Bioinformatics, 33(21), 2017, 3387–3395

doi: 10.1093/bioinformatics/btx431
Advance Access Publication Date: 7 July 2017
Original Paper

Sequence analysis

DeepLoc: prediction of protein subcellular

localization using deep learning

Downloaded from [Link] by guest on 24 September 2025

José Juan Almagro Armenteros1,2,*, Casper Kaae Sønderby2,
Søren Kaae Sønderby2, Henrik Nielsen1 and Ole Winther2,3
1
Department of Bio and Health Informatics, Technical University of Denmark, 2800 Kgs. Lyngby, Denmark, 2The
Bioinformatics Centre, Department of Biology, University of Copenhagen, 2200 Copenhagen N, Denmark and 3DTU
Compute, Technical University of Denmark, 2800 Kgs. Lyngby, Denmark
*To whom correspondence should be addressed.
Associate Editor: John Hancock
Received on March 16, 2017; revised on June 6, 2017; editorial decision on June 29, 2017; accepted on July 3, 2017

Abstract
Motivation: The prediction of eukaryotic protein subcellular localization is a well-studied topic in
bioinformatics due to its relevance in proteomics research. Many machine learning methods have
been successfully applied in this task, but in most of them, predictions rely on annotation of homo-
logues from knowledge databases. For novel proteins where no annotated homologues exist, and
for predicting the effects of sequence variants, it is desirable to have methods for predicting protein
properties from sequence information only.
Results: Here, we present a prediction algorithm using deep neural networks to predict protein sub-
cellular localization relying only on sequence information. At its core, the prediction model uses a
recurrent neural network that processes the entire protein sequence and an attention mechanism
identifying protein regions important for the subcellular localization. The model was trained and
tested on a protein dataset extracted from one of the latest UniProt releases, in which experimen-
tally annotated proteins follow more stringent criteria than previously. We demonstrate that our
model achieves a good accuracy (78% for 10 categories; 92% for membrane-bound or soluble), out-
performing current state-of-the-art algorithms, including those relying on homology information.
Availability and implementation: The method is available as a web server at [Link]
dk/services/DeepLoc. Example code is available at [Link]
ization. The dataset is available at [Link]
Contact: jjalma@[Link]

1 Introduction been extensively studied in bioinformatics (Emanuelsson et al.,

Proteins fulfil a wide diversity of functions inside the various com- 2007; Imai and Nakai, 2010; Wan and Mak, 2015).
partments of eukaryotic cells. The function of a protein depends on Most of the current machine learning methods for subcellular lo-
the compartment or organelle where it is located, as it provides a calization prediction extract a fixed number of features from the
physiological context for its function. However, aberrant protein protein sequences and use this fixed length representation as input
subcellular localization can affect the function that a protein ex- to a non-linear classifier such as a support vector machine (SVM).
hibits and contributes to the pathogenesis of many human diseases; However, sequence-based models, which process one position at a
such as metabolic, cardiovascular and neurodegenerative diseases, time, are more natural for this task as they can learn and make infer-
as well as cancer (Hung and Link, 2011). Therefore, predicting the ences from input of varying length. Unfortunately, these models
subcellular localization of the proteins is an essential task which has have not been competitive with non-linear classifiers up until

C The Author 2017. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: [Link]@[Link]
V 3387
3388 [Link] Armenteros et al.

recently. In this paper we take advantage of progress in deep learn- test set, which results in overly optimistic performances that do not
ing, specifically recurrent neural networks (RNNs) with long short- reflect the true generalization to new unseen proteins. An example
term memory (LSTM) cells, attention models and convolutional of a state-of-the-art method that uses this dataset set is Sherloc2,
neural networks (CNNs), to propose an end-to-end sequence-based which reports an accuracy of 93% for 11 locations.
model. LSTMs contain memory cells that can hold information This paper has four major contributions:
from past inputs to the network for in principle an arbitrary number
1. We construct a new dataset from a recent version of UniProt
of positions (Hochreiter and Schmidhuber, 1997). Attention
where proteins have experimental evidence for their subcellular
(Bahdanau et al., 2014) makes it possible to detect sorting signals in
locations according to the new stricter definition. We perform
proteins regardless of their position in the sequence. In addition,
stringent homology partitioning to avoid overfitting, providing
CNNs are able to train filters that detect short motifs in the input se-
realistic accuracy measures on new proteins.
quence irrespectively of where they occur, and have shown promis-
2. We show that models trained on the Höglund dataset have poor
ing performance for protein subcellular localization when combined

Downloaded from [Link] by guest on 24 September 2025

generalization performance on our new dataset. This reflects the
with LSTMs (Sønderby et al., 2015). We also propose a hierarchical
tree likelihood mimicking the biology of the sorting pathway and a high level of homology and possibly erroneous annotations in
transfer learning approach to jointly predict subcellular localization the old dataset.
and whether the protein is membrane-bound or soluble. 3. We develop deep recurrent neural networks for the protein sub-
In the following we discuss some of the caveats with the datasets cellular localization task with a number of novel state-of-the-art
used in previous subcellular localization tools. First, many methods model features. This includes convolutional motif detectors, se-
use homology information for prediction, either by directly using lective attention on sequence regions important for subcellular
annotated subcellular location annotations of retrieved hits in a localization prediction and a novel hierarchical sorting likeli-
database search, as in LocTree3 (with an accuracy of 80% for 18 lo- hood. These features are used for interpretation of the model
cations) (Goldberg et al., 2014), or by taking hints from other types and predictions. Our networks show improved prediction accur-
of annotation such as GO-terms, as in iLoc-Euk and YLoc acy without using homology information.
(Briesemeister et al., 2010; Chou et al., 2011), or PubMed abstracts 4. We implement the resulting model as a user-friendly web-server
linked to the protein’s Swiss-Prot entry, as in SherLoc (Briesemeister called DeepLoc (Concurrently with our work, Kraus et al.
et al., 2009). These methods are appropriate for annotated proteins (2017) has introduced a method for protein subcellular location
or proteins with annotated close homologues. Nonetheless, it should from cell image data also called DeepLoc).
be taken into account that the performance will be much lower for
sequences without well-annotated homologues—precisely the se-
quences for which it would be most relevant to have working predic- 2 Materials and methods
tion methods. In addition, any homology-based method will have
very limited chance of being able to predict the consequences of mu- 2.1 Neural network models
tations affecting sorting signals because the wild-type and the vari- The deep learning neural network model used is described in detail
ant probably would pick up the same homologues in a database below. Figure 1 and the following description gives a summary of
search. the architecture used: The input is sequence length (¼1000)  size of
Second, the performances of machine learning algorithms are amino acid vocabulary (¼20). The CNN extracts motif information
crucially dependent on the datasets used to train and test them. For using 120 filters of different sizes (20 for each of the sizes 1, 3, 5, 9,
protein subcellular localization a key aspect is that proteins should 15 and 21). This gives a 1000  120 feature map. Another convolu-
have experimental evidence for their subcellular location, so that tional layer of 128 filters of size 3  120 is applied to this feature
predictions are not based on predictions in a circular fashion. map. This gives a 1000  128 feature map which is used as input to
However, current methods use data from UniProt (The UniProt the recurrent layer. The recurrent neural network scans the sequence
Consortium, 2017) prior to release 2014_09, where a major change using 256 LSTM units in both directions giving in total a
in the annotation standards took place. Before the change, an anno- 1000  512 dimensional output. The attention decoding layer uses
tation was regarded as experimental if it lacked qualifiers such as an LSTM with 512 units through 10 decoding steps and the atten-
‘Potential’, ‘Probable’ or ‘By similarity’; after the change, only anno- tion mechanism feedforward neural network (FFN) has 256 units.
tations with a specific literature reference were annotated as being The final fully connected dense layer is composed by 512 and the
experimental (evidence code ECO:0000269). This resulted in a con- two output layers have one unit (membrane-bound) and 10 units
siderable decrease in the number of proteins with subcellular loca- (subcellular localization).
tion regarded as experimentally confirmed, thus raising the issue We learn a subcellular localization model which predicts the sub-
that current methods may in part be trained and tested on question- cellular localization using the amino acids sequence as input:
able examples. y ¼ fh ðX Þ ; (1)
Another aspect of the dataset issue is that the amount of hom-
ology between training data and test data should be kept at a min- where y is the predicted localization, f is the prediction model para-
imum (Hobohm et al., 1992). The measured test performance metrized by parameters h and X is the input data sequence of size
should be a true measure of the predictive performance on new pro- L  N where L is the protein length and N is the number of input
teins and not just a measure of how good the method is at finding features per sequence position. The parameters h are optimized
homologues with the same subcellular location. Unfortunately, the using stochastic gradient descent with cross entropy loss between
Höglund dataset (Höglund et al., 2006) which has been used in the the true and predicted localization distribution.
training and test of several methods (Blum et al., 2009; In practice, the length of protein sequences can vary from tens to
Briesemeister et al., 2009, 2010; Shatkay et al., 2007; Sønderby thousands of amino acids posing a challenge for many prediction al-
et al., 2015) is only homology reduced to 80% identity. This means gorithms requiring a fixed size input representation. Instead, recur-
that rather close homologues to the training data will occur in the rent neural networks (RNN) that naturally handle varying input
Prediction of protein subcellular localization using deep learning 3389

Downloaded from [Link] by guest on 24 September 2025

Fig. 1. (A) The convolutional neural network (CNN) extracts motif information using different motif sizes. (B) The recurrent neural network scans the sequence in
both directions, extracting the spatial dependencies between amino acids. (C) The attention mechanism assigns higher importance to amino acids that are rele-
vant for the prediction. At each decoding step, the attention weights a are generated based on the hidden states from the RNN and the hidden states from the pre-
vious decoding step. The weighted average of these weights at the last decoding step is used as input to a fully connected dense layer. (D) All the information
gathered from the protein sequence is passed to a softmax function and a hierarchical tree of sorting pathways to calculate the final prediction

sequence lengths were used. The networks applies a recurrent calcu- two separate RNNs and the input to the final classifier is then the
lation at each sequence position t concatenated outputs of the last hidden state of both RNNs. The
forwards and backwards RNNs will then be better at remembering
ht ¼ fE ðxt ; ht1 Þ; t ¼ 1 . . . L (2)
motifs in the C-terminus and N-terminus respectively. Nevertheless,
where fE is an RNN denoted the encoder, xt is the input features of for long sequences these algorithms still have to remember informa-
X at position t and h ¼ ½h1 ; . . . ; hL  is the hidden states of the RNN tion across many steps. To solve this problem, as well as identify
where ht is a vector of same length as the number of hidden units in protein regions important for classification, we augmented the bidir-
the RNN. The encoder can be viewed as a trainable feature extrac- ectional RNN encoder with an attentive decoder (Bahdanau et al.,
tor encoding the amino acid sequence into a feature space suitable 2014). Using the last hidden state of the encoder hL as input the at-
for subcellular localization prediction. Naively, the final subcellullar tentive decoder fD is run for D decoding steps. Note that D does not
location y could be predicted by applying a classifier fy to the final depend on the input sequence length L. At each step, the hidden
hidden state of the encoder hL state of the attentive decoder d r is used by an attention function fA
to assign a normalized importance weight to each sequence position
y ¼ fy ðhL Þ : (3) of the encoder hidden states h ¼ ½h1 ; . . . ; hL  as
However, this approach is not ideal for several reasons. Firstly the d r ¼ fD ðhL ; d r1 ; cr1 Þ; r ¼ 1 . . . D (4)
RNN has to remember all useful information across the entire, often
very long, input sequence. In subcellular localization this is espe- et;r ¼ fA ðht ; d r Þ ¼ tanhðht W e þ d r1 W d ÞvT (5)
cially problematic since most of the information is known to reside
 
in the beginning (N-terminus) and end (C-terminus) of the sequence. exp et;r
Secondly all information about the protein has to be compressed at;r ¼ ; (6)
P
L  
into the same size vector regardless of the length of the protein. exp et0 ;r
t0 ¼1
Two different solutions were used to alleviate these problems,
Bidirectional RNNs and Attention RNNs. In bidirectional RNNs, where d r is the hidden state of the decoder at step r, matrices W d
the protein sequence is processed both forwards and backwards by and W e and column vector v are the trainable parameters of the
3390 [Link] Armenteros et al.

attention function. d r is vector of same size as the number of hidden PðyjXÞ ¼ PðyjB; XÞPðBjA; XÞpðAjXÞ : (9)
units in the decoder LSTM, which can be different from the dimen-
An example path is A ¼ Non-Secretory Pathway, B ¼ N-terminal
sionality of the encoder ht . at;r is the normalized importance weights
Sequence and y ¼ Mitochondria. Each of the nine binary classifiers
and cr is a weighted average of the encoder RNN hidden states cal-
is implemented by a logistic output connected to the fully connected
culated as
dense layer. By construction, the tree probabilities are normalized
P
X y pðyjXÞ ¼ 1.
L
cr ¼ at;r ht : (7)
t¼1
2.3 Datasets
The initial value of cr , i.e. c0 , is a learned parameter vector that is 2.3.1 DeepLoc dataset
trained as part of the neural network model. The subcellular local- The protein data used to train DeepLoc were extracted from the
ization is then predicted using the weighted average of the encoder UniProt database, release 2016_04 (The UniProt Consortium,

Downloaded from [Link] by guest on 24 September 2025

RNN hidden states at the last step of the decoder 2017). The protein dataset was filtered using the following criteria:
eukaryotic, not fragments (they could have the N-terminal or C-ter-
y ¼ fy ðcD Þ : (8)
minal missing), encoded in the nucleus, longer than 40 amino acids
This allows the model to selectively assigns weight to sequence pos- and experimentally annotated (ECO:0000269). Similar locations or
itions important for classification, which reduces the need for re- subclasses of the same location were mapped to 10 main locations
membering all information across the entire length of the sequence. in order to increase the number of proteins per compartment.
Both fE and fD are implemented as a special type of RNN units Furthermore, proteins were classified as membrane or soluble if they
called Long-Short Term Memory (LSTM) cells (Hochreiter and were found on either the membrane or the lumen of the organelle; if
Schmidhuber, 1997). LSTMs share the same chain structure as no information was provided, they were tagged as unknown.
RNNs, but the recurrent calculation is augmented with an internal Finally, proteins with more than one subcellular localization were
memory cell capturing long range dependencies. filtered out. A total of 13 858 proteins were obtained after the filter-
Furthermore, convolutional filters were used to detect protein ing process. The mapped sublocations and the number of proteins in
motifs. Here a filter, akin to position specific scoring matrices, is each main localization are summarized in Table 1.
slid across the sequence. It will then detect a motif regardless of To ensure that the model generalizes to new data a stringent
its position in the sequence. The weights of each filter can be ad- homology partitioning was performed. Homologous proteins that
justed to find the motifs that help to better predict each class. fulfil a certain threshold of similarity were clustered as detailed
These new features created with a CNN can represent the inputs in below. Then, each cluster of homologous proteins was assigned to
a more abstract way, which, in combination with LSTMs, has been one of the five folds, ensuring that similar proteins were not mixed
shown to be beneficial for protein classification (Sønderby et al., between the different folds. PSI-CD-HIT (Li and Godzik, 2006) was
2015). used to cluster proteins with 30% of identity or 106 E-value cutoff
and the alignment must cover 80% of shorter (redundant) se-
quences, which produced 8410 clusters for the whole dataset. The
2.2 Hierarchical tree likelihood five folds generated had approximately the same number of proteins
To include information from protein sorting pathways into our in each location. Four were used for the training and validation and
model, a hierarchical tree with multiple nodes was developed. Each one heldout set for testing.
node represents a binary decision attempting to assign the protein to
the right pathway from high-level to detailed classification. As an 2.3.2 Höglund dataset
example, the first binary decision in the tree classifies proteins in the The Höglund dataset (Höglund et al., 2006) have been used to train
secretory or non-secretory pathway, whereas the last nodes separate both the MultiLoc and RNN prediction methods in Höglund et al.
related compartments such as mitochondria and chloroplasts, see (2006) and Sønderby et al. (2015). This dataset consist of 5959 pro-
Figure 1 panel D. The leaf nodes correspond to the final subcellular teins with 11 possible locations (cytoplasm, nucleus, extracellular,
localizations, and the likelihood is calculated as the joint probability mitochondria, plasma membrane, ER, chloroplast, Golgi apparatus,
of decisions in the tree. So for example, if we have decisions A, B, y lysosome, vacuole and peroxisome) and is homology reduced to
then according to the tree decomposition the probability of y given 80% identity. Apart from grouping together lysosomal and vacuolar
input sequence X is given by proteins no modifications were made to the dataset.

Table 1. Number of proteins in each location and sublocations that were grouped together under the same main location

Location No. of proteins Sublocations

Nucleus 4043 Envelope, inner and outer membrane, matrix, lamina, chromosome, nucleus speckle
Cytoplasm 2542 Cytoplasm (cytosol and cytoskeleton)
Extracellular 1973 Extracellular
Mitochondrion 1510 Envelope, inner and outer membrane, matrix, intermembrane space
Cell membrane 1340 Apical, apicolateral, basal, basolateral, lateral, cell membrane, cell projection
Endoplasmic reticulum (ER) 862 ER membrane and lumen, microsome, rough ER, smooth ER, Sarcoplasmic reticulum
Plastid 757 Plastid membrane, stroma and thylakoid
Golgi apparatus 356 Golgi apparatus membrane and lumen
Lysosome/Vacuole 321 Contractile, lytic and protein storage vacuole, vacuole lumen and membrane, lysosome lumen and
membrane
Peroxisome 154 Peroxisome matrix and membrane
Prediction of protein subcellular localization using deep learning 3391

2.4 Comparison to current prediction algorithms never optimized any parameters on the test set leaving the reported
The performance of our models were compared with a number of performances unbiased.
current prediction algorithms using the following approaches: To decrease the training time, the maximum protein length was
LocTree2 (Goldberg et al., 2012), MultiLoc2 (Blum et al., 2009) 1000. If a protein exceeded this length, amino acids from the middle
and SherLoc2 (Briesemeister et al., 2009) were run with local of the sequence were removed in order to not to lose information
command-line versions installed on our own server, while CELLO about the N-terminal and C-terminal sorting signals. 9.98% of the
(Yu et al., 2006), iLoc-Euk (Chou et al., 2011) and WoLF PSORT proteins were truncated using this rule.
(Horton et al., 2007) were run on their web servers. YLoc The performance measurements used to assess the performance
(Briesemeister et al., 2010) was run offline by the maintainer of the of our models were accuracy and the Gorodkin measure (Gorodkin,
web service. Results for YLoc are given with the option to include 2004). For the binary prediction, the accuracy and the Matthew’s
GO terms turned on. For MultiLoc2 and SherLoc2, a newer version Correlation Coefficient (Matthews, 1975) (MCC) were used. The
of InterProScan (5.21-60) was used instead of the recommended one Gorodkin measure can be seen as a generalization of MCC that

Downloaded from [Link] by guest on 24 September 2025

(4.4) due to compatibility problems with the older version. As a ref- applies to K-categories, which is more informative than the accuracy
erence the performance of Höglund test set was measured on our when there is an imbalance of classes. For K¼2, the Gorodkin meas-
local installation obtaining an accuracy of 0.8300 for Multiloc2 and ure squared is the ‘generalized squared correlation’ (GC2) of Baldi
0.9179 for SherLoc2. et al. (2000).
In the cases where current methods predict more than ten loca- All models were implemented in Python 2.7.11 using the neural
tions, the predicted locations were mapped onto our ten locations. network library Lasagne 0.2 (Dieleman et al., 2015) and Theano 0.9.0
Two of the methods, iLoc-Euk and WoLF PSORT, in some cases (Theano Development Team, 2016) for efficient GPU implementation.
predict dual locations (such as cytoplasm/nucleus). Since pro-
teins with dual locations were filtered out in the construction of
the dataset, those predictions were counted as erroneous, unless 3 Results
both the predicted locations mapped to the same location in our We designed experiments to address the following questions:
classification.
• What are the relative performances of the proposed neural net-
work model architectures? ! Section 3.1
2.5 Experiments • How does the generalization performances of models trained on
Two different set of experiments were carried out. The first experi- either the DeepLoc or Höglund datasets compare? ! Section 3.2
ments were used for model selection comparing the relative perform- • How does the final DeepLoc model compare to current state-of-
ances of the following model architectures: the-art protein subcellular prediction models? ! Section 3.3
• Feedforward neural network (FFN)
• Bidirectional LSTM neural network (BLSTM) 3.1 Model selection
• BLSTM neural network with attention mechanism (A-BLSTM) In Table 2 we compare the performances of different model archi-
• Convolutional BLSTM neural network with attention mechan- tectures trained on the DeepLoc dataset. Note that we are interested
ism (Conv A-BLSTM) in the relative performance of the models. Due to this, we only used
BLOSUM62 encodings as input features, which resulted in a slightly
Using the best model architectures the second set of experiments is
degraded performance compared to the final performances
designed to test the generalization performance of models trained on
described in the following sections.
either our new DeepLoc dataset or the Höglund dataset.
The A-BLSTM and the CONV A-BLSTM models achieved the
Hyperparameters were optimized on three of four splits of the
highest performance predicting the subcellular localization with
training data and the performance was evaluated on the last valid-
accuracies of 0.7290 and 0.7289, respectively. Comparing these re-
ation split. The hyperparameter selection was done using uni-
sults with the performance of the BLSTM without attention (accur-
dimensional search where one hyperparameter was changed and the
acy 0.6925), we see that attention improves performance. These
rest were kept fixed. If a hyperparameter had not yet been tested,
results confirm the benefit of selective, context dependent, attention
the median value in the range of that hyperparameter was chosen.
for protein classification. All of the A-BLSTM models performed
Each hyperparameter setting was run for 150 epochs (epoch ¼ full
significantly better than the baseline FFN model which achieved an
pass over the training set) and the performance was measured as the
accuracy of 0.5234. This is expected since FFN models do not take
highest seen performance on the validation set. This strategy was
into account the order of the amino acids, whereas the LSTM mod-
used for computational reasons since a full grid search over all par-
els naturally consider the relationships between amino acids.
ameters was not computationally feasible. After the best hyperpara-
Furthermore, we observed that including 10 decoding steps in the at-
meters were identified, a final run of experiments were used to
tention mechanism increased the accuracy (a difference of 1%) in
identify the best combination of amino acid encodings among
BLOSUM62 (Henikoff and Henikoff, 1992), sparse, protein profiles
Table 2. Comparison of performances for different model architec-
or HSDM encoding (Prlic et al., 2000). We further found that pro-
tures using BLOSUM62 input features
tein profiles gave the highest performance and included these as in-
put features for the final models. The profiles were generated using Model Subcellular location Membrane
the same method as the TOPCONS web server (Tsirigos et al.,
Accuracy Gorodkin Accuracy MCC
2015).
The test performance was measured by training four models on FFN 0.5234 0.4229 0.7301 0.4509
the training set using the four different combinations of training and BLSTM 0.6925 0.6278 0.9004 0.8023
validation set. The reported test performance is the average of the A-BLSTM 0.7290 0.6729 0.9163 0.8345
four models evaluated on the held-out test set. We stress that we CONV A-BLSTM 0.7289 0.6780 0.9111 0.8218
3392 [Link] Armenteros et al.

comparison with a single decoding step. Increasing the decoding prediction accuracy we trained an ensemble of 16 models using
steps beyond 10 resulted in a reduction in the accuracy. Lastly, the nested cross validation. Eight of the models were trained using a
A-BLSTM models predicted whether the proteins were membrane- softmax output distribution (class probability from softmax func-
bound or soluble with accuracies of 0.9163 and 0.9111 respectively. tion) and eight of the models using the hierarchical tree distribution
From the amino acid encoding comparison, we found that the (joint probability of multiple logistic functions). Further we mitigate
CONV A-BLSTM model using protein profiles encoding had the the effect of the class imbalances by using a cost matrix (Zhou and
highest accuracy, with a difference of 2% compared to the A- Liu, 2006) to recalculate the class probabilities based on the number
BLSTM model. Therefore, we decided to use this encoding and this of samples in the training set. The full ensemble achieved an accur-
model for the rest of the experiments. acy of 0.7797 and Gorodkin of 0.7347 on the subcellular localiza-
tion and an accuracy of 0.9234 and a MCC of 0.8435 on the
3.2 Dataset comparison membrane-bound or soluble prediction. We found that the softmax
To compare the generalization performance of models trained on either models had a slightly higher accuracy than the hierarchical tree

Downloaded from [Link] by guest on 24 September 2025

the DeepLoc or the Höglund datasets, we trained a CONV A-BLSTM model with the 8-ensembles achieving an accuracy of 0.7717 and
model on each dataset and evaluated the performances on the test sets 0.7695, respectively. We show in Table 4 the accuracy and the
from both datasets. Table 3 shows that (i) the Höglund training set MCC for each binary decision in the hierarchical tree model. We ex-
achieves a good test performance only on the Höglund test set and perimented with increasing the ensemble size but found no improve-
(ii) the DeepLoc training set achieves a good test performance on test ment in performance.
sets with stringent independence between training and test sets. The training time for the full ensemble was 80 hours, approxi-
These results show that models trained on the Höglund dataset mately five hours per model. When testing, the ensemble takes
generalize poorly compared to models trained on the DeepLoc data- three seconds per protein on average to perform a prediction.
set. As a qualitative comparison of the two datasets we visualized Nonetheless, this ensemble used protein profiles, which were already
the context vectors cr for CONV A-BLSTM models trained on both generated for this dataset. This profile generation is the most time-
datasets as seen in Figure 2. The compartments are notably more consuming step usually taking approximately 30 seconds per
separated for the model trained on the Höglund dataset compared protein. If a hit with the PFAM database is not found the profile
to the model trained on the DeepLoc dataset generation uses Uniref90 instead. This can take even longer and
therefore can be problematic for large protein datasets. To solve this
3.3 DeepLoc model we trained the same ensemble using BLOSUM62 encoding. This
From the model comparisons we identified the CONV A-BLSTM as model has an accuracy of 0.7360 and Gorodkin of 0.6832 on the
the best performing model architecture. To further improve
Table 4. Accuracy and MCC of each node in the hierarchical tree
Table 3. Comparison of generalization performances using the
Node Accuracy MCC
CONV A-BLSTM model between the DeepLoc dataset and the
Höglund dataset Secretory/Non-secretory pathway 0.9502 0.8902
Intracellular/Extracellular 0.9507 0.8979
Training set Test set Accuracy Gorodkin
N-terminal sequences 0.9544 0.8784
DeepLoc DeepLoc 0.7511 0.6988 Intermediate compartment 0.7982 0.5824
Höglund DeepLoc 0.6426 0.5756 PTS 0.9784 0.4085
Mitochondrion/Chloroplast signals 0.9537 0.8955
DeepLoc Höglund 0.8301 0.8010 Cell membrane/Lysosome 0.8575 0.5002
Höglund Höglund 0.9138 0.8979 ER/Golgi 0.8559 0.6376
NLS 0.8138 0.6031
Note: Sequence profiles were used as input features.

Fig. 2. t-SNE representation of the context vector cr for a Conv A-BLSTM trained on the DeepLoc and Höglund dataset and visualized for the respective test sets
Prediction of protein subcellular localization using deep learning 3393

Table 5. Confusion matrix of the test set on the final DeepLoc model using profiles encoding

Location Number of predicted proteins Sens. MCC

Nucleus 680 103 4 5 2 8 1 2 2 1 0.842 0.784

Cytoplasm 94 361 7 18 5 4 3 8 1 7 0.711 0.608
Extracellular 3 5 365 5 5 4 2 0 4 0 0.929 0.907
Mitochondrion 9 21 0 247 0 5 14 2 1 3 0.818 0.812
Cell membrane 5 15 6 1 203 20 1 4 18 0 0.744 0.732
Endoplasmic reticulum 3 6 6 3 18 120 1 7 8 1 0.694 0.654
Plastid 1 2 0 8 0 0 140 0 1 0 0.921 0.883
Golgi apparatus 4 17 1 0 9 8 1 26 4 0 0.371 0.414
Lysosome/Vacuole 0 7 11 1 20 9 0 4 12 0 0.188 0.194
Peroxisome 0 13 0 4 1 4 0 0 0 8 0.267 0.321

Downloaded from [Link] by guest on 24 September 2025

Note: Sens., sensitivity.

Table 6. Confusion matrix for the membrane-bound predictor proteins, which generates a model that learns to predict the wrong
labels. Secondly, the homology reduction threshold 80% used for
Type Number of predicted proteins constructing the Höglund dataset might not be stringent enough,
since it produces similar training and test examples.
Soluble 968 38
Membrane-bound 96 647 In Figure 2, we compared the attention context vector for models
trained on either the DeepLoc or Höglund datasets. For the
Höglund trained model, all locations are almost perfectly separated
subcellular localization and an accuracy of 0.9130 and a MCC of implying that there is little variation within the Höglund dataset
0.8237 on the membrane-bound or soluble prediction. By omitting classes and that the training and test sets are relatively similar. This
the profile generation, we achieved a faster prediction at the cost of supports the finding of poor generalization performance for models
decrease in accuracy. trained on the Höglund dataset. Hence, we believe that the high per-
Tables 5 and 6 show the confusion matrices of the full ensemble formance reported for algorithms trained on this dataset is actually
described above for subcellular localization and membrane-bound results from overfitting. The true variation within each protein class
prediction respectively. The primary sources of error are confusion is larger as indicated by the better generalization performance for
of the nucleus and cytoplasm, lysosome/vacuole misclassified as cell- models trained on the DeepLoc dataset. This is further corroborated
membrane and Golgi misclassified as cytoplasm. In Figure 3, we by the poorer separation of classes for the DeepLoc trained models
show the attention vector a, i.e. how important different regions of in the same figure.
the sequence are for the classification. In general, the DeepLoc We compared the performance of the final DeepLoc model with
model assigns large importance to the N-terminal for secreted pro- other current prediction algorithms in Table 7. We found that the
teins whereas e.g. membrane proteins have regions of importance DeepLoc model performs significantly better than the other
interspersed across the protein length. approaches. Here we note that the DeepLoc performance is a true
To compare the performance of the final DeepLoc model to test set performance, whereas the performances of the other meth-
other approaches we benchmarked a number of current prediction ods may be overestimated since some sequences in our test set may
algorithms on the DeepLoc test set as seen in Table 7. The accuracy have been included in their training sets. Further we emphasize that
of the final DeepLoc model (0.7797) is significantly better than all the DeepLoc method is a purely sequence-based method and does
other methods with iLoc-Euk achieving the second best accuracy of not rely on annotation information from homologous proteins. Due
0.6820. to the stringent homology partitioning applied in the dataset con-
struction, the model should generalize to new proteins without
known close homologues.
4 Discussion We note that we also compared the performance against the
In this paper we have introduced the DeepLoc dataset: a well LocTree3 prediction method (Goldberg et al., 2014), which is a
assembled protein collection with reliable subcellular localization in- combination of LocTree2 and a BLAST search of a database of pro-
formation. Secondly we have provided a deep neural network based teins with known subcellular location. However, as 75% of the pro-
prediction algorithm achieving state-of-the-art performance on this teins in the DeepLoc test set are also in the LocTree3 BLAST
new dataset. The context-dependent annotation vector generated by database, the measured accuracy was artificially high at 91%, since
the attention mechanism is able to represent a protein based on its LocTree3 simply retrieves the same subcellular location used for
subcellular localization. In addition, the attention based prediction labelling our test set.
method allows visualization of the biologically plausible regions The compartment specific prediction performance of the final
used to predict the subcellular localization of the proteins which we DeepLoc model is shown in Table 5. The main source of error is the
believe will provide relevant information. low performance on the Golgi apparatus, lysosome/vacuole and per-
The comparison of the generalization performances for models oxisome. One possible cause is the low number of samples used to
trained on our new DeepLoc dataset and the Höglund dataset train these classes. However, this finding could also be associated
showed that DeepLoc trained models generalized much better than with the similarity between the proteins from these locations and
the Höglund trained model. Here we discuss a number of explan- other compartments. For example, Table 5 shows that the lysosome/
ations for these findings. Firstly, with the Uniprot database change vacuole is usually misclassified as cell membrane and the peroxi-
the Höglund dataset could contain many wrongly annotated some as cytoplasm.
3394 [Link] Armenteros et al.

Downloaded from [Link] by guest on 24 September 2025

Fig. 3. Sequence importance across the protein sequence of DeepLoc test set when making the prediction. The x-axis is the sequence position and alomg the
y-axis we have the proteins in the test set sorted according to protein localization. For visualization proteins shorter than 1000 amino acids are padded from the
middle, so the N-terminus and C-terminus align. Proteins longer than 1000 amino acids have the middle part removed

Table 7. Accuracy and Gorodkin measure achieved by current Figure 3 allows us to visualize what regions in the sequence are
predictors and the final DeepLoc model on the DeepLoc test set relevant for each subcellular localization to perform the prediction.
For the extracellular proteins, the model focuses mainly on the signal
Method Accuracy Gorodkin
peptide, which can be seen as a small region at the N-terminus of the
LocTree2 0.6120 0.5250 sequence. In contrast, the attention is scattered across the sequence for
MultiLoc2 0.5592 0.4869 plasma membrane proteins, which could indicate that the algorithm is
SherLoc2 0.5815 0.5112 detecting the transmembrane helices. For the ER proteins we can see
YLoc 0.6122 0.5330 attention at the N-terminus, where the signal peptide is located, and
CELLO 0.5521 0.4543 also some attention at the C-terminus, which could mean the presence
iLoc-Euk 0.6820 0.6412
of KDEL or KKXX signals. Golgi proteins have the importance on the
WoLF PSORT 0.5671 0.4785
N-terminus slightly shifted to the right, in comparison with other pro-
DeepLoc 0.7797 0.7347
teins from the secretory pathway, as they are mostly type II transmem-
The highest scores are shown in bold. brane proteins with signal anchors. Mitochondrial and chloroplastic
proteins have large regions at the N-terminus, which clearly correlates
to the mitochondrial and chloroplastic transit peptides. The lyso-
In addition to the mentioned under-represented classes, proteins
somal/vacuolar proteins do not seem to have a clear important region
from the cytoplasm and nucleus are also difficult to differentiate
across their sequences. Finally, for peroxisomal proteins, some regions
(Fig. 2, Table 4) because they both lack N-terminal sorting signals.
at the N-terminus and at the C-terminus are observed, which could
The only difference between them is the nuclear localization signal
mean that the model is detecting PTS2 and PTS1 signals.
(NLS), which is a highly variant short sequence that can be located
in multiple regions of the protein sequence, making it hard to recog-
nize. Figure 3 shows the positions in the sequence that the attention
mechanism focuses on to generate the attention context vector cr.
For the nucleic and cytoplasmic proteins, the model focuses on the
5 Conclusion
beginning of the sequence (checking for the absence of an N-ter- We have shown that convolutional BLSTM neural networks with at-
minal sorting signal). Moreover, the model also gives importance to tention mechanism are able to accurately predict the protein subcellular
small regions across the sequence. The main difference is that there localization and if a protein is membrane-bound or soluble just using
is a higher density of these regions in nucleus examples than in cyto- the sequence information. Further we have introduced the DeepLoc
plasm, which could indicate that the model is able to identify some dataset. The DeepLoc model trained on this dataset is able to generalize
of the most represented NLS. better than using previous datasets for subcellular localization. In
Prediction of protein subcellular localization using deep learning 3395

addition, DeepLoc obtained the highest accuracy using the independent Dieleman,S. et al. (2015) Lasagne: First Release. Geneva, Switzerland,
test set, when compared with the current methods. Zenodo.
There are several perspectives of this project that we would like Emanuelsson,O. et al. (2007) Locating proteins in the cell using TargetP,
SignalP and related tools. Nature Protoc., 2, 953–971.
to pursue in the future. One of those is to make better use of existing
Goldberg,T. et al. (2012) LocTree2 predicts localization for all domains of
knowledge about sorting signals. DeepLoc 1.0 is trained in a rela-
life. Bioinformatics, 28, i458–i465.
tively ‘naive’ way, where the networks have been provided only
Goldberg,T. et al. (2014) Loctree3 prediction of localization. Nucleic Acids
with protein profiles and their location labels. It would be beneficial Res., 42, W350–W355.
to explicitly model known sorting signals such as N-terminal signal Gorodkin,J. (2004) Comparing two k-category assignments by a k-category
peptides and transit peptides. correlation coefficient. Comput. Biol. Chem., 28, 367–374.
In addition, it should be investigated whether performance can Henikoff,S. and Henikoff,J.G. (1992) Amino acid substitution matrices from
be enhanced by training several models with a narrower taxonom- protein blocks. Proc. Natl. Acad. Sci. USA, 89, 10915–10919.
ical scope instead of treating all eukaryotes by one model. Hobohm,U. et al. (1992) Selection of representative protein data sets. Protein

Downloaded from [Link] by guest on 24 September 2025

Obviously, animals and fungi do not have plastids, and some false Sci., 1, 409–417.
Hochreiter,S. and Schmidhuber,J. (1997) Long short-term memory. Neural
predictions could be avoided by disallowing plastid predictions for
Comput., 9, 1735–1780.
these groups, but more subtle differences between sorting signals are
Höglund,A. et al. (2006) Multiloc: prediction of protein subcellular localiza-
also known to exist. However, there is a trade-off between the preci-
tion using N-terminal targeting sequences, sequence motifs and amino acid
sion of the taxonomical scope and the sizes of the training datasets. composition. Bioinformatics, 22, 1158–1165.
For taxonomic groups with limited numbers of data with experi- Horton,P. et al. (2007) WoLF PSORT: protein localization predictor. Nucleic
mentally known subcellular location, it may be necessary to employ Acids Res., 35, W585–W587.
semisupervised learning, where unlabelled data from genome se- Hung,M.-C., and Link,W. (2011) Protein localization in disease and therapy.
quences are used along with labelled data. J. Cell Sci., 124, 3381–3392.
Imai,K. and Nakai,K. (2010) Prediction of subcellular locations of proteins:
where to proceed? Proteomics, 10, 3970–3983.
Acknowledgements Kraus,O.Z. et al. (2017) Automated analysis of high-content microscopy data
with deep learning. Mol. Syst. Biol., 13, 924.
The authors wish to thank Konstantinos Tsirigos and Arne Elofsson of
Li,W. and Godzik,A. (2006) Cd-hit: a fast program for clustering and compar-
Stockholm University for permission to use their fast profile construction
ing large sets of protein or nucleotide sequences. Bioinformatics, 22,
method in DeepLoc, even though it has not been published yet. In addition,
1658–1659.
they want to thank Fabian Aicheler of University of Tübingen for kindly run-
Matthews,B.W. (1975) Comparison of the predicted and observed secondary
ning the DeepLoc test set on YLoc.
structure of T4 phage lysozyme. Biochim. Biophys. Acta (BBA)-Protein
Struct., 405, 442–451.
Prlic,A. et al. (2000) Structure-derived substitution matrices for alignment of
Funding distantly related sequences. Protein Eng., 13, 545–550.
S.K.S. and O.W. were supported by a grant from the Novo Nordisk Foundation Shatkay,H. et al. (2007) Sherloc: high-accuracy prediction of protein subcellu-
and the NVIDIA Corporation with the donation of TITAN X GPUs. lar localization by integrating text and protein sequence data.
Bioinformatics, 23, 1410–1417.
Conflict of Interest: none declared. Sønderby,S.K. et al. (2015) Convolutional LSTM networks for subcellular lo-
calization of proteins. In: International Conference on Algorithms for
Computational Biology, volume 9199 of Lecture Notes in Computer
References Science, pp. 68–80. Springer.
Bahdanau,D. et al. (2014) Neural machine translation by jointly learning to The UniProt Consortium (2017) UniProt: the universal protein knowledge-
align and translate. arXiv preprint arXiv:1409.0473. base. Nucleic Acids Res., 45, D158–D169.
Baldi,P. et al. (2000) Assessing the accuracy of prediction algorithms for classi- Theano Development Team (2016) Theano: A Python framework for fast
fication: an overview. Bioinformatics, 16, 412–424. computation of mathematical expressions. arXiv e-prints, abs/1605.02688.
Blum,T. et al. (2009) Multiloc2: integrating phylogeny and gene ontology terms Tsirigos,K.D. et al. (2015) The TOPCONS web server for consensus predic-
improves subcellular protein localization prediction. BMC Bioinformatics, tion of membrane protein topology and signal peptides. Nucleic Acids Res.,
10, 1. 43, W401–W407.
Briesemeister,S. et al. (2009) Sherloc2: a high-accuracy hybrid method for pre- Wan,S. and Mak,M.-W. (2015) Machine Learning for Protein Subcellular
dicting subcellular localization of proteins. J. Proteome Res., 8, 5363–5366. Localization Prediction. De Gruyter, Berlin, Germany.
Briesemeister,S. et al. (2010) YLoc–an interpretable web server for predicting Yu,C.-S. et al. (2006) Prediction of protein subcellular localization. Proteins,
subcellular localization. Nucleic Acids Res., 38, W497–W502. 64, 643–651.
Chou,K.-C. et al. (2011) iLoc-Euk: a multi-label classifier for predicting the Zhou,Z.-H. and Liu,X.-Y. (2006) Training cost-sensitive neural networks
subcellular localization of singleplex and multiplex eukaryotic proteins. with methods addressing the class imbalance problem. IEEE Trans.
PLoS ONE, 6, e18258. Knowledge Data Eng., 18, 63–77.