standOff. When we have changed our mind. Some entries in the table begin with a word in capitals and the reason for it. EVIDENCE CORRECTED — the finding still stands, but we had quoted the wrong sentence from the paper, and the quotation has been replaced with the right one. RETRACTED — the entry itself was wrong; it stays on the page, marked, with the reason, rather than being deleted. WITHDRAWN AND REPLACED — the entry was wrong, and a corrected version of it appears in the entries below. Each mark carries its own reason, in brackets, because the reasons differ from one entry to the next. An entry we have WITHDRAWN is removed from this page and kept in the record, with its reason. A line reading Audit and a date is a later re-reading of the entry against its whole sentence: DOUBTFUL means it was kept, and the reason says what is loose about it. Quotation marks inside the tagger's sentence. Where a tag is the tagger's own sentence, a phrase in quotation marks inside it is borrowed from the paper — the tagger's wording around it is not. The paper's own sentence is the evidence quote shown on the same row, so the borrowing can be checked against the thing borrowed from without leaving this page. Nothing inside quotation marks anywhere on this page is ours: if it is quoted, it is the paper's. Whose words a concept carries. A concept entry records a term the paper itself uses, so the term, the sense given for it, and its evidence are the paper's own words, quoted where the paper fixes them. This is the opposite of a sense entry in the tag table, which is the tagger's sentence about how the paper uses a term. Where a concept entry could not be grounded in the paper's words, the entry says so on its own row.| second reading | alternatives | an independent re-reading of a facet, recorded whether or not it agrees |
| relation to other work | claim_relations | how this paper stands to another: agrees, contradicts, extends, qualifies |
| claim type | claim_type | the shape of the claim: existence, mechanism, method, critique, prediction, null |
| concept edge | concept_edges | an asserted relation between two concepts, with the sentence asserting it |
| concept | concepts | an idea the paper works with, quoted where the paper uses it |
| content tag | content_tags | a finding, definition, quantity, statistic or variable named in the paper |
| control | controls | an experiment run to eliminate a rival explanation |
| limitation | limitations | what the paper says its evidence cannot support |
| mechanism | mechanisms | how the reported effect is said to come about |
| method | method | how the paper produced its evidence |
| object of study | object_of_study | what the paper is about — the thing it investigates |
| phenomenon | phenomenon | the effect or regularity under study |
| sense | senses | a distinct use of a term, quoted where the paper fixes it |
| transfer | transfers | a method or concept carried in from another field |
| unrecordable relation | unhoused_relations | a relation the paper asserts that this vocabulary cannot express, with the reason |
The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.
Abstract
claim_typeWe introduce a new representation and feature extraction method for biological sequences. concept_edgesconcept_edgescontent_tags↗Named bio-vectors (BioVec) to refer to biological sequences in general with protein-vectors (ProtVec) for proteins (amino-acid sequences) and gene-vectors (GeneVec) for gene sequences, this representation can be widely used in applications of deep learning in proteomics and genomics. In the present paper, we focus on protein-vectors that can be utilized in a wide array of bioinformatics investigations such as family classification, protein visualization, structure prediction, disordered protein identification, and protein-protein interaction prediction. method↗alternativesIn this method, mechanisms↗content_tagswe adopt artificial neural network approaches and represent a protein sequence with a single dense n-dimensional vector. content_tagsTo evaluate this method, we apply it in classification of 324,018 protein sequences obtained from Swiss-Prot belonging to 7,027 protein families, where content_tagscontent_tagscontent_tags↗an average family classification accuracy of 93%±0.06% is obtained, outperforming existing family classification methods. In addition, we use ProtVec representation to predict disordered proteins from structured proteins. Two databases of disordered sequences are used: the DisProt database as well as a database featuring the disordered regions of nucleoporins rich with phenylalanine-glycine repeats (FG-Nups). limitationscontent_tagsUsing support vector machine classifiers, FG-Nup sequences are distinguished from structured protein sequences found in Protein Data Bank (PDB) with a 99.8% accuracy, and unstructured DisProt sequences are differentiated from structured DisProt sequences with 100.0% accuracy. These results indicate that by only providing sequence data for various proteins into this model, accurate information about protein structure can be determined. content_tagsImportantly, this model needs to be trained only once and can then be applied to extract a comprehensive set of information regarding proteins of interest. Moreover, this representation can be considered as pre-training for various applications of deep learning in bioinformatics. The related data is available at Life Language Processing Website: http://llp.berkeley.edu and Harvard Dataverse: http://dx.doi.org/10.7910/DVN/JMFHTN .
Introduction
concepts↗concepts↗unhoused_relationssensesNature uses certain languages to describe biological sequences such as DNA, RNA, and proteins. Much like humans adopt languages to communicate, biological organisms use sophisticated languages to convey information within and between cells. object_of_studyInspired by this conceptual analogy, claim_relationswe adopt existing methods in natural language processing (NLP) to gain a deeper understanding of the “language of life” with the ultimate goal to discover functions encoded within biological sequences [ 1 – 4 ].
content_tags↗Feature extraction is an important step in data analysis, machine learning and NLP. It refers to finding an interpretable representation of data for machines that can increase performance of learning algorithms. Even the most sophisticated algorithms would perform poorly if inappropriate features are used, while simple methods can potentially perform well when they are fed with the appropriate features. Feature extraction can be done manually or in an unsupervised fashion. phenomenonIn this paper, we propose an unsupervised data-driven distributed representation for biological sequences. This method, called bio-vectors (BioVec) in general and more specifically protein-vectors (ProtVec) for proteins, can be applied to a wide range of problems in bioinformatics, such as protein visualization, protein family classification, structure prediction, domain extraction, and interaction prediction. In this approach, each biological sequence is embedded in an n-dimensional vector that characterizes biophysical and biochemical properties of sequences using neural networks. In the following, we first explain how this method works and how it is trained from 546,790 sequences of Swiss-Prot database. Subsequently, we will analyze the biophysical and the biochemical properties of this representation qualitatively and quantitatively. To further evaluate this feature extraction method, we apply it in classification of 7,027 protein families of 324,018 protein sequences in Swiss-Prot. In the next step, we use this approach for visualization and characterization of two categories of disordered sequences: the DisProt database as well as a database of disordered regions of phenylalanine-glycine nucleoporins (FG-Nups). Finally, we classify these protein families using support vector machine (SVM) classifiers [ 5 ]. mechanismsunhoused_relationsAs a key advantage of the proposed method, the embedding needs to be trained only once and then may be used to encode biological sequences in a given problem. The related data and future updated will be available at: http://llp.berkeley.edu and Harvard Dataverse: http://dx.doi.org/10.7910/DVN/JMFHTN .
Distributed Representation
content_tags↗Distributed representation has proved one of the most successful approaches in machine learning [ 6 – 10 ]. concepts↗concepts↗sensesThe main idea in this approach is encoding and storing information about an item within a system through establishing its interactions with other members. transfersDistributed representation was originally inspired by the structure of human memory, where the items are stored in a “content-addressable” fashion. Content-based storing allows for efficiently recalling items from partial descriptions. Since the content-addressable items and their properties are stored within a close proximity, such a system provides a viable infrastructure to generalize features attributed to an item.
sensesContinuous vector representation, transfersas a distributed representation for words, has been recently established in natural language processing (NLP) as an efficient way to represent semantic/syntactic units with many applications. In this model, each word is embedded in a vector in an n-dimensional space. Similar words have close vectors, where similarity is defined in terms of both syntax and semantic. transfersconcepts↗concepts↗unhoused_relationssensesThe basic idea behind training such vectors is that the meaning of a word is characterized by its context, i.e. neighboring words. Thus, words and their contexts are considered to be positive training samples. claim_relationsSuch vectors can be trained using large amounts of textual data in a variety of ways, e.g. neural network architectures like the Skip-gram model [ 10 ].
transfersInteresting patterns have been observed by training word vectors using Skip-gram in natural language. Words with similar vector representations show multiple degrees of similarity. For instance, K i n g → - M a n → + W o m a n → resembles the closest vector to the word Q u e e n → [ 11 ].
In this work, we seek unique patterns in biological sequences to facilitate biophysical and biochemical interpretations. claim_relationsWe show how Skip-gram can be used to train a distributed representation for biological sequences over a large set of sequences, and establish physical and chemical interpretations for such representations. We propose this as a general-purpose representation for protein sequences that can be used in a wide range of bioinformatics problems, including protein family classification, protein interaction prediction, structure prediction, motif extraction, protein visualization, and domain identification. To illustrate, we specifically tackle visualization and protein family classification problems.
Protein Family Classification
concepts↗sensescontent_tags↗A protein family is a set of proteins that are evolutionarily related, typically involving similar structures or functions. The large gap between the number of known sequences versus the amount of known functional information about sequences has motivated family (function) identification methods based on primary sequences [ 12 – 14 ]. Protein Family Database (Pfam) is a widely used source for protein families [ 15 ]. sensesIn Pfam, a family can be classified as a “family”, “domain”, “repeat”, or “motif”. concept_edgescontent_tags↗In this study, we utilize ProtVec to classify protein families in Swiss-Prot using the information provided by Pfam database and we obtain a high classification accuracy.
Protein family classification based on the primary structures (sequences) has been performed using classifiers such as support vector machine classifier (SVM) [ 16 – 18 ]. concepts↗concepts↗sensesclaim_relationsconcept_edgesBesides the primary sequence, the existing methods typically require extensive information for feature extraction, e.g. hydrophobicity, normalized Van der Waals volume, polarity, polarizability, charge, surface tension, secondary structure and solvent accessibility. limitationsThe reported accuracies of a previous study on family classification have been in the range of 69.1–99.6% for 54 protein families [ 16 ]. content_tags↗In another study, researchers used motifs from protein interactions for detecting Structural Classification of Proteins (SCOP) [ 19 ] families for 368 proteins, and obtained a classification accuracy of 75% at sensitivity of 10% [ 20 ]. content_tags↗In contrast, our proposed approach is trained based solely on primary sequence information, yet achieving high accuracy when applied in classifications of protein families.
content_tags↗Disordered Proteins
Proteins can be fully or partially unstructured, i.e. lacking a secondary or ordered tertiary three-dimensional structure. Due to their abundance and the critical roles they play in cell biology, disordered proteins are considered to be an important class of proteins [ 21 ]. Several studies have focused on disordered peptides and their functional analysis in recent years [ 22 – 24 ].
In the present work, we introduce ProtVec for the visualization and characterization of two categories of disordered proteins: DisProt database as well as a database of disordered regions of phenylalanine-glycine nucleoporins (FG-Nups) [ 25 ].
concepts↗concept_edgesDisProt is a database of experimentally identified disordered proteins that categorizes disordered and ordered regions of a collection of proteins [ 26 ]. content_tagsDisProt Release 6.02 consists of 694 proteins presenting 1539 disordered, and 95 ordered regions. FG-Nups dataset is a collection of FG-Nups disordered sequences [ 27 ]. Nucleoporins form the nuclear pore complex (NPC), the sole gateway for bidirectional transport of cargo between the cytoplasm and the nucleus in eukaryotic cells [ 28 ]. limitationslimitationsSince FG-Nups are mostly computationally identified, only 10 sequences out of 1,138 disordered sequences exist in Swiss-Prot. concept_edgesA recent study on features of FG-Nups versus DisProt showed biophysical differences between FG-Nups and average DisProt sequences [ 29 ].
We further propose using protein-vectors for the visualization of biological sequences. Simplicity and biophysical interpretations encoded within ProtVec distinguishes this method from the previous work [ 30 , 31 ]. As an example, we use ProtVec for the visualization of FG-Nups, DisProt, and structured PDB proteins. claim_relationsclaim_relationsThis visualization confirms the results obtained [ 29 ] on the biophysical features of FG-Nups and typical disordered proteins. Furthermore, we employ ProtVec to classify FG-Nups versus random PDB sequences as well as DisProt disordered regions versus disport ordered regions.
Methods
Protein-Space Construction
Our goal is to construct a distributed representation of biological sequences. In the training process of word embedding in NLP, a large corpus of sentences should be fed into the training algorithm to ensure sufficient contexts are observed. Similarly, a large corpus is needed to train distributed representation of biological sequences. content_tagsWe use Swiss-Prot as a rich protein database, which consists of 546,790 manually annotated and reviewed sequences.
concepts↗concepts↗concepts↗sensesThe next step in training distributed representations is to break the sequences into sub sequences (i.e. biological words). content_tags↗The simplest and most common technique in bioinformatics to study sequences involves fixed-length overlapping n-grams [ 32 – 34 ]. concept_edgesHowever, instead of using n-grams directly in feature extraction, we utilize n-gram modeling for training a general purpose distributed representation of sequences. This so-called embedding model needs to be trained only once and may then be adopted in feature extraction part of specific problems.
content_tags↗In n-gram modeling of protein informatics, usually an overlapping window of 3 to 6 residues is used. mechanismsInstead of taking overlapping windows, we generate 3 lists of shifted non-overlapping words, as shown in Fig 1 . limitationscontent_tags↗content_tags↗Evaluating K-nearest neighbors in a 2xfold cross-validation for different window sizes, embedding vector sizes and overlapping versus non-overlapping n-grams showed a more consistent embedding training for a window size of 3 and the mentioned splitting.
Fig 1. Protein sequence splitting.
Open in a new tab
concepts↗concepts↗sensesIn order to prepare the training data, content_tags↗each protein sequence will be represented as three sequences (1, 2, 3) of 3-grams.
sensescontent_tagsThe same procedure is applied on all 546,790 sequences in Swiss-Prot, thus at the end we obtain a corpus consisting of 546,790 × 3 = 1,640,370 sequences of 3-grams (content_tags↗3-gram is a “biological” word consisting of 3 amino acids). The next step is training the embedding based on such data through a Skip-gram neural network. In training word vector representations, Skip-gram attempts to maximize the probability of observed word sequences (contexts). In other words, for a given training sequence of words we would like to find their corresponding n-dimensional vectors maximizing the following average log probability function. mechanismsSuch a constraint allows similar words to assume a similar representation in this space.
argmax v , v 1 N ∑ i = 1 N ∑ - c ≤ j ≤ c , j ≠ 0 log p ( w i + j | w i ) p ( w i + j | w i ) = exp ( v ′ w i + j T v w i ) ∑ k = 1 W exp ( v ′ w k T v w i ) , (1)
where N is the length of the training sequence, 2 c is the window size we consider as the context, w i is the center of the window, W is the number of words in the dictionary and v w and v w ′ are input and output n-dimensional representations of word w , respectively. The probability p ( w i + j | w i ) is defined using a softmax function. Hierarchical softmax or negative sampling are efficient approximations of such a softmax function. In the implementation we use (Word2Vec) [ 10 ] negative sampling has been utilized, which is considered as the state-of-the-art for training word vector representation. Negative sampling uses the following objective function in the calculation of the word vectors:
arg max θ ∏ ( w , c ) ∈ D p ( D = 1 | c , w ; θ ) ∏ ( w , c ) ∈ D ′ p ( D = 0 | c , w ; θ ) , (2)
where D is the set of all word and context pairs (w, c) existing in the training data (positive samples) and D ′ is a randomly generated set of incorrect (w, c) pairs (negative samples).
p ( D = 1| w , c ; θ ) is the probability that (w, c) pair came from the training data and p ( D = 0| w , c ; θ ) is the probability that ( w , c ) did not come from the training data. The term p ( D = 1| c , w ; θ ) can be defined using a sigmoid function on the word vectors:
p ( D = 1 | w , c ; θ ) = 1 1 + e - v c · v w ,
where the parameters θ are the word vectors we train within the optimization framework: v c and v w ∈ R d are vector representations for the context c and the word w respectively [ 35 ]. concepts↗concepts↗concepts↗concepts↗unhoused_relationsunhoused_relationssensesIn Eq (2) , the positive samples maximize the probabilities of the observed (w, c) pairs in the training data, while the negative samples prevent all vectors from having the same value by disallowing some incorrect (w, c) pairs. concepts↗To train the embedding vectors, we consider a vector size of 100 and a context size of 25. Thus each 3-gram is presented as a vector of size 100.
Protein-Space Analysis
To qualitatively analyze the distribution of various biophysical and biochemical properties within the training space, content_tagswe project all 3-gram embeddings from 100-dimensional space to a 2D space using Stochastic Neighbor Embedding [ 36 ]. Mass, volume, polarity, hydrophobicity, charge, and van der Waals volume properties were analyzed. The data is adopted from [ 37 ]. In addition, to quantitatively measure the continuity of these properties in the protein-space, the best Lipschitz constant, i.e. the smallest k satisfying is calculated:
d f ( f p r o p ( w 1 ) , f p r o p ( w 2 ) ) ≤ k × d w ( w 1 , w 2 ) , (3)
where f is the scale of one of the properties of a given 3-grams (e.g., average mass, hydrophobicity, etc.), d is the distance metric, d f is the absolute value of score differences, and d w is Euclidian distance between two 3-grams w 1 and w 2 . The Lipschitz constant is calculated for the aforementioned properties.
Protein Family Classification
To evaluate the strength of the proposed representation, we set up a classification task on protein families. Family information of 324,018 protein sequences in Swiss-Prot is extracted from the Protein Family (Pfam) database, resulting in a total of 7,027 distinct families for Swiss-Prot sequences. concepts↗unhoused_relationssensesEach sequence is represented as the summation of the vector representation of overlapping 3-grams. Thus, each sequence is presented as a vector of size 100. concepts↗concepts↗sensesFor each family type, the same number of instances from Swiss-Prot are selected randomly to form the negative examples. Support vector machine classifiers are used to evaluate the strength of ProtVec in the classification of protein families through 10 × fold cross-validations. We perform the classification over 7,027 protein families consisting of 324,018 sequences. concepts↗concepts↗sensesFor the evaluation we report specificity (true negative rate), sensitivity (true positive rate), and the accuracy of family classifications.
S e n s i t i v i t y = T P r a t e = T P T P + F N
S p e c i f i c i t y = 1 - F P r a t e = T N F P + T N
A c c u r a c y = T P + T N P + N
Visualization and Classification of Disordered Proteins
Two databases of disordered proteins are used for disordered protein prediction: DisProt database (694 sequences) and FG-Nups dataset (1,138 sequences).
FG-Nups Characterization
To distinguish the characteristics of FG-Nups, a collection of 1,138 FG-Nups and two random sets of 1,138 structured proteins from Protein Data Bank (PDB) [ 38 ] are compared. Since PDB sequences on average have a shorter length than disordered proteins, the two sets are selected from PDB in such a way that they have an average length of 900 residues, the same as the average length of the disordered protein dataset. For visualization purposes, the ProtVec is reduced from 100 dimensions to a 2D space using Stochastic Neighbor Embedding [ 36 ].
We quantitatively evaluate how ProtVec can be used to distinguish between FG-Nups versus typical PDB sequences using a support vector machine binary classifier. The positive examples were the aforementioned 1,138 disordered FG-Nups proteins and the negative examples (again 1,138 sequences) are selected randomly from PDB with the same average length of disordered sequences (≈ 900 residues). We present each protein sequence as a summation of its ProtVecs of all 3-grams. Since the average length of structured proteins is shorter than FG-Nups, and to avoid trivial cases, the PDB sequences are selected in a way to maintain the same average length.
DisProt Characterization
To distinguish the characteristics of DisProt sequences, we use DisProt Release 6.02, consisting of 694 proteins presenting 1539 disordered and 95 ordered regions, and perform the same experiment as for FG-Nups with DisProt sequences.
Results
Protein-Space Analysis
Although the protein-space is trained based on only the primary sequences of proteins, it offers several interesting biochemical and biophysical implications. In order to study these features, we visualized the distribution of different criteria, including mass, volume, polarity, hydrophobicity, charge, and van der Waals volume in this space. To do so, for each 3-gram we conducted qualitative and quantitative analyses as described below.
Qualitative Analysis
content_tagsIn order to visualize the distribution of the aforementioned properties, we projected all 3-gram embeddings from 100-dimensional space to a 2D space using Stochastic Neighbor Embedding (t-SNE) [ 36 ]. In the diagrams presented in Fig 2 , each point represents a 3-gram and is colored according to its scale in each property. content_tagsInterestingly, as can be seen in the figure, 3-grams with the same biophysical and biochemical properties were grouped together. This observation suggests that the proposed embedding not only encodes protein sequences in an efficient way that proved useful for classification purposes, but also reveals some important physical and chemical patterns in protein sequences.
Fig 2. Normalized distributions of biochemical and biophysical properties in protein-space.
Open in a new tab
In these plots, each point represents a 3-gram (a word of three residues) and the colors indicate the scale for each property. Data points in these plots are projected from a 100-dimensional space a 2D space using t-SNE. As it is shown words with similar properties are automatically clustered together meaning that the properties are smoothly distributed in this space.
Quantitative Analysis
Although Fig 2 illustrates the smoothness of protein-space with respect to different physical and chemical meanings, concepts↗concepts↗unhoused_relationssenseswe required a quantitative approach to measure the continuity of these properties in the protein-space. To do so, we calculated the best Lipschitz constant. For all 6 properties presented in Fig 2 , we calculated the minimum k . mechanismsconcepts↗sensescontent_tags↗To evaluate this result we made an artificial space called “scrambled space” by randomly shuffling the labels of 3-grams in the 100 dimensional space. Table 1 contains the values of Libschitz constants for protein-space versus the “scrambled space” with respect to different properties and also their ratio.
Table 1. Using Lipschitz number to evaluate the continuity of ProtVec with respect to biophysical and biochemical properties.
Lipschitz Number
Property protein-Space The scrambled space Ratio
Mass 0.3137 0.6605 0.4750
Volume 0.3742 0.6699 0.5586
Van Der Waal Volume 0.3629 0.6431 0.5643
Polarity 0.4757 1.2551 0.3790
Hydrophobicity 0.608 1.448 0.4203
Charge 0.8733 1.3620 0.6412
content_tagsAverage 0.50 1.01 0.51
Open in a new tab
concepts↗concepts↗unhoused_relationsunhoused_relationsNormally if k = 1 the function is called a short map, and if 0 ≤ k < 1 the function is called a contraction. concepts↗concepts↗sensesconcept_edgescontent_tagsThe results suggest that the protein-space is on average 2-times smoother in terms of physical and chemical properties than a random space. This quantitative result supports our qualitative observation of the space structure in Fig 2 , and suggests that our training space encodes, 3-grams in an informative manner.
Protein Family Classification
sensescontent_tagsIn order to evaluate the strength of ProtVec, we performed classifications of 7,027 protein families and obtained a weighted average accuracy of 93 ± 0.06%, which exhibits a more reliable result than the existing methods. In contrast to the existing methods, our proposed approach is trained based on primary sequence information alone.
limitationscontent_tags↗Table 2 shows the sensitivity, specificity, and the accuracy for the most frequent families in Swiss-Prot. These results suggest that structural features of proteins can be accurately predicted from the primary sequence information solely. The results for all 7,027 families can be found in Supplementary Information, see S1 File . content_tagsThe average accuracy for the first 1,000 (261,149 sequences), 2,000 (293,957 sequences), 3,000 (308,292 sequences), and 4,000 (316,135 sequences) frequent families were respectively 94% ± 0.05%,93% ± 0.05%, 92% ± 0.06%, and 91% ± 0.08%. limitationslimitationsTo compute the overall accuracy for all 7,026 families, we calculated the weighted average accuracy, because for the families with number of instances less than 10, the validation set are not statistically sufficient and they should have less contribution in the overall accuracy. content_tags↗The weighted accuracy of all 7,027 families (weighted based on the number of instances) was 93% ± 0.06%.
Table 2. Performance of protein family classification using SVM and ProtVec over some of the most frequent families in Swiss-Prot.
Families are sorted with respect to their frequency in Swiss-Prot.
Training instances Classification Result
Family name # of positive sequences # of negative sequences Specificity Sensitivity Accuracy
50S ribosome-binding GTPase 3,084 3,084 0.95 0.93 0.94
Helicase conserved C-terminal domain 2,518 2,518 0.83 0.80 0.82
ATP synthase alpha-beta family, nucleotide-binding domain 2,387 2,387 0.98 0.97 0.97
7 transmembrane receptor (rhodopsin family) 1,820 1,820 0.95 0.96 0.95
Amino acid kinase family 1,750 1,750 0.91 0.92 0.91
ATPase family associated with various cellular activities (AAA) 1711 1711 0.92 0.90 0.91
tRNA synthetases class I (I, L, M and V) 1,634 1,634 0.97 0.97 0.97
tRNA synthetases class II (D, K and N) 1,419 1,419 0.88 0.83 0.85
Major Facilitator Superfamily 1,303 1,303 0.95 0.97 0.96
Hsp70 protein 1,272 1,272 0.97 0.97 0.97
NADH-Ubiquinone-plastoquinone (complex I), various chains 1,251 1,251 0.97 0.97 0.97
Histidine biosynthesis protein 1,248 1,248 0.96 0.97 0.97
TCP-1-cpn60 chaperonin family 1,246 1,246 0.95 0.96 0.95
EPSP synthase (3-phosphoshikimate 1-carboxyvinyltransferase) 1,207 1,207 0.96 0.96 0.96
Aldehyde dehydrogenase family 1,200 1,200 0.93 0.94 0.94
Shikimate–quinate 5-dehydrogenase 1,128 1,128 0.87 0.89 0.88
GHMP kinases N terminal domain 1,120 1,120 0.88 0.92 0.90
Ribosomal protein S2 1,083 1,083 0.95 0.96 0.95
Ribosomal protein S4–S9 N-terminal domain 1,072 1,072 0.95 0.97 0.96
Ribosomal protein L16p-L10e 1,053 1,053 0.95 0.96 0.96
KOW motif 1,047 1,047 0.93 0.95 0.94
Uncharacterized protein family UPF0004 1,044 1,044 0.95 0.97 0.96
Ribosomal protein S12-S23 1,016 1,016 0.94 0.98 0.96
GHMP kinases C terminal 1,011 1,011 0.88 0.92 0.90
Ribosomal protein S14p-S29e 997 997 0.93 0.98 0.95
Ribosomal protein S11 980 980 0.96 0.98 0.97
UvrB-uvrC motif 968 968 0.94 0.96 0.95
Ribosomal protein L33 958 958 0.96 0.98 0.97
BRCA1 C Terminus (BRCT) domain 956 956 0.94 0.95 0.95
RF-1 domain 950 950 0.93 0.97 0.95
Ankyrin repeats (3 copies) 944 944 0.89 0.88 0.88
Ribosomal protein L20 932 932 0.96 0.99 0.97
RNA polymerase beta subunit 912 912 0.94 0.97 0.95
Ribosomal protein S18 908 908 0.93 0.97 0.95
ATP synthase B-B CF(0) 900 900 0.92 0.94 0.93
Peptidase family M20-M25-M40 889 889 0.92 0.93 0.93
Ribosomal protein L18e-L15 887 887 0.93 0.96 0.95
Glucose inhibited division protein A 886 886 0.95 0.96 0.95
NADH-ubiquinone-plastoquinone oxidoreductase chain 4L 885 885 0.94 0.97 0.96
lactate-malate dehydrogenase, NAD binding domain 880 880 0.92 0.94 0.93
HD domain 879 879 0.93 0.93 0.93
Ribosomal protein S10p-S20e 873 873 0.95 0.97 0.96
Pyridoxal-phosphate dependent enzyme 870 870 0.91 0.91 0.91
Ribosomal L18p-L5e family 860 860 0.93 0.96 0.94
Ribosomal protein L3 855 855 0.94 0.97 0.96
tRNA synthetases class I (M) 843 843 0.94 0.96 0.95
UbiA prenyltransferase family 841 841 0.94 0.95 0.95
Ribosomal protein L4–L1 family 841 841 0.94 0.95 0.95
Ribosomal protein S16 840 840 0.93 0.97 0.95
Ribosomal protein S13-S18 840 840 0.94 0.97 0.95
MraW methylase family 837 837 0.95 0.98 0.96
Ribosomal L32p protein family 825 825 0.94 0.97 0.95
Elongation factor TS 819 819 0.94 0.97 0.96
Tetrahydrofolate dehydrogenase-cyclohydrolase, catalytic domain 817 817 0.94 0.96 0.95
ATP synthase delta (OSCP) subunit 813 813 0.93 0.96 0.94
tRNA synthetases class I (C) catalytic domain 812 812 0.95 0.97 0.96
SecA Wing and Scaffold domain 805 805 0.95 0.97 0.96
Ribonuclease HII 795 795 0.93 0.94 0.93
Ribosomal protein L31 795 795 0.97 0.99 0.98
Ribosomal L27 protein 794 794 0.98 0.99 0.99
IPP transferase 794 794 0.93 0.95 0.94
GTP-binding protein LepA C-terminus 793 793 0.96 0.98 0.97
Ribosomal protein L17 791 791 0.92 0.96 0.94
Ribosomal protein L23 790 790 0.91 0.96 0.94
Ribosomal protein L10 781 781 0.90 0.92 0.91
Ribosomal protein L19 780 780 0.94 0.97 0.95
Ribosomal protein S20 774 774 0.95 0.97 0.96
Ribosomal protein L35 769 769 0.93 0.97 0.95
Phosphoglucomutase-phosphomannomutase, C-terminal domain 768 768 0.92 0.96 0.94
AMP-binding enzyme 767 767 0.87 0.89 0.88
Ribosomal prokaryotic L21 protein 766 766 0.93 0.96 0.95
tRNA methyl transferase 759 759 0.94 0.96 0.95
Ribosomal L29 protein 757 757 0.95 0.97 0.96
Glycosyl transferase family, a-b domain 754 754 0.90 0.91 0.91
Translation initiation factor IF-2, N-terminal region 750 750 0.96 0.98 0.97
Ribosomal L28 family 749 749 0.93 0.98 0.95
Glycosyl transferase family 4 739 739 0.96 0.98 0.97
tRNA synthetases class I (R) 736 736 0.93 0.96 0.95
Bacterial trigger factor protein (TF) C-terminus 733 733 0.95 0.96 0.95
For the first 1,000 families 261,149 261,149 0.92 0.95 0.94
For the first 2,000 families 293,957 293,957 0.90 0.96 0.93
For the first 3,000 families 308,292 308,292 0.89 0.96 0.92
For the first 4,000 families 316,135 316,135 0.87 0.96 0.91
Weighted average for all 7,027 families 324,018 324,018 0.91 0.95 0.93
Open in a new tab
Disordered Proteins Visualization and Classification
Due to the functional importance of disordered proteins, prediction of unstructured regions of disordered proteins and determining the sequence patterns featured in disordered regions is a critical problem in protein bioinformatics. We evaluated the ability of ProtVec to characterize and discern disordered protein sequences from structured sequences.
FG-Nups Characterization
concept_edgesIn this case study, we used the FG-Nups collection of 1,138 disordered proteins containing disorder regions with a fraction of at least one third of the sequence length. For comparison purposes, we also collected two sets of structured proteins from Protein Data Bank (PDB).
In order to visualize each dataset, we reduced the dimensionality of the protein-space using Stochastic Neighbor Embedding [ 36 , 39 ] and then generated the 2D histogram of all overlapping 3-grams occurring in each dataset. As shown in Fig 3 (see column (b)), the two random sets from structured proteins had nearly identical patterns. However, the FG-Nups dataset exhibits a substantially different pattern. To amplify the characteristic of disordered sequences we have also examined the histogram of disordered regions of FG-Nups (see Fig 3 , column (a)).
Fig 3. Visualization of protein sequences using ProtVec can characterize FGNUPs versus Disport disordered sequences and structured sequences.
Open in a new tab
Column (a) compares FG Nup sequences 2D histogram (at the bottom) with 2D histogram of FG Nup disordered regions (on top). Column (b) compares 2D histogram two random sets of structured sequences with the same average length as the FG-Nups. Column (c) compares between 2D histogram of DisProt sequences (at the bottom) and 2D histogram of DisProt disordered regions (on top).
In the next step, we quantitatively evaluated how ProtVec can be used to distinguish between FG-Nups versus typical PDB sequences using a support vector machine binary classification. The positive examples were the above mentioned 1,138 disordered FG-Nups proteins and negative examples (again 1,138 sequences) were selected randomly from PDB with the same average length of disordered sequences (≈ 900 residues). We represented each protein sequence as a summation of its ProtVecs of all 3-grams. Since on average the length of structured proteins were shorter than FG-Nups, controlsin order to avoid trivial cases, the PDB sequences were selected in such a way as to maintain the same average length. But still, an accuracy of 99.81% was obtained with high sensitivity and specificity ( Table 3 ). The distribution of the classified proteins in a 2D space is shown in Fig 4 .
Table 3. The performance of FG-Nups disordered protein classification in a 10xFold cross-validation using SVM.
Sensitivity Specificity Accuracy
0.9987 0.9974 0.9981
Open in a new tab
Fig 4. Classification of FG-Nups versus PDB structured sequences.
Open in a new tab
In this figure, each point presents a protein projected into a 2D space.
DisProt characterization
In this part, we used DisProt consisting of 694 proteins presenting 1539 disordered, and 95 ordered regions. We performed the same analysis as we did for FG-Nups with DisProt sequences (see Fig 3 column (c)). limitationsSince the size of DisProt was relatively small compared to that of the FG-Nups, the scales of columns (a),(b) were not comparable with column (c) (see Fig 3 ). The visualization of disordered regions of DisProt sequences ( Fig 3 column (c), on top) revealed a different characteristic than FG-Nups disordered regions ( Fig 3 column (a), on top). content_tags↗A visual comparison between Figs 3 and 2 suggest that the FG-Nups have a significantly higher amount of hydrophobic residues and less polar residues in their disordered regions than the experimentally identified disordered proteins in DisProt [ 27 , 29 ]. Additionally, the DisProt disordered regions versus DisProt ordered regions can be classified with 100% accuracy respectively using SVM and ProtVec.
Conclusions
concept_edgesAn unsupervised data-driven distributed representation, called ProtVec, was proposed for application of machine learning approaches in biological sequences. concepts↗By training this representation solely on protein sequences, our feature extraction approach was able to capture a diverse range of meaningful physical and chemical properties. concepts↗We demonstrated that ProtVec can be used as an informative and dense representation for biological sequences in protein family classification, and obtained an average family classification accuracy of 93%.
We further proposed ProtVec as a powerful approach for protein data visualization and showed the utility of this approach by providing an example in characterization of disordered protein sequences vs. structured protein sequences. Our results suggest that ProtVec can characterize protein sequences in terms of biochemical and biophysical interpretations of the underlying patterns. In addition, this dense representation of sequences can help to discriminate between various categories of sequences, e.g. disordered proteins. Furthermore, we demonstrated that ProtVec was able to identify disordered sequences with an accuracy of nearly 100%. The related data is available at: http://llp.berkeley.edu and Harvard Dataverse: http://dx.doi.org/10.7910/DVN/JMFHTN .
Another advantage of this method is that embeddings could be trained once and then used to encode biological sequences in any given problem. In general, machine learning approaches in bioinformatics can widely benefit from bio-vectors (ProtVec and GeneVec) representation. concept_edgesThis representation can be considered as pre-training for various applications of deep learning in bioinformatics. In particular, ProtVec can be used in protein interaction predictions, structure prediction, and protein data visualization.
Supporting Information
S1 File. The results of family classification task for all 7,027 families. (XLSX)
Click here for additional data file. (316.7KB, xlsx)
Acknowledgments
Fruitful discussions with Kiavash Garakani, Mohammad Soheilypour, Zeinab Jahed, Mohaddeseh Peyro, Hengameh Shams and other members of the Molecular Cell Biomechanics Lab at the University of California Berkeley are gratefully acknowledged.
Data Availability
All relevant data are within the paper and its Supporting Information files. Our web-based tools and trained data are available at Life Language Processing website: http://llp.berkeley.edu and through the Dataverse database: http://dx.doi.org/10.7910/DVN/JMFHTN , and will be regularly updated for calculation/classification of ProtVecs as well as visualization of biological sequences.
Funding Statement
Financial support from National Science Foundation through a CAREER Award (CBET-0955291) is gratefully acknowledged. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Relations on this page carry a quotation for the claim. Where a relation was separately evidenced — and so tested against the vocabulary for its type — the row says so; 2 of 11 relations here carry that mark. The rest are the tagger’s reading of the sentence quoted beside them.
| facet | tag | field | conf. | whose words |
|---|---|---|---|---|
| object_of_study | Protein and gene sequences treated as text, and whether a word-embedding trained on them captures biophysical and biochemical structure. | object_of_study | the paper's words | |
| claim_type | method | claim_type | the paper's words | |
| phenomenon | distributed vector representation of protein sequences | phenomenon | the paper's words | |
| method | system-or-tool-building also in: system-or-tool-building | method | the paper's words | |
| mechanisms | mechanisms | mechanisms[0] | the tagger's words | |
| mechanisms | A protein is split into three shifted non-overlapping 3-gram sequences rather than one overlapping stream, which triples the corpus and was chosen by cross-validation | mechanisms[1] | the tagger's words | |
| mechanisms | Because Skip-gram places sequences with similar contexts near each other, 3-grams sharing biophysical properties end up adjacent without ever being told those properties | mechanisms[2] | the tagger's words | |
| mechanisms | The embedding is trained once on the whole database and then reused as a feature extractor, so downstream tasks need no retraining | mechanisms[3] | the tagger's words | |
| mechanisms | Continuity of a property over the space is measured by the smallest Lipschitz constant, and read against the same measure on a space with the labels shuffled | mechanisms[4] | the tagger's words | |
| limitations | Families with fewer than ten instances have validation sets the authors describe as statistically insufficient, so a weighted rather than plain average accuracy had to be used. what this limits: the per-family accuracy figures for small families: FOR THE FAMILIES WITH NUMBER OF INSTANCES LESS THAN 10, THE VALIDATION [set is statistically insufficient], which is why a weighted average is reported | limitations[0] | the tagger's words | |
| limitations | The DisProt visualisations are not on a comparable scale to the FG-Nups ones because the DisProt set is much smaller. what this limits: any visual comparison between the DisProt and FG-Nups panels: THE SCALES OF COLUMNS (A),(B) WERE NOT COMPARABLE WITH COLUMN (C) | limitations[1] | the tagger's words | |
| limitations | Almost all FG-Nup sequences used are computational predictions rather than curated entries. what this limits: the disorder demonstration as an empirical result: ONLY 10 SEQUENCES OUT OF 1,138 DISORDERED SEQUENCES EXIST IN SWISS-PROT - the set is mostly computational prediction, not curated observation | limitations[2] | the tagger's words | |
| limitations | The headline accuracy is the mean over 7,027 SEPARATE binary classifications, each with equal numbers of positive and negative sequences, so the chance baseline for each is 50% - not an assignment among 7,027 families what this limits: the headline accuracy: it is the mean over 7,027 separate binary family classifications, each built with equal numbers of positives and negatives | limitations[3] | the tagger's words | |
| limitations | That figure is nonetheless compared against studies doing different and harder tasks - 54 families in one case, SCOP family detection in another what this limits: the comparison against prior work: THE REPORTED ACCURACIES OF A PREVIOUS STUDY ... HAVE BEEN IN THE RANGE OF 69.1-99.6% FOR 54 PROTEIN FAMILIES - a wide band, and a different family count | limitations[4] | the tagger's words | |
| limitations | Accuracy falls monotonically as rarer families are included, from 94% over the top thousand to 91% over the top four thousand, and the overall figure is weighted to reduce the influence of the rarest what this limits: the headline accuracy figure: it is a WEIGHTED AVERAGE ACCURACY, and accuracy falls as rarer families are included | limitations[5] | the tagger's words | |
| limitations | The disordered-protein results are at or near 100%, which for a sequence task on two databases invites the question of what distinguishes the two sets besides disorder what this limits: the disordered-protein result, and this is an IDENTIFIED concern rather than an author concession: near-total separation (99.8%) of two databases invites the question whether something other than the biology distinguishes them | limitations[6] | the tagger's words | |
| limitations | The FG-Nups set is largely computational rather than experimental - only ten of its 1,138 sequences are in Swiss-Prot duplicate of: limitations[2] · what this limits: the same scope as limitations[2], with the count: the FG-Nup set is largely computational, only ten of its 1,138 sequences being curated entries | limitations[7] | the tagger's words | |
| limitations | Window size and embedding size were chosen on the same cross-validation used to justify the representation what this limits: the choice of window and embedding size: they were selected on the SAME 2-FOLD CROSS-VALIDATION used to justify the representation, so the reported consistency is partly a property of the selection | limitations[8] | the tagger's words | |
| transfers | distributed word representations / the Skip-gram model, applied to protein sequences | transfers[0] | the tagger's words | |
| transfers | The Skip-gram model and the whole distributional idea that meaning is fixed by context, carried over by treating a 3-residue k-mer as a word and a protein as a sentence | transfers[1] | the tagger's words | |
| transfers | The word-analogy demonstration (king - man + woman = queen) imported as the evidence that embedding spaces carry structure worth looking for in sequences | transfers[2] | the tagger's words | |
| transfers | Content-addressable memory as the original motivation for distributed representation, credited explicitly rather than assumed | transfers[3] | the tagger's words | |
| controls | the PDB comparison set chosen so its AVERAGE LENGTH MATCHES the FG-Nup set, which is longer on average eliminates: sequence LENGTH as the feature separating the two classes - a classifier handed systematically shorter negatives learns to count residues, and would score well while knowing nothing about the biology · outcome (stated by the tagger, no quotation): held | controls[0] | 0.95 | the tagger's words |
| concepts | protein family also in: protein family | concepts[12] | 0.85 | the paper's words |
| concepts | DisProt also in: DisProt | concepts[17] | 0.85 | the paper's words |
| concepts | language of life also in: language of life | concepts[24] | 0.95 | the paper's words |
| concepts | conceptual analogy also in: conceptual analogy | concepts[25] | 0.95 | the paper's words |
| concepts | content-addressable also in: content-addressable | concepts[26] | 0.95 | the paper's words |
| concepts | partial descriptions also in: partial descriptions | concepts[27] | 0.95 | the paper's words |
| concepts | context also in: context | concepts[28] | 0.95 | the paper's words |
| concepts | positive training samples also in: positive training samples | concepts[29] | 0.95 | the paper's words |
| concepts | general-purpose representation also in: general-purpose representation | concepts[30] | 0.95 | the paper's words |
| concepts | sub sequences also in: sub sequences | concepts[31] | 0.95 | the paper's words |
| concepts | shifted non-overlapping words also in: shifted non-overlapping words | concepts[32] | 0.95 | the paper's words |
| concepts | three sequences also in: three sequences | concepts[33] | 0.95 | the paper's words |
| concepts | corpus consisting of 546,790 × 3 = 1,640,370 sequences of 3-grams | concepts[34] | 0.95 | the paper's words |
| concepts | negative sampling also in: negative sampling | concepts[35] | 0.95 | the paper's words |
| concepts | positive samples also in: positive samples | concepts[36] | 0.95 | the paper's words |
| concepts | negative samples also in: negative samples | concepts[37] | 0.95 | the paper's words |
| concepts | incorrect (w, c) pairs also in: incorrect (w, c) pairs | concepts[38] | 0.95 | the paper's words |
| concepts | vector of size 100 also in: vector of size 100 | concepts[39] | 0.95 | the paper's words |
| concepts | summation of the vector representation of overlapping 3-grams | concepts[40] | 0.95 | the paper's words |
| concepts | negative examples also in: negative examples | concepts[41] | 0.95 | the paper's words |
| concepts | 10 × fold cross-validations also in: 10 × fold cross-validations | concepts[42] | 0.95 | the paper's words |
| concepts | specificity also in: specificity | concepts[43] | 0.95 | the paper's words |
| concepts | sensitivity also in: sensitivity | concepts[44] | 0.95 | the paper's words |
| concepts | primary sequence information alone also in: primary sequence information alone | concepts[45] | 0.95 | the paper's words |
| concepts | extensive information for feature extraction | concepts[46] | 0.95 | the paper's words |
| concepts | continuity of these properties also in: continuity of these properties | concepts[47] | 0.95 | the paper's words |
| concepts | best Lipschitz constant also in: best Lipschitz constant | concepts[48] | 0.95 | the paper's words |
| concepts | protein-space also in: protein-space | concepts[49] | 0.95 | the paper's words |
| concepts | random space also in: random space | concepts[50] | 0.95 | the paper's words |
| concepts | randomly shuffling the labels of 3-grams | concepts[51] | 0.95 | the paper's words |
| concepts | short map also in: short map | concepts[52] | 0.95 | the paper's words |
| concepts | contraction also in: contraction | concepts[53] | 0.95 | the paper's words |
| concepts | physical and chemical properties also in: physical and chemical properties | concepts[54] | 0.95 | the paper's words |
| concepts | informative and dense representation also in: informative and dense representation | concepts[55] | 0.95 | the paper's words |
| unhoused_relations | continuity of these properties → best Lipschitz constant The best Lipschitz constant is calculated to measure property continuity; measured-by is absent from the edge registry. | unhoused_relations[0] | 0.95 | the tagger's words |
| unhoused_relations | summation of the vector representation of overlapping 3-grams → vector of size 100 Summing the overlapping 3-gram vectors produces a sequence vector of size 100; represented-as is absent from the edge registry. | unhoused_relations[1] | 0.95 | the tagger's words |
| unhoused_relations | language of life → human languages The paper draws an analogy between biological and human languages; analogous-to is absent from the edge registry. | unhoused_relations[2] | 0.95 | the tagger's words |
| unhoused_relations | context → meaning of a word Context characterizes word meaning; characterizes is absent from the edge registry. | unhoused_relations[3] | 0.95 | the tagger's words |
| unhoused_relations | negative sampling → positive samples Negative sampling uses observed positive pairs; composed-from is absent from the edge registry. | unhoused_relations[4] | 0.95 | the tagger's words |
| unhoused_relations | negative sampling → negative samples Negative sampling uses generated negative pairs; composed-from is absent from the edge registry. | unhoused_relations[5] | 0.95 | the tagger's words |
| unhoused_relations | embedding → encode biological sequences in a given problem Training once enables later reuse for sequence encoding; enables-reuse is absent from the edge registry. | unhoused_relations[6] | 0.95 | the tagger's words |
| unhoused_relations | k = 1 → short map A function with k equal to one is called a short map; threshold-name is absent from the edge registry. | unhoused_relations[7] | 0.95 | the tagger's words |
| unhoused_relations | 0 ≤ k < 1 → contraction A function with k below one is called a contraction; threshold-name is absent from the edge registry. | unhoused_relations[8] | 0.95 | the tagger's words |
| senses | word - in natural language: a semantic or syntactic unit whose meaning is fixed by the words around it | senses[0] | 1.0 | the tagger's words |
| senses | word - biological: three consecutive amino acids, defined outright as the unit that plays the role a word plays in text | senses[1] | 1.0 | the tagger's words |
| senses | family - the broad sense: a set of evolutionarily related proteins, which is what the classification task means by it | senses[2] | 1.0 | the tagger's words |
| senses | family - the Pfam sense: one of four sibling categories, so that a Pfam 'family' is a species of the very thing the word names | senses[3] | 1.0 | the tagger's words |
| senses | accuracy - per-family binary: one classifier per family with equal positives and negatives, which is what every number in Table 2 is | senses[4] | 1.0 | the tagger's words |
| senses | smoothness - measured against a shuffled control rather than in absolute terms, so the reported figure is a ratio and not a value | senses[5] | 1.0 | the tagger's words |
| senses | Language of life - conceptual analogy mapping biological sequences to human language | senses[6] | 0.95 | the tagger's words |
| senses | Content-addressable representation - nearby storage of items and properties enabling partial-cue recall | senses[7] | 0.95 | the tagger's words |
| senses | Embedding context - neighboring words that characterize a word’s meaning | senses[8] | 0.95 | the tagger's words |
| senses | Shifted 3-gram construction - three non-overlapping offset tokenizations per protein | senses[9] | 0.95 | the tagger's words |
| senses | ProtVec training corpus - 1,640,370 shifted 3-gram sequences derived from Swiss-Prot | senses[10] | 0.95 | the tagger's words |
| senses | Negative sampling - observed pairs pulled together and incorrect pairs used to prevent vector collapse | senses[11] | 0.95 | the tagger's words |
| senses | Protein sequence vector - sum of all overlapping 3-gram representations | senses[12] | 0.95 | the tagger's words |
| senses | Balanced one-family classifier - equal positive and randomly sampled negative instances | senses[13] | 0.95 | the tagger's words |
| senses | Specificity / sensitivity - true-negative and true-positive family-classification rates | senses[14] | 0.95 | the tagger's words |
| senses | Sequence-only feature learning - contrast with methods requiring extensive engineered properties | senses[15] | 0.95 | the tagger's words |
| senses | Best Lipschitz constant - smallest bound quantifying continuity of embedding-space properties | senses[16] | 0.95 | the tagger's words |
| senses | Scrambled space - control created by randomly shuffling 3-gram property labels | senses[17] | 0.95 | the tagger's words |
| alternatives | phenomenon: distributed vector representation of protein sequences a second reading reaching the same value. Recorded: distributed vector representation of protein sequences — note: a second reading reaching the same value from a different sentence. Not applied - a person decides. | alternatives.phenomenon | the tagger's words | |
| claim_relations | builds-on → NLP / word2vec methods | claim_relations[0] | the tagger's words | |
| claim_relations | builds-on → the Skip-gram / Word2Vec word embedding model | claim_relations[1] | the tagger's words | |
| claim_relations | replicates → the prior study of biophysical differences between FG-Nups and typical DisProt disordered proteins (ref. 29) | claim_relations[2] | the tagger's words | |
| claim_relations | extends → the Skip-gram word-embedding model of Mikolov et al. | claim_relations[3] | the tagger's words | |
| claim_relations | critiques → existing family classification methods that require hand-engineered biophysical features | claim_relations[4] | the tagger's words | |
| claim_relations | supports → the earlier finding of biophysical differences between FG-Nups and typical DisProt sequences | claim_relations[5] | the tagger's words | |
| concept_edges | ProtVec —is-a→ BioVec strength of the relation: 1 | concept_edges[0] | the tagger's words | |
| concept_edges | GeneVec —is-a→ BioVec strength of the relation: 1 | concept_edges[1] | the tagger's words | |
| concept_edges | EVIDENCE CORRECTED (RE-GROUNDED) · distributed representation —operationalised-as→ ProtVec EVIDENCE CORRECTED: RE-GROUNDED. The old span named only 'distributed representation'. This sentence names both and identifies the one as the other: the distributed representation 'called ProtVec'. — strength of the relation: 1
Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the representation is CALLED ProtVec; same-as fits, operationalised-as is the wrong type. also in: distributed representation · ProtVec | concept_edges[4] | the tagger's words | |
| concept_edges | EVIDENCE CORRECTED (RE-GROUNDED) · protein family —operationalised-as→ Pfam EVIDENCE CORRECTED: RE-GROUNDED. The old span named only ProtVec's use. This sentence names both and states what Pfam is FOR protein families - a source of them. — strength of the relation: 1
Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the study used Pfam information to classify protein families; operationalised-as is loose. also in: protein family · Pfam | concept_edges[6] | the tagger's words | |
| concept_edges | FG-Nups —is-a→ disordered proteins strength of the relation: 1 also in: FG-Nups · disordered proteins | concept_edges[7] | the tagger's words | |
| concept_edges | DisProt —operationalised-as→ disordered proteins strength of the relation: 1
Audit 2026-09-28: DOUBTFUL - read against its span and kept, but DisProt is a database OF disordered proteins; operationalised-as is the wrong type. also in: DisProt · disordered proteins | concept_edges[8] | the tagger's words | |
| concept_edges | FG-Nups —distinct-from→ DisProt strength of the relation: 1 | concept_edges[9] | the tagger's words | |
| concept_edges | ProtVec —operationalised-as→ pre-training strength of the relation: 0.9 · note: DOUBTFUL, found by an audit that read the edge against its span: type: 'can be considered as' is closer to is-a. The edge is kept live; a reader should weigh it with this in mind.
Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the span says ProtVec 'can be considered' pre-training; operationalised-as overstates a hedge. also in: ProtVec · pre-training | concept_edges[13] | the tagger's words | |
| concept_edges | EVIDENCE CORRECTED (RE-GROUNDED) · feature extraction —contrasted-with→ distributed representation EVIDENCE CORRECTED: RE-GROUNDED. The old span named neither endpoint. This sentence names both and states the contrast in the paper's own voice: n-grams used for TRAINING A REPRESENTATION rather than AS FEATURES. — strength of the relation: 0.9 also in: feature extraction · distributed representation | concept_edges[14] | the tagger's words | |
| concept_edges | primary sequence information alone —contrasted-with→ extensive information for feature extraction the relation itself is separately evidenced · strength of the relation: 1 · note: The paper explicitly contrasts its sequence-only method with feature-rich existing methods. | concept_edges[18] | 0.95 | the tagger's words |
| concept_edges | protein-space —distinct-from→ random space the relation itself is separately evidenced · strength of the relation: 1 · note: The protein space is reported as twice as smooth as the random control space. also in: protein-space · random space | concept_edges[19] | 0.95 | the tagger's words |
| content_tags | 93% ± 0.06% — average protein-family classification accuracy of ProtVec about this quote: QUOTE RECOVERED FROM THE PAPER'S OWN CHARACTERS. The recorded quote was LETTERS-ONLY IDENTICAL to this stretch of the staged text but not an exact substring of it - spaces, hyphens, ligature glyphs or mathematical operators differed, so it could never anchor. This is the file's own text between the same two letters: not produced by a rule about the damage, copied out of the file. Round-trip checked (the recovered span is letters-only equal to the original) and length-checked. The text as originally recorded is kept in evidence as extracted. | content_tags.quantities[0] | the tagger's words | |
| content_tags | 100-dimensional — dimensionality of the 3-gram embedding space | content_tags.quantities[1] | the tagger's words | |
| content_tags | 546,790 sequences — the Swiss-Prot training corpus | content_tags.quantities[2] | the tagger's words | |
| content_tags | 1,640,370 — training sequences after the three-way shifted split | content_tags.quantities[3] | the tagger's words | |
| content_tags | 324,018 protein sequences in 7,027 families — the classification benchmark | content_tags.quantities[4] | the tagger's words | |
| content_tags | 100-dimensional — the embedding size before projection | content_tags.quantities[5] | the tagger's words | |
| content_tags | 0.50 versus 1.01, ratio 0.51 — average Lipschitz constant in the protein space against the scrambled control | content_tags.quantities[6] | the tagger's words | |
| content_tags | 694 proteins, 1539 disordered and 95 ordered regions — the DisProt release used | content_tags.quantities[7] | the tagger's words | |
| content_tags | A protein sequence is represented as one dense n-dimensional vector learned with neural-network methods. | content_tags.findings[0] | the tagger's words | |
| content_tags | On protein family classification, ProtVec reaches 93% ± 0.06% average accuracy, outperforming existing methods. about this quote: QUOTE RECOVERED FROM THE PAPER'S OWN CHARACTERS. The recorded quote was LETTERS-ONLY IDENTICAL to this stretch of the staged text but not an exact substring of it - spaces, hyphens, ligature glyphs or mathematical operators differed, so it could never anchor. This is the file's own text between the same two letters: not produced by a rule about the damage, copied out of the file. Round-trip checked (the recovered span is letters-only equal to the original) and length-checked. The text as originally recorded is kept in evidence as extracted. | content_tags.findings[1] | the tagger's words | |
| content_tags | 3-grams sharing biophysical and biochemical properties cluster together in the embedding. | content_tags.findings[2] | the tagger's words | |
| content_tags | the protein space is on average about twice as smooth in physical and chemical properties as a random space | content_tags.findings[3] | the tagger's words | |
| content_tags | weighted average accuracy of 93% across 7,027 protein families, more reliable than existing methods | content_tags.findings[4] | the tagger's words | |
| content_tags | accuracy declines steadily as rarer families are added, 94% to 91% between the top 1,000 and top 4,000 | content_tags.findings[5] | the tagger's words | |
| content_tags | disordered sequences were separated from structured ones at 99.8% and 100.0% | content_tags.findings[6] | the tagger's words | |
| content_tags | the embedding needs training only once and can then serve any downstream sequence task | content_tags.findings[7] | the tagger's words | |
| content_tags | EVIDENCE CORRECTED (RE-GROUNDED) · 3-gram EVIDENCE CORRECTED: RE-GROUNDED. The published string was not verbatim in EITHER copy of the paper. A quote field promises the author's words; this held 'each protein sequence IS represented'. The paper says WILL BE - it is describing a procedure it is about to apply, not a property that holds. also in: 3-gram | content_tags.variables[0] | the tagger's words | |
| content_tags | family classification accuracy about this quote: QUOTE RECOVERED FROM THE PAPER'S OWN CHARACTERS. The recorded quote was LETTERS-ONLY IDENTICAL to this stretch of the staged text but not an exact substring of it - spaces, hyphens, ligature glyphs or mathematical operators differed, so it could never anchor. This is the file's own text between the same two letters: not produced by a rule about the damage, copied out of the file. Round-trip checked (the recovered span is letters-only equal to the original) and length-checked. The text as originally recorded is kept in evidence as extracted. also in: family classification accuracy | content_tags.variables[1] | the tagger's words | |
| content_tags | primary sequence also in: primary sequence | content_tags.variables[2] | the tagger's words | |
| content_tags | protein family also in: protein family | content_tags.variables[3] | the tagger's words | |
| content_tags | window size also in: window size | content_tags.variables[4] | the tagger's words | |
| content_tags | embedding vector size also in: embedding vector size | content_tags.variables[5] | the tagger's words | |
| content_tags | ProtVec also in: ProtVec | content_tags.definitions[0] | the tagger's words | |
| content_tags | EVIDENCE CORRECTED (RE-GROUNDED) · n-gram EVIDENCE CORRECTED: RE-GROUNDED. The published string was not verbatim in EITHER copy of the paper. A quote field promises the author's words; this held 'Every biological sequence involves fixed-length overlapping n-grams'. THE PAPER'S SENTENCE IS ABOUT A TECHNIQUE, NOT ABOUT SEQUENCES: 'the simplest and most common technique in bioinformatics to study sequences involves fixed-length overlapping n-grams'. The rewrite turned a statement about method into a statement about nature. also in: n-gram | content_tags.definitions[1] | the tagger's words | |
| content_tags | distributed representation also in: distributed representation | content_tags.definitions[2] | the tagger's words | |
| content_tags | protein family also in: protein family | content_tags.definitions[3] | the tagger's words | |
| content_tags | biological word also in: biological word | content_tags.definitions[4] | the tagger's words | |
| content_tags | disordered proteins also in: disordered proteins | content_tags.definitions[5] | the tagger's words | |
| content_tags | feature extraction also in: feature extraction | content_tags.definitions[6] | the tagger's words | |
| content_tags | scrambled space also in: scrambled space | content_tags.definitions[7] | the tagger's words | |
| content_tags | statistical-significance also in: statistical-significance | content_tags.statistics[1] | the tagger's words | |
| content_tags | specificity also in: specificity | content_tags.statistics[2] | the tagger's words | |
| content_tags | sensitivity also in: sensitivity | content_tags.statistics[3] | the tagger's words | |
| content_tags | cross-validation also in: cross-validation | content_tags.statistics[4] | the tagger's words | |
| content_tags | weighted average also in: weighted average | content_tags.statistics[5] | the tagger's words |
Tags that could not be placed in the text:
What was deliberately not tagged, and why, is recorded per paper by the tagger. Where a paper has no such record, this page says so rather than implying that nothing was declined.