1 de 13

Prediction of response to melanoma immunotherapy using gut metagenomic data

Authors

M. Filippov,

T. Ziubko

​

Supervisor

Eugeniy Olekhnovich

FRCC PCM, Moscow, Russia

2 de 13

Immunotherapy has revolutionized the field of oncology. It has been demonstrated highly effective against various types of cancer, but a considerable number of patients were shown to be resistant.

​

Identification of individuals susceptible or resistant to cancer immunotherapy is a crucial point in response enhancement and patient survival.

​

The intestinal microbiota has been shown to significantly affect the immunotherapy of malignant tumours, which can be considered as a basis for the prediction of clinical response status.

​

Here we developed Machine learning-based predictive models utilizing gut metagenomic data to identify responded patients under melanoma immunotherapy. Microbial signatures as well as taxonomic assignments associated with it were also identified.

​

Introduction

3 de 13

Goal:

confirm the hypothesis of melanoma immunotherapy response status

can be predicted from gut microbiomes using machine learning models

​

​

Objectives:

  • Build a predictive model based on Random Forest classifier;
  • Validate model with cross-validation;
  • Identify taxonomical as well as functional signatures, which can be used as biomarkers to predict the clinical response status

Goals and objectives

4 de 13

Methods and approaches

Collecting raw data

MetaPhlAn (MPA)

​

KEGG Orthology (KO)

​

metagenome-assembled genomes (MAGs)

ConQuR package (v2.0)

https://wdl2459.github.io/ConQuR/ConQuR.Vignette.html

​

R-4.3.0

680 metagenomic stool samples from 7 studies published in 2017-2022

Frankel et al. 2017

Gopalakrishnan et al. 2018

Matson et al. 2018

Davar et al. 2021

Baruch et al. 2021

Spencer et al. 2021

Lee et al. 2022

Arranged in tables

with feature frequencies

Batch effect correction

RFC trained with sklearn package on python 3.9

​

Model validation on corrected and uncorrected data

Random forest classifier (RFC)

Extraction of the importances

​

Analysis of the most important features toward literature data

​

Identification of feature signatures associated with positive immunotherapy response

​

​

Annotation and analysis

5 de 13

Results: Batch effect correction

ConQuR

In attempt to get the most accurate feature importances, we corrected batch effect with ConQuR tool.

​

Using quantile regression it diminishes the dispersion explained by the batch and keeps the dispersion explained by the target variable (R/NR).

6 de 13

Results: model validation

  • The model trained on batch-uncorrected data demonstrated acceptable accuracy with souffle-split CV, but very poor results in 6-vs-1 testing.

​

  • the most important features were obtained on batch-corrected data

​

  • Batch correction procedure pre-trains the model

7 de 13

Results: importance extraction

We extracted the most important feature from each of MetaPhlAn, KEGG Orthology and MAGs datasets.

8 de 13

Results: interpretation

  • 29 of 70 most important MAGs and MetaPhlAn features (corresponding to bacterial species) are common

​

  • KO annotation cannot be linked to any cancer pathology pathways

KO annotation

29

bacterial species

​

from 17 genuses

MPH/MAGs genuses crossing tree

9 de 13

  • 29 of 70 most important MPH/MAGs common bacterial species from 17 genuses

Bacteroides

Alistipes

Blautia

Parabacteroides

Roseburia

Akkermansia

Barnesiella

Bifidobacterium

Dorea

Escherichia

Faecalibacterium

Flavonifractor

Fusicatenibacter

Gemmiger

Odoribacter

Ruminococcus

Ruthenibacterium

Results: interpretation

10 de 13

Results: interpretation

  • 29 of 70 most important MPH/MAGs common bacterial species from 17 genuses

​

  • Major representation in Alistipes, Bacteroides, Blautia, Roseburia and Parabacteroides

Bacteroides: caccae, cellulosilyticus, ovatus, stercoris, thetaiotaomicron, uniformis, xylanisolvens

7

Alistipes: finegoldii, indistinctus, putredinis, shahii

4

Blautia obeum, wexlerae

2

Parabacteroides: distasonis, merdae

2

Roseburia intestinalis, inulinivorans

2

Akkermansia muciniphila

1

Barnesiella intestinihominis

1

Bifidobacterium longum

1

Dorea longicatena

1

Escherichia coli

1

Faecalibacterium prausnitzii

1

Flavonifractor plautii

1

Fusicatenibacter saccharivorans

1

Gemmiger formicilis

1

Odoribacter splanchnicus

1

Ruminococcus bromii

1

Ruthenibacterium lactatiformans

1

11 de 13

Lachnospiraceae

Bacteroidaceae

Ruminococcaceae

Rikenellaceae

Bifidobacteriaceae

Tannerellaceae

Akkermansiaceae

Enterobacteriaceae

Odoribacteraceae

Prevotellaceae

​

​

Results: interpretation

MPH/MAGs genuses crossing tree

  • Top 70 bacteria from MPH & MAGs

belong to 39 & 37 genuses (57 unique in total) with:

​

19 common ones

(from 10 families)

12 de 13

Conclusion: machine learning approaches can reveal gut microbiome-immunotherapy interactions and predict response status in melanoma therapy with good predictive value, which is ough to improve cancer patient outcomes

Conclusion

13 de 13

​

  • Try to use another ML models and create a neural network;
  • Find the best way to correct batch effect with available tools;
  • Validate the model on additional data;
  • Identify functional features associated with clinical response status;
  • Hypothesize metabolic pathways and biological processes that might be important in immunotherapy response;
  • Publish obtained results in a peer-reviewed scientific paper!

Future goals: