​Multi-targeted drug DREAM Challenge: A "crowdsourcing" drug screening program for RET-driven tumors and Tau protein neurodegenerative diseases

On September 14, 2021, the research group of Jiang Hualiang and Zheng Mingyue from the Institute of Immunochemistry of Shanghai University of Science and Technology/Shanghai Institute of Materia Medica of the Chinese Academy of Sciences published the title Crowdsourced identification of multi-target kinase inhibitors for RET- as co-authors in PLOS Computational Biology.
and TAU-based disease: The Multi-Targeting Drug DREAM Challenge article [1]
.

This work reports the results of the multi-target prediction DREAM challenge at the end of 2018.
The first author of the article is Xiong Zhaoping, who won the first place in the challenge.
Challenge the champion of the competition), he was still studying for a doctorate at ShanghaiTech University (supervisors are Academician Jiang Hualiang and researcher Zheng Mingyue), and now Huawei is engaged in the development of AI algorithms related to drug research and development
.

At present, drug development is usually aimed at regulating a single cell pathway or target, but many marketed drugs often regulate multiple cell pathways or targets (http://ruben.
ucsd.
edu/dnet/).
This phenomenon is called "Off-target" effects or multiple pharmacology
.

Although many researchers hope to reduce the impact of secondary targets and devote themselves to the development of highly specific drugs, many studies have shown that multi-target therapy has significant advantages for a variety of complex diseases [2]
.

If drugs have precise and controllable multi-target regulatory effects, they may be more effective in the treatment of diseases, but this also puts forward higher requirements and challenges for drug molecular design
.

The organizers of the multi-target prediction DREAM Challenge hope to use crowdsourcing strategies to design a multi-target drug screening algorithm
.

In the previous research, Saifang used the Dominant genetic modifier screen of Drosophila to identify the key regulatory nodes of RET-driven cancer and Tau protein neurodegenerative diseases (Tauopathy), respectively.
It contains some targets that need to be inhibited to reverse the disease process (Pro-Targets) and targets that cannot be affected (Anti-Targets) [3,4]
.

For example, for RET-driven tumor models, Saifang proved in previous experiments that if the Pro targets of RET, BRAF, SRC and S6K are inhibited, and at the same time the Anti targets of MKNK1, TTK, ERK8, PDK and PAK3 are normal.
In the case of work, the disease process can be reversed
.

Figure 1 DREAM Multi-target Prediction Challenge process.
Based on this, the competition requires participants to develop algorithms for these two disease models and look for inhibitors, hoping to find inhibitors that are active against specific protein targets, but not against other specific targets Active, which requires the compound to have highly accurate multi-target selectivity
.

In order to objectively verify the algorithm, the contestant requires that the compound predicted by the algorithm is a commercially available compound in the ZINC15 database.
The contestant will then base on the innovation and scalability of the algorithm submitted by the contestant, as well as the novelty of the compound structure.
To selectively purchase compound samples for experimental activity testing, and finally score the contestants based on the biological activity test results (Figure 1)
.

Figure 2 RET-driven tumor disease model suppression results
.

(A) The inhibitory effect of the compound submitted by the contestant on the (Pro-) target and the anti-(Anti-) target, respectively.
The smaller the binding constant (Kd) value, the darker the yellow and the better the inhibitory effect
.

(B) Similarity assessment of compounds
.

For the RET-driven tumor disease model, the competition selected 9 commercially available compounds provided by 10 participating teams according to the method description.
The smaller the Kd value, the deeper the yellow and the better the inhibitory effect
.

It can be seen from Figure 2 that the compounds provided by the top two contestants have a good inhibitory effect on Pro-Target, but have poor inhibitory activity on Anti-Target
.

In terms of the structural similarity of the compounds, the compounds selected by the contestants have relatively low structural similarity
.

Figure 3 Inhibition results of Tau protein neurodegenerative disease model
.

(A) The inhibitory effect of the compound submitted by the contestant on the (Pro-) target and the anti-(Anti-) target, the smaller the Kd value, the deeper the yellow, the better the inhibitory effect
.

(B) Similarity assessment of compounds
.

For the Tau protein neurodegenerative disease model, the competition selected 8 commercially available compounds provided by 8 participating teams according to the method description
.

It can be seen from Figure 3 that the compounds provided by the top two contestants have high inhibitory activity on both Pro-Target and Anti-Target, with poor selectivity, and relatively poorly satisfying the requirements for the disease model.

.

Table 1 The scoring list article in the main body of the game also specifically introduces the methods used by the top two teams (Table 1), and the methods of other teams are described in the Supplementary
.

Among them are academic institutions from the University of Michigan, University of Pittsburgh, Korea University, Shanghai University of Science and Technology, and Shanghai Institute of Materia Medica, Chinese Academy of Sciences
.

In addition, it also includes a number of biotech companies and pharmaceutical companies that do not want to disclose specific information
.

Figure 4 The framework diagram of the FEMTD (Zhaoping Xiong's method) model with the highest total score
.

The team with the highest overall score (Zhaoping Xiong) comes from Shanghai University of Science and Technology and Shanghai Institute of Pharmaceutical Sciences, Chinese Academy of Sciences.
The model framework adopted is shown in Figure 4
.

FEMTD (Fused Embedding for Multi-Targeting Drug, Zhaoping Xiong's method) uses graph neural networks to characterize molecules, and doc2vec model for pre-training characterization of protein sequences
.

After that, the two representations are fused together by GRU-gated recurrent neural unit to obtain a fusion representation
.

Fusion characterization divides the inhibitory activity of small molecules on target proteins into three categories: ineffective, weak and effective
.

This model has the highest score in the two tasks of RET-driven tumor model and Tau protein neurodegenerative disease model
.

Figure 5 The framework diagram of the DMIS-MTD model ranked second in total score
.

Although the best solution to the two tasks in this crowdsourcing algorithm challenge appeared in a single model, other teams also put forward very clever research ideas
.

The DMIS-MTD model proposed by the Korean University team, which ranked second, uses transcriptomics data and a Siamese neural network model to perform drug screening by predicting the similarity of the transcription profiles of two compounds
.

It first uses the kinase KIEO database to determine several prototype drugs whose kinase inhibition profile meets most of the (Pro-) target and anti-(Anti-) target conditions required by the competition, and then uses the ReSimNet trained on the CMap dataset to predict the prototype The transcription profile similarity between the drug and the compound from ZINC15 resulted in compounds with higher similarity.
Finally, the top-ranked compounds were further aggregated and filtered through Lipinski rule, patent search exclusion, and molecular docking
.

At present, crowdsourcing and pooling wisdom through the organization of algorithm challenges to overcome some problems in the field is attracting more and more attention
.

For drug research, some artificial intelligence algorithm competitions have also been successfully held in China
.

For example, the theme of the maker track of the recently concluded "Huawei Cloud Cup" 2021 Artificial Intelligence Application Innovation Competition is "Compound-Protein Interaction Prediction", which aims to screen active compounds based on compound SMILES codes and protein amino acid sequence information
.

This competition attracted more than 350 teams composed of more than 2500 people from various professional fields (https://competition.
huaweicloud.
com/information/1000041388/introduction), and many of the teams that won the final It is a well-known IT company in the world and the figure of scientific and technological personnel in the Internet field
.

Academician Hualiang Jiang commented that this competition "is a useful attempt to combine the fields of IT and BT
.

Through this competition, I am more confident about the prospects of AI technology in China's life sciences basic research and drug development"
.

Original link: https://doi.
org/10.
1371/journal.
pcbi.
1009302 Platemaker: 11 References 1.
Xiong Z, Jeon M, Allaway RJ, Kang J, Park D, Lee J, et al.
(2021 ) Crowdsourced identification of multi-target kinase inhibitors for RET- and TAU- based disease: The Multi-Targeting Drug DREAM Challenge.
PLoS Comput Biol 17(9): e1009302.
2.
Davis MI, Hunt JP, Herrgard S, Ciceri P, Wodicka LM, Pallares G, et al.
Comprehensive analysis of kinase inhibitor selectivity.
Nat Biotechnol.
2011;29: 1046–1051.
pmid:22037378.
3.
Read RD, Goodfellow PJ, Mardis ER, Novak N, Armstrong JR, Cagan RL.
A Drosophila model of multiple endocrine neoplasia type 2.
Genetics.
2005;171: 1057–1081.
pmid:15965261.
4.
Sonoshita M, Scopton AP, Ung PMU, Murray MA, Silber L, Maldonado AY, et al.
A whole-animal platform to advance a clinical kinase inhibitor into new disease space.
Nat Chem Biol.
2018.
pmid:29355849.
Reprinting instructions [Non-original articles] The copyright of this article belongs to the author of the article.
Personal forwarding and sharing are welcome.
Reprinting is prohibited without permission.
The author has all legal rights, and offenders must be investigated. .