	Arm	AUC $\uparrow$	ACC $\uparrow$	F1 $\uparrow$	PR-AUC $\uparrow$	RMSE $\downarrow$
main ${}_{H_{0}}$	Full model	$0.9283_{\scriptscriptstyle\pm 0.0044}$	$0.8407_{\scriptscriptstyle\pm 0.0046}$	$0.8387_{\scriptscriptstyle\pm 0.0031}$	$0.9352_{\scriptscriptstyle\pm 0.0034}$	$0.3322_{\scriptscriptstyle\pm 0.0063}$
A2 ${}_{H_{0}}$	seq. $\to$ sinusoidal	$0.9057_{\scriptscriptstyle\pm 0.0045}$	$0.8124_{\scriptscriptstyle\pm 0.0055}$	$0.8159_{\scriptscriptstyle\pm 0.0046}$	$0.9083_{\scriptscriptstyle\pm 0.0046}$	$0.3565_{\scriptscriptstyle\pm 0.0049}$
A3 ${}_{H_{0}}$	seq. $\to$ one-hot	$0.9220_{\scriptscriptstyle\pm 0.0049}$	$0.8348_{\scriptscriptstyle\pm 0.0061}$	$0.8310_{\scriptscriptstyle\pm 0.0078}$	$0.9315_{\scriptscriptstyle\pm 0.0046}$	$0.3447_{\scriptscriptstyle\pm 0.0072}$
A4 ${}_{H_{0}}$	seq. $\to$ ESM-35M	$0.9270_{\scriptscriptstyle\pm 0.0040}$	$0.8385_{\scriptscriptstyle\pm 0.0071}$	$0.8380_{\scriptscriptstyle\pm 0.0063}$	$0.9338_{\scriptscriptstyle\pm 0.0033}$	$0.3354_{\scriptscriptstyle\pm 0.0053}$
A5 ${}_{H_{0}}$	floor	$0.4947$	$0.5000$	$0.6667$	$0.4975$	$0.5371$

Appendix

Table 8 : The sequence arms on the

H_{0}

-only descriptor, against its own full model. Two of them are a different ablation rather than the same one on a different descriptor, which is why both versions are reported: A2

{}_{H_{0}}

substitutes sinusoidal encodings for the language-model embedding (

2{,}908{,}001

parameters) where A2 zeroes it and keeps the projection trained (

2{,}989{,}921

, parameter-matched to the full model), and A3

{}_{H_{0}}

projects a

21

-d one-hot code directly (

2{,}913{,}377

) where A3 places the same code in the leading channels of the full-width input (

2{,}989{,}921

). A4 and A4

{}_{H_{0}}

are the one like-for-like pair, identical at

3{,}030{,}881

parameters, so their difference isolates the descriptor change alone. A5

{}_{H_{0}}

’s

0.4947

is the same constant-output artefact as A5’s

0.5401

, landing on the other side of chance.

Sequence features
Hyperparameter	Value
ESM-2 model	facebook/esm2_t6_8M_UR50D
ESM hidden dimension $d_{\text{esm}}$	$320$
Topological features
Filtration	Vietoris-Rips
Max edge length	$20\,\text{\AA}$

Appendix

Table 9: PIT-GCL hyperparameters used across all experiments unless otherwise stated.

Stage	Cost	Recomputed per candidate pair?
Persistence landscape	$101$ ms / protein (median $49$ )	no
Tower encoding (Phase 1)	$2.3$ ms / protein	no
Fusion $+$ head (Phase 2)	$0.78$ ms / pair	yes

Appendix

Table 10: Per-stage cost. Only the last row scales with the number of candidates.

Figure 3 : Construction of the topological descriptor. A Vietoris-Rips filtration on the C

\alpha

cloud, expanded one dimension above the homology dimension so that triangles are present, gives

H_{0}

and

H_{1}

persistence; each is summarised by three landscapes on a

50

-point grid and the two blocks are concatenated into one

300

-dimensional vector.

Model Name	AUC $\uparrow$	ACC $\uparrow$	F1 $\uparrow$	PR-AUC $\uparrow$	RMSE $\downarrow$
STAG	$0.8564_{\scriptscriptstyle\pm 0.0077}$	$0.7671_{\scriptscriptstyle\pm 0.0073}$	$0.6106_{\scriptscriptstyle\pm 0.0065}$	$0.6801_{\scriptscriptstyle\pm 0.0154}$	$0.3907_{\scriptscriptstyle\pm 0.0059}$
TEPCAM	$0.6634_{\scriptscriptstyle\pm 0.0618}$	$0.7248_{\scriptscriptstyle\pm 0.0856}$	$0.2939_{\scriptscriptstyle\pm 0.0396}$	$0.3794_{\scriptscriptstyle\pm 0.0449}$	$0.4531_{\scriptscriptstyle\pm 0.0443}$
TRAP	$0.6705_{\scriptscriptstyle\pm 0.0956}$	$0.7721_{\scriptscriptstyle\pm 0.0105}$	$0.1668_{\scriptscriptstyle\pm 0.0988}$	$0.4089_{\scriptscriptstyle\pm 0.0996}$	$0.4403_{\scriptscriptstyle\pm 0.0240}$
ConTCR	$0.7033_{\scriptscriptstyle\pm 0.0030}$	$0.6358_{\scriptscriptstyle\pm 0.0174}$	$0.4644_{\scriptscriptstyle\pm 0.0037}$	$0.3982_{\scriptscriptstyle\pm 0.0084}$	$0.4655_{\scriptscriptstyle\pm 0.0019}$
epiTCR	$\mathbf{0.9031}_{\scriptscriptstyle\pm\mathbf{0.0025}}$	$0.8120_{\scriptscriptstyle\pm 0.0034}$	$0.6766_{\scriptscriptstyle\pm 0.0039}$	$\mathbf{0.8187}_{\scriptscriptstyle\pm\mathbf{0.0029}}$	$0.3692_{\scriptscriptstyle\pm 0.0010}$
epiTCR-MHC	$0.8945_{\scriptscriptstyle\pm 0.0020}$	$0.7869_{\scriptscriptstyle\pm 0.0038}$	$0.6468_{\scriptscriptstyle\pm 0.0044}$	$0.7412_{\scriptscriptstyle\pm 0.0032}$	$0.3866_{\scriptscriptstyle\pm 0.0011}$

Appendix

Table 11: Performance on the TCR-pMHC benchmark, STAG test split.

Model	Year	Architecture	Input Repr.
DeepDTA [ 26 ]	2018	CNN	Two protein sequences
PIPR [ 8 ]	2019	RCNN	Two protein sequences
D-SCRIPT [ 33 ]	2021	Pretrained PLM + CNN	Two protein sequences
PPIFormer [ 4 ]	2024	Transformer	3D structure
GIGN [ 41 ]	2023	Geometric Interaction GNN	3D structure
GraphDTA [ 25 ]	2021	GNN + CNN	Two protein sequences

Appendix

Table 12: Summary of the PPI baselines. GraphDTA is evaluated on PPB-Affinity only; the others are evaluated on PPIRef and, where marked in Table 2 , retrained on PPB-Affinity as well.

Model	Year	Architecture
STAG [ 34 ]	2025	Graph Neural Network
TEPCAM [ 7 ]	2024	CNN + attention
TRAP [ 16 ]	2025	Transformer + contrastive learning
ConTCR [ 32 ]	2025	Cross-attention Transformer + contrastive learning
epiTCR [ 29 ]	2023	Random Forest
epiTCR-MHC [ 29 ]	2023	Random Forest (MHC-aware)

Appendix

Table 13: Summary of TCR-pMHC baseline model architectures.

Model	Sequence				Structure
Model	Peptide	MHC	CDR3 $\alpha$	CDR3 $\beta$	Peptide	MHC	CDR3 $\alpha$	CDR3 $\beta$
STAG					✓	✓	✓	✓
TEPCAM	✓			✓
TRAP	✓			✓	✓ ‡	✓ ‡
ConTCR	✓	✓		✓
epiTCR	✓			✓

Appendix

Table 14: Input representations used by TCR-pMHC baseline models. † TITAN represents epitopes using SMILES rather than standard amino-acid sequence encoding. ‡ TRAP is the one baseline whose structural input is predicted rather than resolved: it folds each pMHC with AlphaFold-Multimer and extracts features from the epitope’s local environment, so the MHC enters through that structure and not as a sequence embedding.

Explore similar work

May 19, 2026cs.LG

Hierarchical Contrastive Learning for Multi-Domain Protein-Ligand Binding

Predicting protein-ligand binding affinity remains intractable for multi-domain proteins, where inter-domain dynamics govern molecular recognition. Existing geometric deep learning methods typically treat proteins as monolithic static graphs, suffering from rigid-body assumptions and aleatoric noise in flexible regions. To address this, we introduced HCLBind, a self-supervised framework that decouples geometric representation learning from affinity regression. HCLBind leverages a general-to-specific pre-training paradigm on the Q-BioLiP database to learn a robust physical grammar of binding. We propose a novel hierarchical decoy strategy: the model learns local physicochemical constraints through protein coordinate perturbation in single-domain proteins and global conformational geometry through inter-domain rotation in multi-domain complexes. Our hybrid architecture integrates a domain-gated graph attention network and cross-modal attention to explicitly prioritize domain interfaces. Furthermore, we employ LoRA on protein and ligand foundation models, ensuring efficient optimization while preserving evolutionary knowledge. Experiments on PDBBind demonstrate that HCLBind effectively learns discriminative interface features and provides robust uncertainty estimation, overcoming the limitations of standard supervised learning. The code is available at https://github.com/jiankliu/HCLBind.

Shuo Zhang, Rongqi Hong, Huifeng Zhang +1

University of Birmingham, UK

Protein Structure Prediction Geometric Deep Learning Contrastive Learning Interleaved Graph Attention Wireless Foundation Models

Ask PDF

May 19, 2026q-bio.QM

Protein Thoughts: Interpretable Reasoning with Tree of Thoughts and Embedding-Space Flow Matching for Protein-Protein Interaction Discovery

Protein-protein interactions (PPIs) govern nearly all cellular processes, yet computational methods for identifying binding partners typically produce ranked predictions without mechanistic justification. This creates a fundamental barrier to adoption because biologists cannot assess whether predictions reflect genuine biochemical insight or spurious correlations. We present \textbf{Protein Thoughts}, a framework that reformulates PPI discovery as an interpretable search problem with explicit reasoning. The system decomposes binding evidence into four biologically meaningful signals: sequence similarity reflecting evolutionary relationships, structural complementarity capturing geometric fit, interface balance, and chemical compatibility encoding residue-level interactions. Rather than collapsing these signals into an opaque score, we preserve their individual contributions through a transparent value function that enables both ranking and auditing. To navigate large candidate spaces efficiently, we introduce hypothesis-guided entropy-regularized Tree-of-Thoughts search. A fine-tuned language model generates search directives from embedding-derived features, classifying candidates as high-priority, exploratory, or skippable. These directives condition a Boltzmann policy that balances exploitation with entropy-driven exploration, while hypothesis-aware pruning prevents premature abandonment of promising candidates. For candidates exhibiting score disagreement, hypothesis-conditioned embedding-space flow matching transports protein embeddings toward the binder manifold. On the SHS148k benchmark, Protein Thoughts achieves mean best-binder rank of 11.2 versus 47.7 for an entropic tree search baseline, a 76% improvement, and for binding prediction the trained value function achieves

91.08 \pm 0.19

Micro-F1, outperforming existing PPI methods on the same dataset.

Kingsley Yeon, Xuefeng Liu, Promit Ghosal

Department of Statistics and CCAM University of Chicago Chicago, IL 60637 · School of Medicine Stanford University Stanford, CA 94305 · Department of Statistics University of Chicago Chicago, IL 60637

Biomedical Knowledge Graph Tree Search Thoughts

Ask PDF

Aug 10, 2026cs.LG

RAVEN: Frozen Random Graph Reservoirs with Physics-Informed Interaction Fingerprints for Protein-Ligand Binding Affinity Prediction

Quantitative estimation of protein-ligand binding affinity from three-dimensional complex structures is a fundamental task in structure-based computational chemistry and molecular modeling. Reliable prediction remains challenging because available structure-affinity data are limited, experimentally heterogeneous, conformation-dependent, and sensitive to dataset partitioning. RAVEN (Randomized Atomistic Views with Ensemble Neural Reservoirs) utilizes a multihead reservoir of independently initialized and fully frozen atomistic graph encoders to generate diverse structural projections without end-to-end optimization of the graph representation. These projections are integrated with a deterministic physicochemical interaction fingerprint and processed by heterogeneous supervised readers, including neural and tree-based regressors, whose outputs are combined through validation-based nonnegative fusion. The random reservoir expands structural feature coverage across independent encoder realizations, whereas the explicit physicochemical descriptors and heterogeneous readers contribute complementary information and distinct inductive biases. Evaluation on a similarity-isolated PDBbind 2020R1 split reconstructed using GEMS similarity resources, together with the protected CASF-2016 subset, demonstrated strong predictive performance. The results indicate that frozen multi-view graph representations, explicit physicochemical statistics, and heterogeneous model fusion provide a robust and flexible framework for protein-ligand binding-affinity prediction.

Qingyang Zou, Jiaye Huang, Hangbo Xie +3

School of Science, China Pharmaceutical University, Nanjing, 210009, China · School of Basic Medicine and Clinical Pharmacy, China Pharmaceutical University, Nanjing, 210009, China

Protein Structure Prediction Conformational Ensembles Molecular Property Prediction Inhomogeneous Random Graphs Physics-Informed Learning Reservoir Computing

Ask PDF