Multi-scale Convolutional Neural Networks for the Prediction of

Human-virus Protein Interactions

Xiaodi Yang

, Ziding Zhang

1,* b

and Stefan Wuchty

2,3* c

State Key Laboratory of Agrobiotechnology, College of Biological Sciences, China Agricultural University, Beijing

100193, China

Dept. of Computer Science, University of Miami, Miami FL, 33146, U.S.A.

Dept. of Biology, University of Miami, Miami FL, 33146, U.S.A.

Keywords: Human-virus PPI, Prediction, Deep Learning, PSSM, CNN, Transfer Learning.

Abstract: Allowing the prediction of human-virus protein-protein interactions (PPI), our algorithm is based on a

Siamese Convolutional Neural Network architecture (CNN), accounting for pre-acquired protein evolutionary

profiles (i.e. PSSM) as input. In combinations with a multilayer perceptron, we evaluate our model on a variety

of human-virus PPI datasets and compare its results with traditional machine learning frameworks, a deep

learning architecture and several other human-virus PPI prediction methods, showing superior performance.

Furthermore, we propose two transfer learning methods, allowing the reliable prediction of interactions in

cross-viral settings, where we train our system with PPIs in a source human-virus domain and predict

interactions in a target human-virus domain. Notable, we observed that our transfer learning approaches

allowed the reliable prediction of PPIs in relatively less investigated human-virus domains, such as Dengue,

Zika and SARS-CoV-2.

1 INTRODUCTION

Deep learning as a branch of machine learning

represents information through artificial neural

network modules, which share similar properties with

neural modules in the brain (Kriegeskorte and

Douglas, 2018; Yamins and DiCarlo, 2016). In the

past decade, applications of deep learning approaches

demonstrated improved performance in many fields

(e.g. biomedicine, image, speech recognition, etc)

(Karimi et al., 2019; Pospisil et al., 2018; Sainath et

al., 2015). In particular, convolutional neural

networks (CNN) (Hashemifar et al., 2018) and

recurrent neural networks (RNN) (Zhang et al., 2016)

automatically capture local features in images as well

as preserve contextualized/long-term ordering

information in sequence data. In addition, many

recent studies adopt a Siamese network architecture

based on CNN or RNN to capture mutual influence

between two individual inputs (Chen et al., 2019;

Hashemifar et al., 2018).

https://orcid.org/0000-0002-3229-5865

https://orcid.org/0000-0002-9296-571X

https://orcid.org/0000-0001-8916-6522

In general, traditional machine learning/deep

learning can only perform well, if training and test

sets are cut from the same feature space, ensuring

similar statistical distributions of feature values.

(Shao et al., 2015). While the rigid application of a

trained model on data sets with different distributions

usually perform poorly, transfer learning methods

utilize prior knowledge from a ‘source’ to train in a

‘target’ task domain (Chang et al., 2018; Shao et al.,

2015). In particular, transfer learning approaches

have been successfully applied to tackle problems in

many fields, such as medical imaging (Cheplygina et

al., 2019), biomedicine (Taroni et al., 2019), and

visual categorization (Shao et al., 2015). A regular

phenomenon appears in various training objectives

(Le et al., 2011; Lee et al., 2009) in that the first layers

of deep neutral networks (DNN) usually capture

standard features of the training data, providing a

foundation for transfer learning. Specifically, a deep

neural network can be trained on a source task,

establishing the parameters of the first layers.

Yang, X., Zhang, Z. and Wuchty, S.

Multi-scale Convolutional Neural Networks for the Prediction of Human-virus Protein Interactions.

DOI: 10.5220/0010185300410048

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 2, pages 41-48

ISBN: 978-989-758-484-8

Subsequently, parameters of late layers are trained on

the target task. Depending on the size of the target

dataset and number of parameters of the DNN, first

layers of the target DNN can either remain unchanged

during training on the new dataset (i.e. frozen), or

fine-tuned towards the new task, indicating a balance

between specificity and generality of derived prior

knowledge.

Here, we propose a framework to predict

interactions between virus and human proteins that is

based on a Siamese Convolutional Neural Network

architecture (CNN), accounting for pre-acquired

protein evolutionary profiles (i.e. PSSM) as protein

sequence input. In combination with a multilayer

perceptron (MLP), we assess the prediction

performance of our model on different human-virus

PPI datasets, outperforming other prediction

frameworks. Allowing to predict interactions in a

target domain of human-virus interactions, we

propose two types of transfer learning methods where

we freeze/fine-tune weights learned in the Siamese

CNN. Notably, the transfer of prior knowledge

learned from a large-scale human-virus PPI dataset

allowed the reliable prediction of PPIs between

human and proteins of less well investigated viruses

such as Dengue, Zika and SARS-CoV-2.

2 MATERIALS AND METHODS

2.1 Deep Neural Networks Framework

Representing interactions between human and viral

proteins through their amino-acid sequences, we

introduce an end-to-end deep neural network

framework, called a Siamese-based CNN that

consists of a pre-acquired protein sequence profile

module, a Siamese CNN module and a prediction

module (Fig. 1). In particular, the Siamese

architecture of the CNN module allows us to account

for residual relationships between interacting viral

and human protein sequences through protein

sequence profiles (i.e. PSSM) that capture

evolutionary relationships between proteins. Such

latent protein profile representations of interacting

protein pairs are fed to the Siamese CNN module to

generate respective high-dimensional sequence

embeddings. Finally, output embeddings of two

proteins are combined to form a sequence pair vector

as the input of a multilayer perceptron (MLP) with an

appropriate loss function to predict the

presence/absence of an interaction between a viral

and a human protein.

2.1.1 Pre-acquired Protein Sequence Profile

Module

For each protein sequence with variable lengths, we

generate a sequence profile, called PSSM. In

particular, we performed PSI-BLAST searches with

default parameters applying a threshold of E-value <

0.001 in the UniRef50 protein sequence database

(Suzek et al., 2015) as PSI-BLAST allows us to

discover protein sequences that are evolutionary

linked to the search sequence (Hamp and Rost, 2015;

Figure 1: Overall deep learning architecture to predict

interactions between viral and human host proteins.

Hashemifar et al., 2018). Sequence profiles for each

search sequence were processed by truncating

profiles of long sequences to a fixed length n and

zero-padding short sequences, a method widely used

for data pre-processing and effective training

(Matching, 2018; Min et al., 2017). As a result, we

obtained a 𝑛×20 dimensional array S for each

protein sequence, capturing the probability 𝑠

,

that

the residue in the i

position of the sequence is the j

out of the alphabet of 20 amino acids.

𝑆=

⎣

⎢

⎡

𝑠

,

⋯𝑠

,

⋯𝑠

,

⋮⋯⋮⋯⋮

𝑠

,

⋯𝑠

,

⋯𝑠

,

⋮⋯⋮⋯⋮

𝑠

,

⋯𝑠

,

⋯𝑠

,

⎦

⎥

⎤

2.1.2 Siamese CNN Module

To capture complex relationship between two

proteins we employ a Siamese CNN architecture with

two identical CNN sub-networks that share the same

parameters for a given pair of protein profiles 𝑆,𝑆



ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

Each sub-network produces a sequence embedding of

a single protein profile that are then concatenated.

While each single CNN module consists of a

convolutional and pooling layer, we leveraged four

connected convolutional modules to capture the

patterns in an input sequence profile.

Specifically, we use 𝑋, a 𝑛×𝑠 array of length 𝑛

with 𝑠 features in each position. The convolution

layer applies a sliding window of length 𝑤 (the size

of filters/kernels) to convert 𝑋 into a

(

𝑛−𝑤+1

)

𝑓𝑛 array 𝐶 where 𝑓𝑛 represents the number of

filters/kernels. Let 𝐶

,

denote the score of

filter/kernel 𝑘 , 1≤𝑘≤𝑓𝑛, that corresponds to

position 𝑖 of array 𝑋. Moreover, the convolutional

layer applies a parameter-sharing kernel 𝑀, a 𝑓𝑛 ×

𝑚 × 𝑠 array where 𝑀

,,

is the coefficient of pattern

𝑘 at position 𝑗 and feature 𝑙. The calculation of 𝐶 is

defined as

𝐶=𝐶𝑜𝑛𝑣



(

𝑆

)

𝐶

,

= 𝑀

,,









𝑋

,

Furthermore, the pooling layer is utilized to reduce

the dimension of 𝐶 to a

(

𝑛−𝑝+1

)

×𝑓𝑛 array 𝑃

where p is the size of pooling window. Array 𝑃=

𝑃𝑜𝑜𝑙

(

𝐶

)

is calculated as the maximum of all

positions 𝑖≤𝑗≤𝑖+𝑝 over each feature 𝑘 where

1≤𝑖≤

(

𝑛−𝑚+1

)

−𝑝,

𝑃

,

=𝑚𝑎𝑥𝐶

,

,…,𝐶

,

.

2.1.3 Prediction Module

The prediction module concatenates a pair of protein

sequence embedding vectors into a sequence pair

vector as the input of fully connected layers in an

MLP and computes the probability that two proteins

interact. The MLP contains three dense layers with

leaky ReLU where cross-entropy loss is optimized for

the binary classification objective defined as

𝐿𝑜𝑠𝑠 = −

𝐾

 𝑦





𝑙𝑜𝑔





∈

𝑠





where 𝑦



is numerical class label of the protein pair 𝑝.

The output of the MLP for the protein pair 𝑝 is a

probability vector 𝑠̂



, whose dimensionality is the

number of classes 𝑚. s is normalized by a softmax

function, where the normalized probability value for

the 𝑖



class is defined as 𝑠





exp (𝑠̂





)

∑

exp (𝑠̂





)



.

2.1.4 Implementation Details

As for pre-acquired sequence profile construction, we

consider a fixed sequence length of 2,000. As for the

construction of our learning approach, we employ

four convolutional modules, with input size 20, 64,

128 and 256. The convolution kernel size is set to 3

while the size of pooling window is set to 2 with 3

max-pooling layers and a global max-pooling layer.

To optimize the cross-entropy loss function we use

AMSGrad (Reddi et al., 2018) and set the learning

rate to 0.0001. The batch size was set to 64, while the

number of epochs was 100. The fully connected

layers contain three dense layers with input size

1,024, 512, 256 and output a two-dimensional vector

with the last softmax layer. The whole procedure was

implemented with keras (https://keras.io/) with GPU

configuration.

2.2 Data Set Construction

We collected experimentally verified human-virus

PPI data capturing 9,880 interactions in HIV, 5,966

in Herpes, 5,099 in Papilloma, 3,044 in Influenza,

1,300 in Hepatitis, 927 in Dengue and 709 in Zika

from five public databases, including HPIDB

(Ammari et al., 2016), VirHostNet (Guirimand et al.,

2015), VirusMentha (Calderone et al., 2015),

PHISTO (Durmuş Tekir et al., 2013) and PDB

(Altunkaya et al., 2017). As for interactions of

proteins of SARS-CoV-2, we used two recently

published interaction sets (Gordon et al., 2020; Liang

et al., 2020) that captured 291 and 598 PPIs,

respectively. To obtain high-quality PPIs, we

removed interactions from large-scale mass

spectroscopy experiments that were detected only

once, non-physical interactions and interactions

between proteins without available PSSM features.

Sampling negative interactions, we applied our

‘Dissimilarity-Based Negative Sampling’ method as

outlined in our previous work (Yang et al., 2020).

Briefly, we sampled a negative training set of PPIs

(i.e. pairs of proteins that do not interact) by

considering interactions in the positive training set.

Given that we found a protein B with a sequence that

was similar to interacting protein A, we considered

B and C non-interacting. In particular, we sampled a

negative PPI set that was 10 times larger than the

positive PPI training set.

2.3 Transfer Learning

To further improve the performance of our deep

neural network especially when dealing with smaller

datasets, we propose two transfer learning methods

that keep the weights constant (i.e. frozen) or allow

Multi-scale Convolutional Neural Networks for the Prediction of Human-virus Protein Interactions

their fine-tuning in the early layers and applied them

to eight human-virus PPI sets. (i) We used the

proposed DNN architecture to train the models based

on a given source set of human-virus interactions to

obtain pre-trained weights in the CNN layers that

learn the representation of the protein sequences. (ii)

In subsequent transfer learning steps, we keep the

weights of these CNN layers constant (i.e. frozen) and

only re-train parameters of the fully connected layers

of the MLP to predict interactions in a target human-

viral interaction set. As an alternative, our fine-tuning

approach allows us to retrain the weights of CNN

layers that we obtained from the initial training step

and change such weights by learning the interactions

in a target set of human-virus interactions. In analogy

to the ‘frozen’ approach, we re-train parameters of the

fully connected layers of the MLP as well.

2.4 Alternative Machine Learning and

Feature Encoding Methods

A great amount of research demonstrates that

Random Forest (RF) algorithms perform better than

other machine learning methods when applied to

binary classification problems (Chen et al., 2019; Wu

et al., 2009; Yang et al., 2020). Therefore, we

compare the performance of our deep learning

approaches to this representative state-of-art

classifier. Moreover, we consider three widely-used

encoding methods for feature representations as the

input to the RF classifier.

2.4.1 Random Forest

Random Forest (F) (Hamp and Rost, 2015; Wu et al.,

2009) is an ensemble learning method where each

decision tree is constructed using a different bootstrap

sample of the data (‘bagging’). In addition, random

forests change how decision trees are constructed by

splitting each node, using the best among a subset of

predictors randomly chosen at that node (‘boosting’).

Compared to many other classifiers this strategy turns

out to be robust against over-fitting, capturing

aggregate effects between predictor variables. We

utilize the GridSearchCV function to optimize the

parameters for the RF algorithm and set the

‘neg_log_loss’ scoring function as the assessment

criterion.

2.4.2 Alternative Feature Encoding

Approaches

Amino acid sequences provide primary structure

information of a protein that work well as feature

representations of binary PPIs. Here, we use three

commonly used sequence-based encoding schemes

including Local Descriptor (LD) (Cui et al., 2007;

Davies et al., 2008; Tong and Tammi, 2008; Yang et

al., 2010), Conjoint Triad (CT) (Sun et al., 2017) and

Auto Covariance (AC) (Guo et al., 2008; You et al.,

2013). Generally, these features cover specific, yet

different aspects of protein sequences such as

physicochemical properties of amino acids,

frequency information of local patterns, and

positional distribution information of amino acids.

3 RESULTS AND DISCUSSION

3.1 Performance of the Proposed Deep

Learning Method

Applying our deep learning approach to a set of

different human-viral protein interaction data sets, we

observed generally high prediction performance of

our deep learning approach (Table 1). However, we

also found that small training data sets such as

Dengue, Zika and SARS-CoV-2 translated into

decreasing prediction performance.

Table 1: Performance of our deep learning architecture

(PSSM+CNN+MLP) using 5-fold cross validation.

Human-viral

PPI dataset

Sensitivity Specificity AUPRC

HIV

89.72 99.54 0.974

Herpes

68.10 97.98 0.768

Papilloma

70.48 98.53 0.818

Influenza

70.30 98.68 0.834

Hepatitis

49.77 97.79 0.636

Dengue

45.85 98.04 0.605

Zika

59.94 98.96 0.746

SARS-CoV-2

55.12 98.53 0.672

To compare the performance of our proposed

deep learning method (i.e. PSSM+CNN+MLP), we

trained a RF model using three widely used sequence-

based feature encoding schemes (i.e. LD, CT and AC)

on human-virus PPI datasets using 5-fold cross

validation. Comparing corresponding AUPRC

values, we observe that our method generally

outperformed other those RF based classifiers

especially when applied to comparatively large

datasets (Table 2). To further assess the impact of our

encoding scheme to represent the features of

interacting proteins, we compared the performance of

our deep learning architecture using PSSMs and a

different word embedding technique, word2vec+CT

one-hot. Specifically, this method considers each

amino acid as a word and learns a word-embedding

of sequences based on the training data, where each

amino acid is finally encoded by a 5-dimensional

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

Table 2: Performance comparison of our deep learning

architecture (PSSM + CNN + MLP) and random forests

(RF) that were combined with three sequence encoding

schemes (LD, CT, AC) using 5-fold cross validation.

AUPRC

Human-

viral PPI

dataset

Our

method

LD+RF CT+RF AC+RF

HIV 0.974 0.972 0.97 0.972

Herpes 0.768 0.741 0.737 0.699

Papilloma 0.818 0.74 0.724 0.656

Influenza 0.834 0.813 0.795 0.713

Hepatitis 0.636 0.571 0.58 0.537

Dengue 0.605 0.526 0.505 0.456

Zika 0.746 0.720 0.718 0.698

SARS-

CoV-2

0.672 0.668 0.678 0.652

vector. Moreover, the 20 amino acids can be clustered

into 7 groups based on their dipoles, volumes of the

side chains and other chemical descriptors.

Furthermore, CT one-hot is a 7-dimensional one-hot

encoding based on the classification of these 20

amino acids. As a result, word2vec+CT one hot is the

concatenation of pre-trained word embeddings and

CT one-hot encodings for each protein that is

represented by a 𝑛×13 dimensional array. As noted

previously, we considered a fixed sequence length of

n = 2,000 and zero-padded smaller sequences. In

comparison to word2vec+CT one hot, Table 3

indicates that our learning approach combined with

PSSM allows better prediction performance

especially in comparatively small datasets such as

Dengue, Zika and SARS-CoV-2.

3.2 Comparison with Several Existing

Human-virus PPI Prediction

Methods

To further assess the performance of our proposed

method, we compared our method with three existing

human-virus PPI prediction approaches. Recently, we

proposed a sequence embedding-based RF method to

predict human-virus PPIs with comparatively

promising performance (Yang et al., 2020). The main

point of our approach is the application of an

unsupervised sequence embedding technique (i.e.

doc2vec) to represent protein sequences as low-

dimensional vectors with rich features. Such

representations of protein pairs were subjected to a

RF method that predicted the presence/absence of an

interaction. In Alguwaizani et al.’s work

(Alguwaizani et al., 2018), the authors utilized a

Support Vector Machine (SVM) model to

predict human-virus PPIs based on a simple way to

feature-encode protein sequences through repeat

patterns and local patterns of amino acid

combinations. As for the DeNovo method (Eid et al.,

2016), the authors introduced a domain/linear motif-

based SVM approach to predict human-virus PPIs. To

compare, we first constructed the PSSMs of the

Table 3: Performance comparison of combinations of

different feature encodings (PSSM, word2vec+CT one-hot)

and our deep learning architecture (CNN + MLP).

AUPRC

Human-viral

PPI dataset

PSSM

word2vec+

CT one hot

HIV 0.974 0.968

Herpes 0.768 0.734

Papilloma 0.818 0.778

Influenza 0.834 0.808

Hepatitis 0.636 0.587

Dengue 0.605 0.481

Zika 0.746 0.662

SARS-CoV-2 0.672 0.602

protein sequences of DeNovo’s PPI dataset to train

our learning model. Finally, we assessed the

performance of our reconstructed deep learning

model on the test set provided in (Eid et al., 2016)

including 425 positive and 425 negative samples.

Table 4: Performance comparison of our method (PSSM +

CNN+MLP) with existing human-virus PPI prediction

methods.

Method

Accuracy

(%)

Sensitivity

(%)

Specificity

(%)

Our model 94.12 90.82 97.41

doc2vec+RF

93.23 90.33 96.17

SVM

86.47 86.35 86.59

DeNovo

81.90 80.71 83.06

The corresponding values were retrieved from (Yang et

al., 2020).

The corresponding values were retrieved from

(Alguwaizani et al., 2018).

The corresponding values were

retrieved from (Eid et al., 2016).

Furthermore, we tested our previous RF based

prediction method and Alguwaizani et al’s SVM

approach on these data sets as well. Table 4 clearly

suggests that our deep learning and previously

published RF based method outperformed

Alguwaizani et al.’s SVM and the DeNovo approach.

Multi-scale Convolutional Neural Networks for the Prediction of Human-virus Protein Interactions

Table 4: Performance comparison of our method (PSSM + CNN+MLP) with existing human-virus PPI prediction methods.

3.3 Cross-viral Tests and Transfer

Learning

To explore potential factors that affect prediction

performance in a cross-viral setting, we trained our

deep learning model on one human-virus PPI data set

and predicted protein interactions in a different

human-virus system. Expectedly, such cross-viral

tests dropped considerably in performance compared

to training and testing in the same human-viral system

(Fig. 2). To allow reliable cross-viral predictions of

PPIs, we introduce two transfer learning methods

where we trained the parameters of CNN layers of the

DNN model on a source human-virus PPI dataset.

Subsequently, we transfer all parameters to initialize

a new model (i.e. frozen or fine tuning) to train on a

target human-virus PPI dataset. To comprehensively

test our transfer learning approaches, we considered

each combination of human-viral PPI sets as source

and target data. The left panel in Fig. 3 indicates that

a relatively rigid transfer learning methodology by

keeping the parameters of the feature encoding CNN

untouched (i.e. frozen) strongly outperformed

baseline performance as shown in Fig. 2. In turn, fine-

tuning parameters using a given target human-viral

domain allowed for another marked increase in

performance (right panel, Fig. 3) compared to the

‘frozen’ approach. As for individual pairs of human-

viral domains, we also observed that the frozen

transfer methodology worked well if the target

domain data set was large, independently of the

training domain. In turn, performance dropped when

the target human-viral domain datasets of PPIs were

small. Notably, prediction performance improved

when we applied our fine-tuning transfer learning

approach on small target domains data sets such as

human-Hepatitis, human-Dengue, human-Zika and

human-SARS-CoV-2.

Figure 2: AUPRC performance of cross-viral tests. Rows

indicate human-viral PPIs that were used for training while

columns indicate human-viral PPI test sets.

4 CONCLUSIONS

Here, we proposed a Siamese-based multi-scale CNN

architecture by using PSSM to represent the

sequences of interacting proteins, allowing us to

predict interactions between human and viral proteins

with an MLP approach. In comparison, we observed

that our model outperformed previous state-of-the-art

human-virus PPI prediction methods. Furthermore,

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

we confirmed that the performance of the

combination of our deep learning framework and the

representation of the protein features as PSSMs was

mostly superior to combinations of other machine

learning and pre-trained feature embeddings. While

we found that our model that was trained on a given

source human-viral interaction data set performed

dismally in predicting protein interactions of proteins

in a target human-virus domain, we introduced two

transfer learning methods (i.e. frozen type and fine-

tuning type). Notably, our methods increased the

cross-viral prediction performance dramatically,

compared to the naïve baseline model. In particular,

for small target datasets, fine-tuning pre-trained

parameters that were obtained from larger source sets

increased prediction performance.

REFERENCES

Alguwaizani, S., Park, B., Zhou, X., Huang, D.S., and Han,

K. (2018). Predicting interactions between virus and

host proteins using repeat patterns and composition of

amino acids. J. Healthc. Eng. 2018, 1391265.

Altunkaya, A., Bi, C., Bradley, A.R., Rose, P.W., Prli, A.,

Christie, H., Costanzo, L. Di, Duarte, J.M., Dutta, S.,

Feng, Z., et al. (2017). The RCSB protein data bank:

integrative view of protein, gene and 3D structural

information. Nucleic Acids Res. 45, D271–D281.

Ammari, M.G., Gresham, C.R., McCarthy, F.M., and

Nanduri, B. (2016). HPIDB 2.0: a curated database for

host-pathogen interactions. Database 2016, baw103.

Calderone, A., Licata, L., and Cesareni, G. (2015).

VirusMentha: a new resource for virus-host protein

interactions. Nucleic Acids Res. 43, D588–D592.

Chang, H., Han, J., Zhong, C., Snijders, A.M., and Jian-

Hua, M. (2018). Unsupervised transfer learning via

multi-scale convolutional sparse coding for biomedical

applications. IEEE Trans. Pattern Anal. Mach. Intell.

40, 1182–1194.

Chen, M., Ju, C.J.T., Zhou, G., Chen, X., Zhang, T., Chang,

K.W., Zaniolo, C., and Wang, W. (2019). Multifaceted

protein-protein interaction prediction based on Siamese

residual RCNN. Bioinformatics 35, i305–i314.

Cheplygina, V., de Bruijne, M., and Pluim, J.P.W. (2019).

Not-so-supervised: a survey of semi-supervised, multi-

instance, and transfer learning in medical image

analysis. Med. Image Anal. 54, 280–296.

Cui, J., Han, L.Y., Li, H., Ung, C.Y., Tang, Z.Q., Zheng,

C.J., Cao, Z.W., and Chen, Y.Z. (2007). Computer

prediction of allergen proteins from sequence-derived

protein structural and physicochemical properties. Mol.

Immunol. 44, 514–520.

Davies, M.N., Secker, A., Freitas, A.A., Clark, E., Timmis,

J., and Flower, D.R. (2008). Optimizing amino acid

groupings for GPCR classification. Bioinformatics 24,

1980–1986.

Durmuş Tekir, S., Çakir, T., Ardiç, E., Sayilirbaş, A.S.,

Konuk, G., Konuk, M., Sariyer, H., Uǧurlu, A.,

Karadeniz, I., Özgür, A., et al. (2013). PHISTO:

pathogen-host interaction search tool. Bioinformatics

29, 1357–1358.

Eid, F., Elhefnawi, M., and Heath, L.S. (2016). DeNovo:

virus-host sequence-based protein-protein interaction

prediction. 32, 1144–1150.

Gordon, D.E., Jang, G.M., Bouhaddou, M., Xu, J.,

Obernier, K., White, K.M., O’Meara, M.J., Rezelj, V.

V., Guo, J.Z., Swaney, D.L., et al. (2020). A SARS-

CoV-2 protein interaction map reveals targets for drug

repurposing. Nature 583, 459–468.

Guirimand, T., Delmotte, S., and Navratil, V. (2015).

VirHostNet 2.0: surfing on the web of virus/host

molecular interactions data. Nucleic Acids Res. 43,

D583–D587.

Guo, Y., Yu, L., Wen, Z., and Li, M. (2008). Using support

vector machine combined with auto covariance to

predict protein-protein interactions from protein

sequences. Nucleic Acids Res. 36, 3025–3030.

Hamp, T., and Rost, B. (2015). Evolutionary profiles

improve protein-protein interaction prediction from

sequence. 31, 1945–1950.

Hashemifar, S., Neyshabur, B., Khan, A.A., and Xu, J.

(2018). Predicting protein-protein interactions through

sequence-based deep learning. Bioinformatics 34,

i802–i810.

Karimi, M., Wu, D., Wang, Z., and Shen, Y. (2019).

DeepAffinity: interpretable deep learning of

compound-protein affinity through unified recurrent

and convolutional neural networks. Bioinformatics 35,

3329–3338.

Kriegeskorte, N., and Douglas, P.K. (2018). Cognitive

computational neuroscience. Nat. Neurosci. 21, 1148–

1160.

Le, Q. V., Karpenko, A., Ngiam, J., and Ng, A.Y. (2011).

ICA with reconstruction cost for efficient overcomplete

feature learning. Adv. Neural Inf. Process. Syst. 24 25th

Annu. Conf. Neural Inf. Process. Syst. 2011, NIPS

2011 2027–2035.

Lee, H., Grosse, R., Ranganath, R., and Ng, A.Y. (2009).

Convolutional deep belief networks for scalable

unsupervised learning of hierarchical representations.

Proc. 26th Annu. Int. Conf. Mach. Learn. ICML 54,

609–616.

Liang, Q., Li, J., Guo, M., Tian, X., Liu, C., Wang, X.,

Yang, X., Wu, P., Xiao, Z., Qu, Y., et al. (2020). Virus-

host interactome and proteomic survey of PMBCs from

COVID-19 patients reveal potential virulence factors

influencing SARS-CoV-2 pathogenesis. BioRxiv

2020.03.31.019216.

Matching, S. (2018). Neural article pair modeling for

wikipedia sub-article matching. In: ECML-PKDD 3–

19.

Min, X., Zeng, W., Chen, N., and Chen, T. (2017).

Chromatin accessibility prediction via convolutional

long short-term memory networks with k -mer

embedding. Bioinformatics 33, i92–i101.

Multi-scale Convolutional Neural Networks for the Prediction of Human-virus Protein Interactions

Pospisil, D.A., Pasupathy, A., and Bair, W. (2018).

’Artiphysiology’ reveals V4-like shape tuning in a deep

network trained for image classification. Elife 7,

e38242.

Reddi, S.J., Kale, S., and Kumar, S. (2018). On the

convergence of Adam and Beyond. 6th Int. Conf.

Learn. Represent. ICLR 2018 - Conf. Track Proc. 1–23.

Sainath, T.N., Kingsbury, B., Saon, G., Soltau, H.,

Mohamed, A. rahman, Dahl, G., and Ramabhadran, B.

(2015). Deep convolutional neural networks for large-

scale speech tasks. Neural Networks 64, 39–48.

Shao, L., Zhu, F., and Li, X. (2015). Transfer learning for

visual categorization: a survey. IEEE Trans. Neural

Networks Learn. Syst. 26, 1019–1034.

Sun, T., Zhou, B., Lai, L., and Pei, J. (2017). Sequence-

based prediction of protein protein interaction using a

deep-learning algorithm. BMC Bioinformatics 18, 277.

Suzek, B.E., Wang, Y., Huang, H., McGarvey, P.B., and

Wu, C.H. (2015). UniRef clusters: a comprehensive and

scalable alternative for improving sequence similarity

searches. Bioinformatics 31, 926–932.

Taroni, J.N., Grayson, P.C., Hu, Q., Eddy, S., Kretzler, M.,

Merkel, A., Greene, C.S., Therapeutics, T., Diseases,

S., Arbor, A., et al. (2019). MultiPLIER: a transfer

learning framework for transcriptomics reveals

systemic features of rare disease Jaclyn. 8, 380–394.

Tong, J.C., and Tammi, M.T. (2008). Prediction of protein

allergenicity using local description of amino acid

sequence. 13, 6072–6078.

Wu, J., Liu, H., Duan, X., Ding, Y., Wu, H., Bai, Y., and

Sun, X. (2009). Prediction of DNA-binding residues in

proteins from amino acid sequences using a random

forest model with a hybrid feature. Bioinformatics 25,

30–35.

Yamins, D.L.K., and DiCarlo, J.J. (2016). Using goal-

driven deep learning models to understand sensory

cortex. Nat. Neurosci. 19, 356–365.

Yang, L., Xia, J.-F., and Gui, J. (2010). Prediction of

protein-protein interactions from protein sequence

using local descriptors. Protein Pept. Lett. 17, 1085–

1090.

Yang, X., Yang, S., Li, Q., Wuchty, S., and Zhang, Z.

(2020). Prediction of human-virus protein-protein

interactions through a sequence embedding-based

machine learning method. Comput. Struct. Biotechnol.

J. 18, 153–161.

You, Z.-H., Li, L., Ji, Z., Li, M., and Guo, S. (2013).

Prediction of protein-protein interactions from amino

acid sequences with ensemble extreme learning

machines and principal component analysis. BMC

Bioinformatics 14, 80–85.

Zhang, S., Zhou, J., Hu, H., Gong, H., Chen, L., Cheng, C.,

and Zeng, J. (2016). A deep learning framework for

modeling structural features of RNA-binding protein

targets. Nucleic Acids Res. 44, e32.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence