Neuro-symbolic XAI for Computational Drug Repurposing

Martin Dranc

, Marina Boudin, Fleur Mougin

and Gayo Diallo

Inserm U1219, Bordeaux Population Health Research Center, Team ERIAS, University of Bordeaux, France

Keywords:

Artiﬁcial Intelligence, XAI, Drug Repurposing, Knowledge Graph, Bioinformatics.

Abstract:

Today in the health domain, the challenge is to build a more transparent artiﬁcial intelligence, less affected by

the opacity intrinsic to the mathematical concepts it uses. Among the ﬁelds which use AI techniques, is drug

development, and more speciﬁcally drug repurposing. DR involves ﬁnding a new indication for an existing

drug. The hypotheses generated by DR techniques must be validated. Therefore, the mechanism of generation

must be understood. In this paper, we describe the use of a state-of-the-art neuro-symbolic algorithm in order

to explain the process of link prediction in a knowledge graph-based computational drug repurposing. Link

prediction consists of generating hypotheses about the relationships between a known molecule and a given

target. More speciﬁcally, the implemented approach allows to understand how the organization of data in a

knowledge graph changes the quality of predictions.

1 INTRODUCTION

Today, it costs about 1.5 billion dollars and about ﬁf-

teen years of research to develop a new drug (DiMasi

et al., 2016). These constraints are limiting factors that

do not encourage the pharmaceutical industry, which

is primarily concerned with cost-effectiveness. In par-

ticular, they do not invest in the development of new

drugs for so-called rare diseases. In this context, over

the last decade, drug repurposing (DR), which con-

sists of searching for a new indication for an exist-

ing drug (link between an existing molecule and a

given indication), has been attracting growing inter-

est. Advances in Artiﬁcial Intelligence (AI) methods

have made it possible to develop AI-based computa-

tional DR with very encouraging results (Ashburn and

Thor, 2004)(Stokes et al., 2020).

However, the use of “black-box” models, which

are completely opaque to the understanding of their

users (biomedical experts, clinicians, etc.), is currently

one of the main obstacles to the use of AI in the health

ﬁeld. In this ﬁeld indeed, AI should not be seen as

a decision maker per-se, but rather as a transparent

support to the users’ decision process. Recently, new

explainable AI (XAI) models, referred to as neuro-

symbolic models, have been developed. These latter

AI models are transparent by nature, i.e. users can

https://orcid.org/0000-0001-6365-531X

https://orcid.org/0000-0002-7436-3010

https://orcid.org/0000-0002-9799-9484

Table 1: Type and number of entities (nodes) in the

OREGANO KG.

Type Number

ATC Code 1,005

Drug 42,856

Gene 35,602

Effect 153

Target 254,289

Disease 8,997

Pathway 297

Activity 74

Phenotype 11,202

Indication 2,154

Side Effect 5,556

ﬁnd and understand the mechanisms leading to a pre-

diction without having to modify, simulate or process

any information about the model’s operation itself.

Neuro-symbolic models combine symbolic and statis-

tical learning (Das et al., 2018). This combination al-

lows for robust predictions made by a neural network,

supported by the transparency provided by the use of

logical rules, understandable by humans. Among the

many advantages of using these neuro-symbolic mod-

els, a major beneﬁt is the possibility of interaction be-

tween the model and users during the learning pro-

cess. Indeed, models that provide an explanation of

how they work can more easily be adjusted in the event

that some bias is detected.

220

Drancé, M., Boudin, M., Mougin, F. and Diallo, G.

Neuro-symbolic XAI for Computational Drug Repurposing.

DOI: 10.5220/0010714100003064

In Proceedings of the 13th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2021) - Volume 2: KEOD, pages 220-225

ISBN: 978-989-758-533-3; ISSN: 2184-3228

 2021 by SCITEPRESS – Science and Technology Publications, Lda. All r ights reser ved

These models are now being applied to knowl-

edge graphs (KGs) in the context of DR. According

to (Paulheim, 2017), a KG has the following charac-

teristics: (i) it describes real-world entities and their

interactions within a graph, (ii) it deﬁnes classes and

relationships between these entities in the form of a

schema, (iii) it allows entities to be arbitrarily linked

together, and (iv) it can aggregate information from

different domains. These data structures are well

suited for describing biomedical information, as they

are usually derived from multiple databases and KGs

are perfect for maintaining the semantic relationship

between entities. Recently, as proved by (Himmel-

stein et al., 2017) and (Zhang et al., 2021), KG-based

DR has demonstrated interesting results through the

use of AI models and algorithms to predict new links

between existing entities in KGs. Unfortunately, the

lack of transparency of the models prevents the results

from being used more widely by experts in the ﬁeld of

DR, including biochemists and clinicians.

The current study is part of the OREGANO project

which aims to provide an end-to-end framework for

KG-based computational DR (Boudin, 2020). Specif-

ically, we describe the exploitation of a state-of-

the-art neuro-symbolic model and algorithm in the

OREGANO workﬂow to explain the results of the

link prediction process. Link prediction here consists

of generating hypotheses about the relationships be-

tween a known molecule and a given target. To do

this, we used PoLo (Policy-guided walks with Logi-

cal rules) (Liu et al., 2021), a neuro-symbolic model

achieving state-of-the-art results in KG link predic-

tion. It has already been used in the context of the

Hetionet KG (Himmelstein et al., 2017) to repurpose

drugs. In our work, we particularly investigated how

the structure of data in a KG changes the quality of

predictions.

2 MATERIALS AND METHODS

2.1 Data

The data that are used to develop the OREGANO KG

have been previously collected in the context of the

work described in (Boudin, 2020). They come from

various free and/or open source knowledge bases and

Linked Open Data. They have been then curated and

integrated into the OREGANO KG depicted in Fig-

ure 1. Currently, this graph has a total of 362,186

nodes of 11 different types and 802,949 relations of

19 different types. The detailed number of nodes per

data type can be found in Table 1 while the number of

relations (edges) is provided in Table 2.

Table 2: Type and number of relations (edges) in the

OREGANO KG.

Type Number

associated to 250,000

causes condition 8,584

decreases activity 2,446

decreases effect 102

decreases efﬁcacy 106,558

has activity 5,565

has code 3,150

has effect 11,251

has indication 8,307

has phenotype 76,177

has side effect 129,330

has target 50,132

increases activity 6,636

increases effect 10,041

increases efﬁcacy 34,071

is affecting 5,341

is substance that treats 60,889

participates in 24,919

2.2 Link Prediction

PoLo (Liu et al., 2021) is an algorithm that was pro-

posed in March 2021 to perform link prediction from

a KG. It is one of the only fully neuro-symbolic algo-

rithms that relies as much on mathematical methods as

on the use of logical rules from the information con-

tained in the KG on which it operates. Its operation

is inspired by MINERVA (Das et al., 2018), both of

which are based on the use of reinforcement learning

in order to optimize the movements of an agent in the

graph and thus make link prediction. From an initial

entity (here, a drug to be repurposed), an agent moves

in the graph from node to node. The decision of which

node to choose for the next move is determined by a

policy set by an long short-term memory (Hochreiter

and Schmidhuber, 1997). The transitions between the

nodes accumulate and form a path that represents a

chain of reasoning. These iterations are repeated a ﬁ-

nite number of times, until the agent obtains a reward,

based on the path taken and the associated prediction.

This learning task is modeled as a Markov decision

process.

2.3 Logical Rules

PoLo uses meta-paths, in the form of Horn rules, to

evaluate the gain provided to the agent according to

its predictions. In order to work, PoLo needs a set

of meta-paths that are considered reliable and that de-

scribe possible paths between the entities to be re-

Neuro-symbolic XAI for Computational Drug Repurposing

221

Figure 1: Schema of the OREGANO KG.

purposed, in this case between a drug and a disease.

These meta-paths must be “scored” according to their

reliability, a very reliable path having a score close to

1 and a very unreliable path a score close to 0. An

example of a meta-path is described as follows:

Aspirin

treats

−−−→ Headache

caused by

−−−−−−→ Dehydration

and becomes :

Drug

treats

−−−→ Symptom

caused by

−−−−−−→ Cause

As mentioned in (Liu et al., 2021), AnyBURL (Any-

time Bottom-Up Rule Learning algorithm) (Meilicke

et al., 2019) was applied to the OREGANO KG to

identify a set of rules intrinsic to the data. These rules

were transformed into meta-paths that were used by

PoLo during its learning phase. Based on these meta-

paths, PoLo evaluates the reward to be given to the

agent responsible for the link prediction.

2.4 Discovery Patterns

In addition to PoLo’s explainable logic rules, discov-

ery patterns have been used as proposed in (Himmel-

stein et al., 2017) and (Zhang et al., 2021). Discovery

patterns based on the exploration of semantic links be-

tween data allow the discovery of mechanistic links

between entities in a KG, but also to strengthen the

explanation of the prediction provided by the AI al-

gorithm. To provide a visual exploration, we relied

on Neo4j

, together with its query language Cypher.

https://neo4j.com/

It is a system for building and manipulating data in

the form of a graph-oriented database (GDB). These

GDBs have the advantage of preserving the structure

of the KG, by storing the information contained in the

nodes but also in the edges. Thus, the modeling of

complex data and queries are facilitated. From the

OREGANO project data, a GDB has been built on

Neo4j. The purpose of this GDB is to bring an ad-

ditional explanation and visualization, in order to in-

validate or validate the predictions made by PoLo. The

advantage of this method is that it offers the possibility

to build precise queries, allowing access to informa-

tion or links that had not been spotted by AnyBURL

or PoLo, but that do exist in the data.

2.5 Impact of the Data in the KG

As with any machine learning task, it is important to

consider the format of the data. Much of the current

research attempts to improve the results of link predic-

tion by improving the performance of the algorithms.

However, in the areas of AI that have grown in recent

years, there has been an evolution of algorithms along

with data processing methods. These methods ensure

optimal learning, for example by modifying or aug-

menting the data. For KGs, as far as we know, little or

no work has been done to determine the best way to

structure the data in KG and its possible impact on the

predictions. A ﬁrst approach studied here concerns the

importance of the ratio between the entities of the KG.

Let r be the relationship of interest linking two types

KEOD 2021 - 13th International Conference on Knowledge Engineering and Ontology Development

222

of entities T 1 and T 2, what is the optimal number of

objects of types T 1 and T 2 in the graph to achieve a

good prediction? To answer this question in the con-

text of DR, several types of link prediction have been

tested. First, AnyBURL and PoLo were used to pre-

dict the following types of links:

• associated to between a protein and a gene;

• causes condition between a gene and a disease;

• has phenotype between a disease and a phenotype;

• has target between a drug and a protein;

• is affecting between a drug and a gene;

• is substance that treats between a drug and a dis-

ease;

• participates in between a gene and a pathway.

In a second approach, the relation between the amount

of training data and the quality of the prediction was

tested. For this purpose, the link prediction task is re-

run focusing exclusively on the links has target and

is substance that treats, but changing the number of

triples of these types. For each type of relation,

the ﬁrst experiments have been launched with 5000

triples, then 5000 triples are added per additional ex-

periment, until all the triples are reached.

3 RESULTS

Table 3 shows the Mean Reciprocal Rank (MRR)

score for each PoLo link prediction based on the

different relation types. These results show that

the quality of predictions cannot be solely related

to the quality of the extracted rules or their num-

ber. Indeed, the best predictions are reached for the

is substance that treats relation for which 630 rules

are available with the best conﬁdence index equal

to 0.70. However, for the is affecting relation, the

MRR is close to 0 while many rules are also avail-

able and have a satisfactory conﬁdence index. In con-

trast, the predictions for the relation causes condition

are ranked third in terms of MRR, with only 112 Horn

rules available and a conﬁdence index three times

lower than that of is affecting. Predictions involving

participates in achieve an MRR of 0.3 based on only

68 rules.

Regarding the impact of the amount of training

data, the results are presented in Figure 2. These

results show that, for both relations, increasing the

amount of training data has rather a negative impact

on the quality of predictions. Furthermore, we can

clearly see the correlation between the MRR and the

number of times a rule given by AnyBURL produces

the correct result.

As a ﬁnal result for the DR problem, best

results were found when predicting the relation

is substance that treats and can be found in Table 4.

These results, which are slightly better than with Het-

ionet, are only partial results, the other (major) part

coming from the explanations provided by PoLo for

each of the predictions made. From this generated ﬁle,

the following results emerged:

• 682 drugs were identiﬁed as candidates for being

repurposed;

• these candidates impact a total of 447 diseases;

• out of 630 Horn rules provided by AnyBURL to

PoLo, only 10 of them were effectively used for

predictions;

• on these 10 rules, the most frequent one has been

used 600 000 times, the second only 21 000 times.

4 DISCUSSION

The use of PoLo to repurpose drugs through link

prediction has produced a signiﬁcant amount of re-

sults. These results reached an accuracy not previ-

ously achieved from the OREGANO KG using other

methods such as TransE (Bordes et al., 2013). More-

over, the use of a neuro-symbolic model offers reliable

explanations for the origin of each prediction. These

results need to be analyzed by DR experts in order to

get the best out of them. It is with this in mind that

these results are currently being further investigated

by the clinical genetics unit of the Bordeaux Univer-

sity Hospital. The aim of this collaboration is to give

access to the results to clinicians who will be able to

evaluate and select those they consider most interest-

ing to explore in their research. The possible explana-

tion for each repurposing will come from logical rules

used by PoLo but also from the post-hoc mechanistic

explanation provided by the visualization capability of

the Neo4j graph database.

Regarding the fact that providing more training

data does not help to obtain better predictions, two im-

portant points should be highlighted:

• The quality and number of rules produced by Any-

BURL are not systematically correlated with the

quality of a prediction;

• The increase in the number of triples in the training

data does not necessarily mean an improvement in

performance.

This last point is particularly interesting since it is

counter-intuitive. We observe that, in both cases, the

MRR drops when the number of rules producing a cor-

rect prediction decreases, which is normal if we con-

sider that the PoLo agent’s gain is directly related to

Neuro-symbolic XAI for Computational Drug Repurposing

223

Table 3: MRR (Mean Reciprocal Rank) produced by PoLo for each type of relation, based on the best quality and number of

rules extracted by AnyBURL.

Relation Number Best Score MRR

associated to 21 0.43 0.0003

causes condition 112 0.14 0.15

has phenotype 96 0.11 0.07

has target 489 0.163 0.07

is affecting 399 0.42 0.004

is substance that treats 630 0.70 0.49

participates in 68 0.92 0.3

Figure 2: Impact of the number of relations in the training data for is substance that treats on the left and has target on the

right.

the path taken for the prediction. These results there-

fore suggest that a surplus of training data makes the

use of the rules proposed by AnyBURL rarer, leading

to less frequent rewards and thus to a decrease in per-

formance of PoLo predictions. Another explanation

for this decrease in performance could be how more

training data brings more noise to the system. PoLo

uses training triples to determine what are the diseases

for which drugs can be repurposed. The more training

data, the more different diseases can be proposed as

a result. Based on this observation, a useful step that

could be added would be to ﬁlter the training data to

keep only those that contain diseases of interest.

The disparity of the metrics according to the type

of relation chosen demonstrates the importance of the

structure of the data in the prediction task. Today,

AI research in the ﬁeld of drug repositioning focuses

mainly on improving prediction algorithms, without

really paying attention to the data and its quality.

These new models are systematically tested on the

same evaluation datasets, without ever reconsidering

their structure. However, it is necessary to under-

stand how the organization of the data within a KG

changes the quality of the predictions, not least be-

cause biomedical data in general are different from the

“test” datasets typically used in the literature. In addi-

tion to these considerations around data, it is important

to remember that the goal of more transparent AI is to

be trusted by clinicians and other healthcare experts.

With this in mind, algorithms should be proposed that

allow for greater interaction with clinicians, in order to

offer greater ﬂexibility in the choice of how they oper-

ate, as here with the logical rules to be followed. Cur-

rently, the set of “good” logical rules to follow is given

to PoLo using AnyBURL. AnyBURL is a powerful

tool for extracting the main logical rules and meta-

paths that exist in a KG, but these rules are ranked

based on their ability to describe the data, not on their

relevance from a medical perspective. Since PoLo is

trained using these rules in the reinforcement learning

process, it will be important to ﬁlter these rules to keep

only those that are medically plausible. Another way

to do this is to have the clinicians construct the rules

themselves, based on their a priori knowledge. The

fact that PoLo only used 10 of the 630 proposed rules

and only one of them most of the time (600 000 times)

shows that these rules are not efﬁcient in repurposing

drugs. This is also a problem when considering the use

of reinforcement learning based on these rules, as the

agent should be rewarded when it uses different med-

ically valid rules and not when it uses only the most

common one among the data.

KEOD 2021 - 13th International Conference on Knowledge Engineering and Ontology Development

224

Table 4: Comparison of the metrics obtained for drug repositioning through the is substance that treats relation on the

OREGANO KG with the results obtained on Hetionet.

Data MRR Hits@1 Hits@3 Hits@10

OREGANO 0.4980 0.3898 0.5551 0.7300

HETIONET 0.4300 0.3370 0.4700 0.6410

5 CONCLUSION

In this paper, we have described a neuro-symbolic

XAI solution applied to the DR task using a KG. The

study has been implemented and evaluated in the con-

text of the OREGANO project using a state-of-the-art

algorithm which enables the explainability. Results

equivalent to those of the state of the art were obtained,

as well as several ways to provide an explanation for

each prediction, i.e. intrinsic to the model’s opera-

tion, but also post-hoc in a mechanistic way using fea-

tures of a graph oriented database. The challenges sur-

rounding this work are numerous. First, it is important

to better understand how the organization of data in a

KG affects the prediction task. This will be partic-

ularly important for the application of these methods

to DR for rare diseases, where data is by deﬁnition

less abundant. Also, more ﬂexible methods must be

thought of, allowing biochemists and other clinicians

to easily participate in the learning process, bringing

their knowledge to the sum of data available for DR.

REFERENCES

Ashburn, T. and Thor, K. (2004). Drug repositioning: Identi-

fying and developing new uses for existing drugs. Na-

ture reviews Drug discovery, 3:673–83.

Bordes, A., Usunier, N., Garc

ıa-Dur

an, A., Weston, J., and

Yakhnenko, O. (2013). Translating embeddings for

modeling multi-relational data. In Advances in Neu-

ral Information Processing Systems 26.

Boudin, M. (2020). Computational approaches for drug

repositioning: Towards a holistic perspective based on

knowledge graphs. Proceedings of the 29th ACM In-

ternational Conference on Information & Knowledge

Management, page 3225–3228.

Das, R., Dhuliawala, S., Zaheer, M., Vilnis, L., Durugkar,

I., Krishnamurthy, A., Smola, A., and McCallum, A.

(2018). Go for a walk and arrive at the answer: Rea-

soning over paths in knowledge bases using reinforce-

ment learning. In International Conference on Learn-

ing Representations.

DiMasi, J. A., Grabowski, H. G., and Hansen, R. W. (2016).

Innovation in the pharmaceutical industry: New es-

timates of r&d costs. Journal of Health Economics,

47:20–33.

Himmelstein, D. S., Lizee, A., Hessler, C., Brueggeman, L.,

Chen, S. L., Hadley, D., Green, A., Khankhanian, P.,

and Baranzini, S. E. (2017). Systematic integration of

biomedical knowledge prioritizes drugs for repurpos-

ing. eLife, 6:e26726.

Hochreiter, S. and Schmidhuber, J. (1997). Long Short-

Term Memory. Neural Computation, 9(8):1735–1780.

Liu, Y., Hildebrandt, M., Joblin, M., Ringsquandl, M., Rais-

souni, R., and Tresp, V. (2021). Neural multi-hop

reasoning with logical rules on biomedical knowledge

graphs. In Verborgh, R., Hose, K., Paulheim, H.,

Champin, P.-A., Maleshkova, M., Corcho, O., Ris-

toski, P., and Alam, M., editors, The Semantic Web,

pages 375–391, Cham. Springer International Publish-

ing.

Meilicke, C., Chekol, M. W., Rufﬁnelli, D., and Stucken-

schmidt, H. (2019). Anytime bottom-up rule learning

for knowledge graph completion. In IJCAI.

Paulheim, H. (2017). Knowledge graph reﬁnement: A sur-

vey of approaches and evaluation methods. Semantic

Web, 8:489–508.

Stokes, J. M., Yang, K., Swanson, K., Jin, W., Cubillos-

Ruiz, A., Donghia, N. M., MacNair, C. R., French, S.,

Carfrae, L. A., Bloom-Ackermann, Z., Tran, V. M.,

Chiappino-Pepe, A., Badran, A. H., Andrews, I. W.,

Chory, E. J., Church, G. M., Brown, E. D., Jaakkola,

T. S., Barzilay, R., and Collins, J. J. (2020). A

deep learning approach to antibiotic discovery. Cell,

180(4):688–702.e13.

Zhang, R., Hristovski, D., Schutte, D., Kastrin, A., Fiszman,

M., and Kilicoglu, H. (2021). Drug repurposing for

covid-19 via knowledge graph completion. Journal of

Biomedical Informatics, 115:103696.

Neuro-symbolic XAI for Computational Drug Repurposing

225