The DOI matches its Crossref record; 36 of the 39 references it cites were found in a registry or an index.
Scoring alignments by embedding vector similarity
Open accessPreprint — published as 10.1093/bib/bbae178
Abstract
Sequence similarity is of paramount importance in biology, as similar sequences tend to have similar function and share common ancestry. Scoring matrices, such as PAM or BLO-SUM, play a crucial role in all bioinformatics algorithms for identifying similarities, but have the drawback that they are fixed, independent of context. We propose a new scoring method for amino acid similarity that remedies this weakness, being contextually dependent. It relies on recent advances in deep learning architectures that employ self-supervised learning in order to leverage the power of enormous amounts of unlabelled data to generate contextual embeddings, which are vector representations for words. These ideas have been applied to protein sequences, producing embedding vectors for protein residues. We propose the E-score between two residues as the cosine similarity between their embedding vector representations. Thorough testing on a wide variety of reference multiple sequence alignments indicate that the alignments produced using the new E -score method, especially ProtT5-score, are significantly better than those obtained using BLOSUM matrices. The new method proposes to change the way alignments are computed, with far reaching implications in all areas of textual data that use sequence similarity. The program to compute alignments based on various E -scores is available as a web server at e-score.csd.uwo.ca . The source code is freely available for download from github.com/lucian-ilie/E-score .
Overview
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404 · cited 1×
“The most successful contextual embeddings rely on the transformer architecture [32], achieving state-of-the-art results in many NLP benchmarks [7, 19, 34, 18].”
BackgroundBasic local alignment search tool
Altschul, S. F., Gish, W., Miller, W., Myers, E. W., & Lipman, D. J. (1990) · Journal of Molecular Biology · doi:10.1006/jmbi.1990.9999 · Cited by 98,669 · cited 1×
“Sequence similarity search is the most widely used procedure in bioinformatics, the BLAST program [2, 3] being one of the most cited contributions in the history of science.”
MethodCochat, P., Vaucoret, L., & Sarles, J. (2012) · Archives de pédiatrie · doi:10.1016/j.arcped.2012.01.013 · Cited by 77,101 · cited 5×
“We have thoroughly tested the new E -score in the context of protein alignments, by comparing the top scoring function, ProtT5-score, against the BLOSUM45 matrix on a large set including a wide variety of reference multiple sequence alignments from NCBI’s Conserved Domain Database.”
MethodInfluentialGapped BLAST and PSI-BLAST: a new generation of protein database search programs
Altschul, S. (1997) · Nucleic Acids Research · doi:10.1093/nar/25.17.3389 · Cited by 73,302 · cited 2×
“1 The most important component of similarity search algorithms, such as BLAST, is the function that describes the similarity scores of various components, such as nucleotides for DNA and amino acids for proteins.”
BackgroundMethodBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325 · cited 2×
“Protein sequences present a similar situation, with entire proteins seen as sentences and residues as words, enabling the use of the embedding techniques developed for NLP to take advantage of the vast amount of protein sequences from UniProt [1].”
BackgroundMethodHighly accurate protein structure prediction with AlphaFold
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., et al. (2021) · Nature · doi:10.1038/s41586-021-03819-2 · Cited by 43,764
Glove: Global Vectors for Word Representation
Pennington, J., Socher, R., & Manning, C. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1162 · Cited by 34,071 · cited 1×
“Embeddings such as word2vec [21] or GloVe [24] are contextual independent as the embedding vector for a given word is the same regardless of context.”
BackgroundEfficient Estimation of Word Representations in Vector Space
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013) · arXiv · doi:10.48550/arxiv.1301.3781 · Cited by 26,568 · cited 3×
“Embeddings such as word2vec [21] or GloVe [24] are contextual independent as the embedding vector for a given word is the same regardless of context.”
BackgroundMethodRoBERTa: A Robustly Optimized BERT Pretraining Approach
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019) · arXiv · doi:10.48550/arxiv.1907.11692 · Cited by 16,515 · cited 1×
“The most successful contextual embeddings rely on the transformer architecture [32], achieving state-of-the-art results in many NLP benchmarks [7, 19, 34, 18].”
BackgroundDeep Contextualized Word Representations
Peters, M., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018) · Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) · doi:10.18653/v1/n18-1202 · Cited by 11,983 · cited 1×
“The next generation of embeddings, called contextual embeddings, overcome this limitation, e.g., ELMo [25], BERT [7], RoBERTa [19], XLNet [34], T5 [26].”
BackgroundNeedleman, S. B., & Wunsch, C. D. (1970) · Journal of Molecular Biology · doi:10.1016/0022-2836(70)90057-4 · Cited by 11,288 · cited 1×
“…less than 300, then we pick all pairs), compute pair-wise alignment of each pair using either the E -score or BLO-SUM matrix, according to the Needleman-Wunsch dynamic programming global alignment algorithm [23], but without any penalties for gaps at the ends; we call this end-gap-free alignment.”
MethodExploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929 · cited 1×
“The next generation of embeddings, called contextual embeddings, overcome this limitation, e.g., ELMo [25], BERT [7], RoBERTa [19], XLNet [34], T5 [26].”
BackgroundUniProt: the Universal Protein knowledgebase
Apweiler, R. (2003) · Nucleic Acids Research · doi:10.1093/nar/gkh131 · Cited by 7,593
XLNet: Generalized Autoregressive Pretraining for Language Understanding
Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R., & Le, Q. V. (2019) · arXiv · doi:10.48550/arxiv.1906.08237 · Cited by 6,687 · cited 2×
“The next generation of embeddings, called contextual embeddings, overcome this limitation, e.g., ELMo [25], BERT [7], RoBERTa [19], XLNet [34], T5 [26].”
BackgroundAmino acid substitution matrices from protein blocks.
Henikoff, S., & Henikoff, J. G. (1992) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.89.22.10915 · Cited by 6,218 · cited 1×
“We consider here the much more complicated case of proteins, where scoring of alignments is done using PAM [6] or BLOSUM [11] matrices, which give the substitution scores for pairs of amino acids.”
THE DISTRIBUTION OF THE FLORA IN THE ALPINE ZONE. 1
Jaccard, P. (1912) · New Phytologist · doi:10.1111/j.1469-8137.1912.tb05611.x · Cited by 4,561 · cited 1×
“For example, . d ssp is defined as the Jaccard distance [14] on the homology sets H ssp : .”
BackgroundUniProt: the universal protein knowledgebase
The UniProt Consortium (2017) · Nucleic Acids Research · doi:10.1093/nar/gkw1099 · Cited by 4,436
ALBERT: A Lite BERT for Self-supervised Learning of Language Representations
Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., & Soricut, R. (2019) · arXiv · doi:10.48550/arxiv.1909.11942 · Cited by 4,308 · cited 1×
“The most successful contextual embeddings rely on the transformer architecture [32], achieving state-of-the-art results in many NLP benchmarks [7, 19, 34, 18].”
BackgroundCDD: NCBI's conserved domain database
Marchler‐Bauer, A., Derbyshire, M. K., Gonzales, N. R., Lu, S., Chitsaz, F., Geer, L. Y., Geer, R. C., He, J., Gwadz, M., Hurwitz, D. I., et al. (2014) · Nucleic Acids Research · doi:10.1093/nar/gku1221 · Cited by 3,468
Improved protein structure prediction using potentials from deep learning
Senior, A. W., Evans, R., Jumper, J., Kirkpatrick, J., Sifre, L., Green, T., Qin, C., Žídek, A., Nelson, A. W. R., Bridgland, A., et al. (2020) · Nature · doi:10.1038/s41586-019-1923-7 · Cited by 3,378
Rives, A., Meier, J., Sercu, T., Goyal, S., Lin, Z., Liu, J., Guo, D., Ott, M., Zitnick, C. L., Ma, J., et al. (2019) · doi:10.1101/622803 · Cited by 3,280
ProtTrans: Toward Understanding the Language of Life Through Self-Supervised Learning
Elnaggar, A., Heinzinger, M., Dallago, C., Rehawi, G., Wang, Y., Jones, L., Gibbs, T., Feher, T., Angerer, C., Steinegger, M., et al. (2022) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2021.3095381 · Cited by 2,329 · cited 1×
“Many protein embeddings have been proposed, including ProtVec [4], SeqVec [10], PRoBERTa [22], MSA-transformer [27], ESM-1b and ESM2 [28], and the models in the ProtTrans project: ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, and ProtT5 [8].”
MethodImproved protein structure prediction using predicted interresidue orientations
Yang, J., Anishchenko, I., Park, H., Peng, Z., Ovchinnikov, S., & Baker, D. (2020) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.1914677117 · Cited by 1,421 · cited 1×
“These protein embeddings have many important applications to various areas, such as: structure prediction [29, 33, 15], function prediction [16, 9, 17], interaction site prediction [13, 35, 12], etc.”
BackgroundStructure-based protein function prediction using graph convolutional networks
Gligorijević, V., Renfrew, P. D., Kościółek, T., Leman, J. K., Berenberg, D., Vatanen, T., Chandler, C., Taylor, B. C., Fisk, I., Vlamakis, H. C., et al. (2021) · Nature Communications · doi:10.1038/s41467-021-23303-9 · Cited by 961 · cited 1×
“These protein embeddings have many important applications to various areas, such as: structure prediction [29, 33, 15], function prediction [16, 9, 17], interaction site prediction [13, 35, 12], etc.”
BackgroundModeling aspects of the language of life through transfer-learning protein sequences
Heinzinger, M., Elnaggar, A., Wang, Y., Dallago, C., Nechaev, D., Matthes, F., & Rost, B. (2019) · BMC Bioinformatics · doi:10.1186/s12859-019-3220-8 · Cited by 606 · cited 1×
“Many protein embeddings have been proposed, including ProtVec [4], SeqVec [10], PRoBERTa [22], MSA-transformer [27], ESM-1b and ESM2 [28], and the models in the ProtTrans project: ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, and ProtT5 [8].”
MethodDeepGOPlus: improved protein function prediction from sequence
Kulmanov, M., & Hoehndorf, R. (2019) · Bioinformatics · doi:10.1093/bioinformatics/btz595 · Cited by 492 · cited 1×
“These protein embeddings have many important applications to various areas, such as: structure prediction [29, 33, 15], function prediction [16, 9, 17], interaction site prediction [13, 35, 12], etc.”
BackgroundBAliBASE 3.0: Latest developments of the multiple sequence alignment benchmark
Thompson, J., Koehl, P., Ripp, R., & Poch, O. (2005) · Proteins: Structure, Function, and Bioinformatics · doi:10.1002/prot.20527 · Cited by 452 · cited 1×
“Most of these databases include comparatively similar sequences [31], albeit with the addition of features that are often difficult to align such as repeats, and many of the benchmarks were created using simulation techniques [30].”
BackgroundRao, R., Liu, J., Verkuil, R., Meier, J., Canny, J. F., Abbeel, P., Sercu, T., & Rives, A. (2021) · doi:10.1101/2021.02.12.430858 · Cited by 350 · cited 1×
“Many protein embeddings have been proposed, including ProtVec [4], SeqVec [10], PRoBERTa [22], MSA-transformer [27], ESM-1b and ESM2 [28], and the models in the ProtTrans project: ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, and ProtT5 [8].”
Method(2026) · Canadian Journal of Emergency Medicine · doi:10.1007/s43678-026-01115-4 · Cited by 303 · cited 1×
“Many protein embeddings have been proposed, including ProtVec [4], SeqVec [10], PRoBERTa [22], MSA-transformer [27], ESM-1b and ESM2 [28], and the models in the ProtTrans project: ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, and ProtT5 [8].”
Rose: generating sequence families
Stoye, J., Evers, D., & Meyer, F. (1998) · Bioinform. · doi:10.1093/bioinformatics/14.2.157 · Cited by 301 · cited 1×
“Most of these databases include comparatively similar sequences [31], albeit with the addition of features that are often difficult to align such as repeats, and many of the benchmarks were created using simulation techniques [30].”
BackgroundStructure-aware protein-protein interaction site prediction using deep graph convolutional network
Yuan, Q., Chen, J.-W., Zhao, H.-Y., Zhou, Y., & Yang, Y. (2021) · Bioinform. · doi:10.1093/bioinformatics/btab643 · Cited by 174 · cited 1×
“These protein embeddings have many important applications to various areas, such as: structure prediction [29, 33, 15], function prediction [16, 9, 17], interaction site prediction [13, 35, 12], etc.”
BackgroundLai, B.-Q., & Xu, J.-B. (2021) · bioRxiv · doi:10.1101/2021.06.16.448727 · Cited by 136 · cited 1×
“These protein embeddings have many important applications to various areas, such as: structure prediction [29, 33, 15], function prediction [16, 9, 17], interaction site prediction [13, 35, 12], etc.”
BackgroundTransforming the Language of Life
Nambiar, A., Heflin, M., Liu, S., Maslov, S., Hopkins, M., & Ritz, A. (2020) · ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics (ACM-BCB) · doi:10.1145/3388440.3412467 · Cited by 104 · cited 1×
“Many protein embeddings have been proposed, including ProtVec [4], SeqVec [10], PRoBERTa [22], MSA-transformer [27], ESM-1b and ESM2 [28], and the models in the ProtTrans project: ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, and ProtT5 [8].”
MethodMeasuring the distance between multiple sequence alignments
Blackburne, B. P., & Whelan, S. (2011) · Bioinformatics · doi:10.1093/bioinformatics/btr701 · Cited by 75 · cited 4×
“We use five distance metrics, three existing ones, d ssp , d seq , and d pos from [5], and two that we introduce here, d cc and d d .”
BackgroundMethodInfluentialSeq-InSite: sequence supersedes structure for protein interaction site prediction
Hosseini, S., Golding, G. B., & Ilie, L. (2023) · doi:10.1101/2023.06.19.545575 · Cited by 26
Self-citationPITHIA: Protein Interaction Site Prediction Using Multiple Sequence Alignments and Attention
Hosseini, S., & Ilie, L. (2022) · International Journal of Molecular Sciences · doi:10.3390/ijms232112814 · Cited by 20 · cited 1×
“These protein embeddings have many important applications to various areas, such as: structure prediction [29, 33, 15], function prediction [16, 9, 17], interaction site prediction [13, 35, 12], etc.”
BackgroundSelf-citation
Flagged first, then by citations
3 of the 39 references couldn’t be matched to a known paper.
Research works printed but not linked: 4
- 10.1371/journal.pone.0141287 · DOI
- 1978, 22 a model of evolutionary change in proteins, Atlas of protein sequence and structure, 5, 345
- 10.1186/s12859-019-2669-9 · DOI
- 10.1073/pnas.2016239118 · DOI · Paper page
Cited by
Sorted by citations
GPR17 - orphan G protein-coupled receptor with therapeutic potential.
Lewash, M., Kostenis, E., & Müller, C. E. (2025) · TIPS - Trends in Pharmacological Sciences · doi:10.1016/j.tips.2025.05.001 · Cited by 17
Guan, J., Yao, L., Xie, P., Zhao, Z., Meng, D., Lee, T.-Y., Wang, J., & Chiang, Y. (2025) · Briefings in Bioinformatics · doi:10.1093/bib/bbaf292 · Cited by 7
“For protein sequences, we used the BLOSUM62 substitution matrix with gap open = –10 and gap extension = –0.5 [34, 35].”
InfluentialEnhancing Data Traceability: A Knowledge Graph Approach with Retrieval-Augmented Generation
Tjokro, V. C., & Sanjaya, S. A. (2024) · 2024 7th International Seminar on Research of Information Technology and Intelligent Systems (ISRITI) · doi:10.1109/isriti64779.2024.10963652 · Cited by 5
“Cosine similarity is calculated based on the formula on equation (2) [21]. where, A and B , each represent a vector data point θ (theta), represent angle between two data point”
Protein embeddings and local alignments
Malec, J., Golding, G. B., & Ilie, L. (2026) · Computational and Structural Biotechnology Journal · doi:10.1016/j.csbj.2025.12.002 · Cited by 5
“…for comparison include the standard for scoring protein alignments, the BLOSUM matrices, which we feel should be replaced by the new method, specialized matrices such as GPCRtm, existing top methods, PEbA [17], ProtT5-score [5], DEDAL [25] and pLM-BLAST [20], as well as the ProstT5-score.”
InfluentialSelf-citationA comprehensive systematic literature review of multiple sequence alignment algorithms
Sabonsolin, J. A., & Lao, D. (2026) · Discover Computing · doi:10.1007/s10791-026-09911-3 · Cited by 4
“The use of embedding vector similarity for scoring alignments is another innovative approach that could lead to more accurate results [83].”
ESM-MHC: An Improved Predictor of MHC Using ESM Protein Language Model
Cai, J., Li, Y., & Chen, D. (2024) · ICBBT · doi:10.1145/3674658.3674674 · Cited by 3
“Pro-tXLNet applies this technique to protein sequence analysis to more effectively handle long-range dependencies in protein sequences [19., 20., 21.].”
Ankh‐Score Produces Better Sequence Alignments Than AlphaFold3
Malec, J., Rusen, K., Golding, G. B., & Ilie, L. (2026) · Proteins: Structure, Function, and Bioinformatics · doi:10.1002/prot.70143 · Cited by 1
“Precisely, the Ankh-score [22, 23] for two residues a 1 and a 2 is given by the cosine similarity between the vectors v i = Ankh ( a i ) , i = 1, 2 , computed by Ankh for the two residues, that is: We prove Ankh to be the best PLM for this task.”
InfluentialSelf-citationDanwada, S., Udomprasert, P., Raychawdhary, N., Seals, C. D., Wu, L., & Bhattacharya, S. (2026) · bioRxiv · doi:10.64898/2026.07.31.742099
“Existing studies largely evaluate PLMs through remote-homology classification [4], [33], residue-level embedding alignment [3], [34]–[37], or end-to-end structure prediction [18] rather than through continuous pairwise structural-similarity scoring across multiple models and embedding-space metrics.”
Hurmach, V., Motuziuk, O., Prylutska, S., Budzanivska, I., & Prylutskyy, Y. (2026) · Journal of Molecular Graphics and Modelling · doi:10.1016/j.jmgm.2026.109529
Oh, S.-H., Park, D., & Song, Y.-K. (2026) · Pharmacoepidemiology and Drug Safety · doi:10.1002/pds.70386
Metadata
- Authors
- Ashrafzadeh, S., Golding, G. B., Ilie, S., & Ilie, L.
- Affiliations
- Western University (CA)
- McMaster University (CA)
- Corresponding author
- Ilie, Lucian
- Venue
- Briefings in Bioinformatics
- Publisher
- openRxiv
- Published
- 2023-09-01
- Volume
- 25
- Issue
- 3
- ISSN
- 1467-5463 (ISSN-L)
- PMC
- PMC11063651
- OpenAlex
- W4396593457
- Cited by
- 8
- References
- 39
- Citations by year
- 2026: 4 · 2025: 3 · 2024: 1
- FWCI
- 1.09
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- Locations
- Briefings in Bioinformatics (publishedVersion, cc-by)
- bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- Funders
- Compute Canada
- Natural Sciences and Engineering Research Council of Canada (RGPIN 2020, rgpin-2020-05733, rgpin-2021-03978, rgpin-2020-05469)
- Topics
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Phylogenetic Studies
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Fractal and DNA sequence analysis
- Keywords
- amino acid similarity, sequence alignment scoring, contextual embeddings, self-supervised learning, ProtT5, cosine similarity, BLOSUM, PAM, multiple sequence alignment, protein residues
- MeSH terms
- Deep Learning, Algorithms, Amino Acid Sequence, Proteins, Software, Sequence Alignment, Computational Biology, Sequence Analysis, Protein, Databases, Protein
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
