Pith. sign in

REVIEW 5 major objections 6 minor 37 references

Aptamer-protein interaction prediction model based on transformer

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that SelfTrans-Ensemble, a model fusing transformer-based sequence encoders with structural features, predicts aptamer–protein binding with 88.0% accuracy, and that its mutation sensitivity is consistent with…

desk verdict A plausible aptamer–protein prediction pipeline whose headline numbers rest on synthetic negatives; the assembled dataset and MD case study have value, but the performance claim needs an external negative set and released artifacts before it means what it says. read the letter →

arxiv 2506.16084 v1 pith:4ZE7GITC submitted 2025-06-19 q-bio.BM

Zhichao Yan , Yue Kang , Buyong Ma This is my paper · ORCID
classification q-bio.BM
keywords aptamer-proteininteractionpredictiontransformerpre-trainedlanguagemodelProtBertRNA-FMsecondarystructurefeaturesmoleculardynamicsvalidationSELEXscreening
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a hybrid model, SelfTrans-Ensemble, can predict whether an aptamer sequence binds a given protein by combining transformer-based sequence embeddings with structural features. ProtBert and RNA-FM provide the sequence representations, while K-mer frequencies, PseAAC descriptors, and predicted secondary structures feed a structural branch. On a held-out aptamer–protein test set the model reaches 88.0% accuracy, 0.896 F1, and 0.9232 AUC, outperforming the two earlier predictors it is compared with. The authors further argue that the model's sensitivity to aptamer mutations is biologically meaningful, because its scores for mutated sequences align with hydrogen-bond changes observed in 500-ns molecular dynamics simulations. The intended use is a fast in-silico pre-screening step that narrows candidate aptamers before experimental SELEX selection.

What carries the argument

The load-bearing mechanism is the ensemble architecture: a sequence component concatenates ProtBert and RNA-FM embeddings and passes them through a three-layer Transformer encoder with multi-head attention, followed by an MLP; a structure component encodes proteins with PseAAC and grouped K-mer features and aptamers/RNAs with K-mer and seven-letter secondary-structure sequences, then runs them through a CNN and a BiLSTM. The two branch scores are mapped through a linear layer, summed, and transformed by a sigmoid into a binding probability between 0 and 1, with 0.5 as the classification threshold.

What would settle it

Assemble a test set of experimentally confirmed non-binding aptamer–protein pairs, for example sequences that failed to enrich in negative SELEX selections or measured as non-binders by binding assays, and compute the model's accuracy and F1 on that set; if these scores fall well below the 88.0% accuracy and 0.896 F1 reported on synthetic negatives, the stated performance is an artifact of the negative construction.

Watch

Extended reading notes

Core claim

The central claim is that fusing two complementary representations—a sequence branch built on the pre-trained protein language model ProtBert and the RNA language model RNA-FM, and a structure branch built from K-mer counts, PseAAC descriptors, and secondary-structure predictions—lets a single classifier recognize aptamer–protein interactions accurately enough to serve as a screening tool. On the independent test set the ensemble achieves an AUC of 0.9232 and an F1 of 0.896, and attention weights in the transformer branch concentrate on the aptamer stem-loop that experimentally contacts the protein. The companion claim is that the model is mutation-sensitive in a biophysically meaningful way: a double-mutated aptamer that the model scores as a non-binder shows reduced hydrogen-bond occupancy at exactly the mutated nucleotides in molecular dynamics, while single mutations that preserve the wild-type secondary structure are still predicted to bind.

Load-bearing premise

The negative training pairs—aptamers with half their bases randomly mutated, and aptamers randomly paired with proteins they were not reported to bind—are treated as true non-binders; if a mutated aptamer still binds its target or a random pairing is a real interaction, the reported accuracy overstates true binding discrimination.

Editorial extensions

If this is right

  • If the reported accuracy holds on experimentally confirmed data, SelfTrans-Ensemble can be used to short-list aptamer candidates for a target protein before SELEX, cutting the number of wet-lab rounds.
  • The same sequence-plus-structure fusion can be transferred to other interaction prediction tasks, such as RNA–protein or peptide–protein binding, where training data are also scarce.
  • Adding short RNA–protein interaction data to the aptamer training set supports the claim that binding mechanisms are shared enough for cross-dataset augmentation to be useful.
  • The alignment between predicted binding scores and molecular-dynamics hydrogen-bond frequencies suggests the model can guide rational mutation design by flagging nucleotides whose loss weakens binding.
  • The attention-weight analysis indicates the model can point to candidate binding motifs for experimental verification.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct blind benchmark would replace the synthetic negatives with experimentally confirmed non-binding aptamer–protein pairs; the drop in accuracy, if any, would reveal how much of the 88% depends on the mutation-based negative construction.
  • The same architecture could be extended to DNA aptamers without the thymine-to-uracil conversion, using a DNA-aware pre-trained encoder, since the current conversion may discard DNA-specific structural information.
  • The gap between wild-type and mutated binding scores could be interpreted as a proxy for binding-affinity change and validated against measured dissociation constants of the same mutants.
  • If the model's insensitivity to structure-preserving single mutations is general, that could be leveraged to identify mutation-tolerant aptamer scaffolds early in design.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The manuscript introduces SelfTrans-Ensemble, a hybrid deep-learning model for aptamer-protein interaction (API) prediction. The model combines sequence embeddings from the pre-trained ProtBert and RNA-FM encoders with structural features derived from k-mer frequencies and predicted secondary structures, processed through transformer, CNN, and BiLSTM components. Training uses 1934 positive samples (aptamer-protein and short RNA-protein pairs) and 1934 synthetic negatives created by mutating half of each aptamer sequence or randomly pairing aptamers with unrelated proteins. On a self-built test set the model reports 88.0% accuracy, 0.896 F1, and 0.9232 AUC, outperforming Aptanet and AptaTrans. The authors also use 500 ns molecular dynamics simulations of the 3HXO aptamer-VWF complex and selected mutants to argue that the model is sensitive to aptamer mutations.

Significance. If the reported accuracy reflects genuine binding specificity, the model would be a useful tool for pre-screening aptamer candidates before SELEX. The paper has several strengths: it integrates strong pre-trained language models, includes an ablation study separating sequence and structure components, performs an attention-layer analysis on a concrete aptamer-protein complex, and combines deep-learning prediction with molecular dynamics validation. These are valuable methodological choices. However, the central quantitative claim rests on the validity of synthetic negatives, and the current manuscript does not provide independent experimental negatives, a sequence-similarity baseline, or a split description. The MD validation is limited to one complex and two mutants. Given these gaps, the headline performance numbers are not yet sufficient to establish that the model learns binding specificity rather than artifacts of the negative-construction procedure.

major comments (5)
  1. [Data collection and pre-processing] The negative dataset is the load-bearing assumption of the entire benchmark. Pairs obtained by mutating half of the bases of known aptamers or by randomly pairing aptamers with proteins not reported to interact are treated as true non-binders. A 50% mutation rate drastically changes k-mer composition and predicted secondary structure, so even a trivial sequence-identity or composition classifier could separate these negatives from the real positives. The manuscript acknowledges the assumption in the Discussion, but it does not test it. I ask for at least one of the following: (i) a baseline that uses sequence identity or k-mer distance to classify the same train/test split, showing that the transformer model adds signal beyond that; (ii) an independent set of experimentally validated non-binding aptamer-protein pairs; or (iii) a demonstration that predicted non-binders do not bind in a small experimental or literature-based validation set. Without this, the 88.0% accuracy and 0.9232 AUC in Table 1 and Fig. 3 do not substantiate a claim of predicting binding specificity.
  2. [Benchmark study (Table 1)] The test set and split are not described. The text states that RPI data were added to the training set and that the test set consists entirely of APIs, but it does not state how the split was made, whether homology-based filtering was applied, or whether any aptamer or protein sequence appears in both training and testing. The reader's concern about leakage is therefore not answerable from the manuscript. Also, the 'average' metrics in Table 1 and Fig. 3 have no error bars or repeated-run variability, so it is impossible to tell whether the differences between SelfTrans-Ensemble, DualBert-Trans, and Structure-Ensemble are significant. Please report the exact split procedure, sequence-overlap analysis, and metrics with standard deviations over multiple seeds.
  3. [Benchmark study (Table 1)] The comparison with AptaTrans is unfair as presented. The text says AptaTrans 'was trained from their training set and did not perform well in both our training set and testset.' Training a model on one distribution and evaluating it on a different, unseen distribution does not measure the model class's capacity; it measures domain shift. Aptanet, in contrast, was retrained on the authors' training set. For a fair comparison, AptaTrans should be retrained or fine-tuned on the same training data and evaluated under the same test protocol, or the authors should clearly state why retraining is not applicable. Without this, the benchmark claim that SelfTrans-Ensemble 'outperforms' AptaTrans is not established.
  4. [Molecular dynamic analysis with deep learning-based validations (Table 3)] The mutation-sensitivity claim is contradicted by the model's behavior on single-point mutations. Table 3 shows that for 18 single-point mutants, SelfTrans-Ensemble predicts 0 as non-binding, while DualBert-SelfTrans predicts 6. The text then states that this illustrates 'the inadequate sensitivity of the DualBert-SelfTrans model,' but the table shows the opposite: the ensemble model is less sensitive, not more, to single mutations. The later MD validation uses a double mutant (aptamer A) and a six-position mutant (aptamer B), so the demonstrated sensitivity is limited to multi-mutant sequences. This needs to be stated clearly, and the argument that the structural module improves mutation sensitivity must be reconciled with the Table 3 numbers.
  5. [Molecular dynamic analysis with deep learning-based validations] The external validation is restricted to a single complex, 3HXO, and MD is run for only two mutant aptamers (A and B, with detailed hydrogen-bond analysis only for A). The prediction score for aptamer A is 0.476, just below the 0.5 threshold, which is a fragile basis for claiming that the model correctly identifies a non-binder. To make the MD cross-validation convincing, the authors should provide more mutants, a quantitative correlation between MD observables (e.g., hydrogen-bond frequency changes or binding free energies) and model scores, and preferably additional complexes. As it stands, the MD experiment supports a qualitative trend, not a quantitative validation of the model's mutation sensitivity.
minor comments (6)
  1. [Methodology, Eq. (12)] The formula for F1 score is missing the factor 2: F1 should be 2 * Precision * Recall / (Precision + Recall). The reported values in Table 1 are consistent with the correct formula, so this appears to be a typographical error, but it should be fixed.
  2. [Methodology, Eq. (7)] In the equation for PseAAC, the second case condition is written as '1≤u≤20' but should be '21≤u≤20+λ'. Also, equation numbering jumps from (4) to (6), with no Eq. (5).
  3. [Results, Table 2 reference] The text in the molecular dynamics section says aptamer B 'was selected, as illustrated in Table 2,' but the binding scores for the mutants are in Table 4. The cross-reference should be corrected.
  4. [Results, Table 1 text] The sentence 'Aptanet ... only achieved an F1 metric of 0.624 with our training set' does not match Table 1, where Aptanet's training-set F1 is 0.767. The text should be aligned with the table.
  5. [Feature extraction] The SPIDER3 description says it predicts 'three classical protein secondary structures (α-helix, β-sheet and helix),' which should probably be coil or loop rather than repeating helix. Also, the paper does not provide a data or code availability statement, which would be important for reproducibility.
  6. [Abstract and Introduction] There are several typos and grammatical issues, including 'the aptamer dataset imbalance' (repeated word), 'apatemer' in the methodology, and 'training accuracy of 98.9%' in the abstract where the model is evaluated on the training set rather than reporting validation accuracy. These should be corrected in a thorough language edit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model's predictions are not defined by its inputs, and the acknowledged synthetic-negative assumption is a validity limitation rather than a by-construction reduction.

full rationale

The paper's prediction pipeline (pretrained BERT embeddings plus k-mer/structural features fed through transformer/CNN/BiLSTM/MLP layers) contains no step where a predicted quantity is defined as a fitted input or where the target result is assumed. The 88.0% test accuracy is measured on a held-out set of aptamer-protein pairs, and the training and test labels both derive from the authors' stated assumption that half-mutated or randomly paired aptamers are non-binders. This is an external-validity limitation, and the authors explicitly flag it in the Discussion: 'We acknowledge that our negative samples were generated based on the assumption that the mutated aptamer would not bind to the target protein if half of its sequence was mutated, and that an aptamer designed to target one protein would not bind to another protein.' Such a limitation concerns whether the benchmark measures true binding specificity, but it is not circular in the sense of the paper's derivation: the model does not use the negative-generation rule as an input feature, and the accuracy is not algebraically equal to the label-construction procedure. Similarly, the mutation-sensitivity analysis is partly in-distribution with the synthetic negatives (multi-mutant sequences resemble the half-mutated training negatives), but the single-mutant result is not forced (0/18 predicted below threshold), and the 500 ns MD simulation on aptamer A provides independent physical evidence of reduced hydrogen bonding. No load-bearing self-citation is present: ProtBert and RNA-FM are external, independently trained models, and no uniqueness theorem or ansatz from the same authors is invoked. The central performance claim therefore has independent empirical content, and the identified concerns are validity caveats rather than circular reductions.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claims rest on three kinds of assumptions: the curated interaction databases are correct, the synthetic negative construction approximates true non-binding, and RNA-protein interaction data can stand in for aptamer-protein data. These are domain assumptions, not mathematical axioms, and they are the main threat to the reported performance. The model's hyperparameters are chosen from prior work or by the authors without systematic tuning.

free parameters (8)
  • K-mer range for nucleic acid sequences = 1 to 4
    Model design choice for the structural component; affects feature vector dimensions in the Structure-Ensemble.
  • K-mer range for protein sequences = 1 to 3
    Selected following Aptanet; contributes primary and structural feature dimensions for proteins.
  • PseAAC lambda = 30
    Order of sequence correlation factors; inherited from Aptanet/Chou and determines the PseAAC feature dimension.
  • PseAAC omega = 0.05
    Weight factor for sequence order effect in PseAAC; inherited from prior work.
  • Transformer encoder layers for RNA = 3 layers, 2 attention heads
    Reported in Methods; no ablation or tuning analysis is provided.
  • Transformer encoder layers for protein = 3 layers, 8 attention heads
    Reported in Methods; no ablation or tuning analysis is provided.
  • Binding score threshold = 0.5
    Sigmoid output threshold for calling a pair binding; all mutation-sensitivity counts in Table 3 depend on this threshold.
  • Molecular dynamics simulation length = 500 ns
    Chosen for the WT and mutant simulations; influences which hydrogen bonds are observed and the RMSD stability conclusions.
assumptions (5)
  • domain assumption Aptamer-protein interactions share a binding mechanism with short RNA-protein interactions, so RNA-protein interaction data can be used as additional positives in training.
    Explicitly hypothesized in the Discussion: 'We hypothesized that the enhancement of our model effect is due to the similar binding mechanism with protein shared by aptamer and short RNA.' This assumption is load-bearing for the data augmentation strategy.
  • domain assumption Negative pairs created by randomly mutating half of the bases of a positive aptamer, or by randomly pairing aptamer and protein sequences with no reported interaction, are true non-binding pairs.
    Acknowledged in the Discussion: 'We acknowledge that our negative samples were generated based on the assumption that the mutated aptamer would not bind to the target protein if half of its sequence was mutated, and that an aptamer designed to target one protein would not bind to another protein.' If this fails, the accuracy and AUC are overestimated.
  • domain assumption DNA aptamer sequences can be converted to RNA sequences by replacing thymine (T) with uracil (U) without losing binding-relevant information.
    Stated in the data curation section: 'for DNA aptamer sequences, thymine (T) is replaced with uracil (U) in order to convert the sequence to RNA sequences.' No experimental validation of this equivalence is provided.
  • domain assumption Predicted secondary structures from RNAStructure, SPOT-RNA, and SPIDER3 are accurate enough for the structural features to be informative.
    The structural component uses these predictions as inputs. The paper does not assess the accuracy of these predictions on aptamers or proteins.
  • domain assumption The curated databases (Aptagen, APTABASE, UTexas, AptaDB, PDB, RPI2825, NPInter) correctly represent true aptamer/RNA-protein interactions.
    All positive samples depend on database annotations; no manual verification or experimental validation is described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Aptamer-protein interaction prediction model based on transformer." pith.science (2026). https://pith.science/paper/4ZE7GITC

@misc{pith2026250616084,
  author       = {Pith},
  title        = {Pith review of: Aptamer-protein interaction prediction model based on transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ZE7GITC}},
  note         = {Machine review of arXiv:2506.16084}
}
read the original abstract

Aptamers are single-stranded DNA/RNAs or short peptides with unique tertiary structures that selectively bind to specific targets. They have great potential in the detection and medical fields. Here, we present SelfTrans-Ensemble, a deep learning model that integrates sequence information models and structural information models to extract multi-scale features for predicting aptamer-protein interactions (APIs). The model employs two pre-trained models, ProtBert and RNA-FM, to encode protein and aptamer sequences, along with features generated from primary sequence and secondary structural information. To address the data imbalance in the aptamer dataset imbalance, we incorporated short RNA-protein interaction data in the training set. This resulted in a training accuracy of 98.9% and a test accuracy of 88.0%, demonstrating the model's effectiveness in accurately predicting APIs. Additionally, analysis using molecular simulation indicated that SelfTrans-Ensemble is sensitive to aptamer sequence mutations. We anticipate that SelfTrans-Ensemble can offer a more efficient and rapid process for aptamer screening.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 36 canonical work pages

  1. [1]

    Ellington, A.D. and J.W. Szostak, In vitro selectio n of RNA molecules that bind specific ligands. Nature, 1990. 346(6287): p. 818-22

    work page 1990

  2. [2]

    The embedding vector that connects the RNA and protein forms a combined vector . The combined vector is then processed through a Transformer Encoder, which consists of an embedding layer and a positional encoding layer , a multi-head attention layer , and a feed -forward layer , allowing the extraction of high-level features and contextual relationships w...

    work page

  3. [3]

    This process generated combinatorial feature vectors based on the K -mer frequencies of both inputs

    Structural Component: 1) The K -mer extraction algorithm is employed to compute the sequence features and structural features of the aptamers and proteins separately, following independent specifications [see K-mer frequency]. This process generated combinatorial feature vectors based on the K -mer frequencies of both inputs . 2) The resulting combinatori...

    work page

  4. [4]

    The n the output final binding score is restricted to a real number between 0 and 1 using the sigmoid activation function

    Ensemble binding scores: This process is to map the binding scores obtained from the two models into a fusion space through a linear layer network. The n the output final binding score is restricted to a real number between 0 and 1 using the sigmoid activation function. Sequences with binding scores above the 0.5 threshold are co nsidered to have binding ...

    work page 2017

  5. [5]

    Tuerk, C. and L. Gold, Systematic evolution of ligands by exponential enrichment: RNA ligands to bacteriophage T4 DNA polymerase. Science, 1990. 249(4968): p. 505-10

    work page 1990

  6. [6]

    Anal Chem, 2020

    Xue, Y., et al., Label-Free and Regenerable Aptasensor for Real -Time Detection of Cadmium(II) by Dual Polarization Interferometry. Anal Chem, 2020. 92(14): p. 10007 - 10015

    work page 2020

  7. [7]

    J Pharm Biomed Anal, 2020

    Mansouri, A., et al., In vitro selection of tacrolimus binding aptamer by systematic evolution of ligands by exponential enrichment method for the development of a fluorescent aptasensor for sensitive detection of tacrolimus. J Pharm Biomed Anal, 2020. 177: p. 112853

    work page 2020

  8. [8]

    Anal Biochem, 2022

    Martinez-Roque, M.A., et al., DNA aptamer selection for SARS-CoV-2 spike glycoprotein detection. Anal Biochem, 2022. 645: p. 114633

    work page 2022

Show all 37 references
  1. [9]

    Mol Ther Nucleic Acids, 2022

    Wu, Y.Y., et al., A novel DNA aptamer targeting lung cancer stem cells exerts a therapeutic effect by binding and neutralizing Annexin A2. Mol Ther Nucleic Acids, 2022. 27: p. 956- 968

    2022

  2. [10]

    Anal Chem, 2021

    Li, L., et al., Identification of a New DNA Aptamer by Tissue-SELEX for Cancer Recognition and Imaging. Anal Chem, 2021. 93(19): p. 7369-7377

    2021

  3. [11]

    Biomicrofluidics, 2018

    Hung, L.Y., et al., Microfluidic platforms for rapid screening of cancer affinity reagents by using tissue samples. Biomicrofluidics, 2018. 12(5): p. 054108

    2018

  4. [12]

    Chem Commun (Camb), 2014

    Nguyen, V.T., et al., Multiple GO-SELEX for efficient screening of flexible aptamers. Chem Commun (Camb), 2014. 50(72): p. 10513-6

    2014

  5. [13]

    Nat Biotechnol, 2013

    Kimoto, M., et al., Generation of high-affinity DNA aptamers using an expanded genetic alphabet. Nat Biotechnol, 2013. 31(5): p. 453-7

    2013

  6. [14]

    Front Cell Dev Biol, 2023

    Chen, Z., et al., Chemically modified aptamers for improving binding affinity to the target proteins via enhanced non-covalent bonding. Front Cell Dev Biol, 2023. 11: p. 1091809

    2023

  7. [15]

    Nucleic Acids Res, 2019

    Xu, G., et al., Structure-guided post-SELEX optimization of an ochratoxin A aptamer. Nucleic Acids Res, 2019. 47(11): p. 5963-5972

    2019

  8. [16]

    Nucleic Acids Res, 2020

    Ishida, R., et al., RaptRanker: in silico RNA aptamer selection from HT-SELEX experiment based on local sequence and structure information. Nucleic Acids Res, 2020. 48(14): p. e82

    2020

  9. [17]

    Roche -Lima, and E.I

    Heredia, F.L., A. Roche -Lima, and E.I. Pares -Matos, A novel artificial intelligence -based approach for identification of deoxynucleotide aptamers. PLoS Comput Biol, 2021. 17(8): p. e1009247

    2021

  10. [18]

    BMC Bioinformatics, 2020

    Li, J., et al., PPAI: a web server for predicting protein -aptamer interactions. BMC Bioinformatics, 2020. 21(1): p. 236

    2020

  11. [19]

    Emami, N. and R. Ferdousi, AptaNet as a deep learning approach for aptamer -protein interaction prediction. Sci Rep, 2021. 11(1): p. 6074

    2021

  12. [20]

    Nat Commun, 2023

    Wang, W., et al., trRosettaRNA: automated prediction of RNA 3D structure with transformer network. Nat Commun, 2023. 14(1): p. 7266

    2023

  13. [21]

    Research (Wash D C), 2023

    Mou, M., et al., A Transformer-Based Ensemble Framework for the Prediction of Protein- Protein Interaction Sites. Research (Wash D C), 2023. 6: p. 0240

    2023

  14. [22]

    Yamada, K. and M. Hamada, Prediction of RNA–protein interactions using a nucleotide language model. Bioinformatics Advances, 2022. 2(1)

    2022

  15. [23]

    BMC Bioinformatics, 2023

    Shin, I., et al., AptaTrans: a deep neural network for predicting aptamer-protein interaction using pretrained encoders. BMC Bioinformatics, 2023. 24(1): p. 447

    2023

  16. [24]

    IEEE Trans Pattern Anal Mach Intell, 2022

    Elnaggar, A., et al., ProtTrans: Toward Understanding the Language of Life Through Self- Supervised Learning. IEEE Trans Pattern Anal Mach Intell, 2022. 44(10): p. 7112-7127

    2022

  17. [25]

    arXiv preprint arXiv:2204.00300, 2022

    Chen, J.a.H., Zhihang and Sun, Siqi and Tan, Qingxiong and Wang, Yixuan and Yu, Qinze and Zong, Licheng and Hong, Liang and Xiao, Jin and King, Irwin and others, Interpretable rna foundation model from unannotated data for highly accurate rna structure and function predictions...

    2022 arXiv

  18. [26]

    Vinay Bachu, L.D., Ayush Kumar, Pooja Rani Kuri, Malaya Mili, Naveen Kumar Singh, & Pranab Goswami, Aptabase: An aptamer database. 2021

    2021

  19. [27]

    Nucleic Acids Res, 2023

    Askari, A., et al., UTexas Aptamer Database: the collection and long-term preservation of aptamer sequence information. Nucleic Acids Res, 2023

    2023

  20. [28]

    Nucleic Acids Res, 2021

    UniProt, C., UniProt: the universal protein knowledgebase in 2021. Nucleic Acids Res, 2021. 49(D1): p. D480-D489

    2021

  21. [29]

    Nucleic Acids Res, 2000

    Berman, H.M., et al., The Protein Data Bank. Nucleic Acids Res, 2000. 28(1): p. 235-42

    2000

  22. [30]

    Gupte, and R

    Jain, D.S., S.R. Gupte, and R. Aduri, A Data Driven Model for Predicting RNA -Protein Interactions based on Gradient Boosting Machine. Sci Rep, 2018. 8(1): p. 9552

    2018

  23. [31]

    Nucleic Acids Res, 2006

    Wu, T., et al., NPInter: the noncoding RNAs and protein related biomacromolecules interaction database. Nucleic Acids Res, 2006. 34(Database issue): p. D150-2

    2006

  24. [32]

    Database (Oxford), 2016

    Hao, Y., et al., NPInter v3.0: an upgraded database of noncoding RNA -associated interactions. Database (Oxford), 2016. 2016

    2016

  25. [33]

    Vaswani, A., et al., Attention is all you need. 2017. 30

    2017

  26. [34]

    Zhang, and K.C

    Ding, Y.S., T.L. Zhang, and K.C. Chou, Prediction of protein structure classes with pseudo amino acid composition and fuzzy support vector machine network. Protein Pept Lett,

  27. [35]

    Huang, J. and A.D. MacKerell, Jr., CHARMM36 all -atom additive protein force field: validation based on comparison to NMR data. J Comput Chem, 2013. 34(25): p. 2135-45

    2013

  28. [36]

    Gállego, and M.R.J.a.p.a

    Ferrando, J., G.I. Gállego, and M.R.J.a.p.a. Costa-jussà, Measuring the mixing of contextual information in the transformer. 2022

    2022

  29. [37]

    Reuter, J.S. and D.H. Mathews, RNAstructure: software for RNA secondary structure prediction and analysis. BMC Bioinformatics, 2010. 11: p. 129

    2010

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.