1 of 15

Modeling microbiome-trait associations with taxonomy-adaptive neural networks

Yifan Jiang

University of Waterloo

2 of 15

Two missions in microbiome research

2

Predict disease status from microbiome profiles

Identify critical microbes associated with diseases

3 of 15

Challenges in modeling microbiome-disease associations

3

High feature dimensionality
Low sample size

Impede the applicability of powerful ML models

Uncertainty in read mapping

Potential

profiling inaccuracies

Curse of

dimensionality

Unbalanced

dataset size

Several challenges impede the accurate modeling of microbiome-disease associations.

First, potential profiling inaccuracies can introduce uncertainty into the read mapping process.

This uncertainty can lead to noisy features and hinder the downstream analysis of microbiome data.

Second, the curse of dimensionality is a common issue in microbiome data analysis, where the number of features far exceeds the number of samples.

This high dimensionality can make it challenging to train machine learning models effectively and can lead to overfitting.

Third, unbalanced dataset sizes can further complicate the analysis of microbiome data.

Many small datasets are available for microbiome research while large datasets are scarce.

This imbalance can make it difficult to train models that generalize well across different datasets and impedes the application of more powerful machine learning techniques, such as deep learning.

4 of 15

An ideal computational model for microbiome-disease associations

4

How can we achieve such a model?

Predictability:

accurate prediction

Interpretability: discover new biology

Trainability:

applicable to small datasets

5 of 15

Modelling dilemma: �predictability vs. trainability vs. interpretability

5

Linear model

Random forest

MIOSTONE

Multi-layer Perceptron

Predictability

Trainability

Interpretability

The modeling dilemma in microbiome-disease associations arises from the trade-offs between predictability, trainability, and interpretability.

Linear models, random forests, and multi-layer perceptrons each have strengths and weaknesses in these areas.

For example, linear models are interpretable and suitable for small datasets but may lack predictive power.

Random forests have high predictability and work well with high-dimensional low-sample-size data but may be challenging to interpret.

Multi-layer perceptrons can capture complex relationships in the data but may require large datasets to train effectively.

Their black-box nature can also hinder interpretability.

MIOSTONE, our proposed model, aims to mitigate this dilemma by achieving a balance among predictability, trainability, and interpretability.

6 of 15

We can mitigate the modeling dilemma by leveraging the inherent correlation structure among taxa

6

Train

Predict

Input

Taxonomy

Output

MIOSTONE

7 of 15

MIOSTONE addresses the challenges in modeling microbiome-disease associations

7

Data-driven taxonomic aggregation

Potential

profiling inaccuracies

Curse of

dimensionality

Unbalanced

dataset size

First, to combat potential profiling inaccuracies, MIOSTONE employs a data-driven taxonomic aggregation strategy to reduce feature noise.

This strategy uses the taxonomic hierarchy to group related microbial features and aggregate their information.

By grouping uninformative features together and preserving informative ones, the impact of profiling inaccuracies is minimized, leading to more accurate predictions.

Second, to address the curse of dimensionality, MIOSTONE utilizes a taxonomy-adaptive pruning mechanism to reduce the number of model parameters.

By pruning the fully connected layers based on the taxonomic hierarchy, MIOSTONE can effectively reduce the model's complexity and prevent overfitting.

Third, to manage unbalanced dataset sizes, MIOSTONE applies a knowledge transfer approach to leverage prior knowledge from large datasets.

The model can be pre-trained on large datasets and fine-tuned on smaller datasets, which is particularly useful for sample-limited tasks.

This is achievable because MIOSTONE's taxonomy-encoded architecture facilitates the encoding of transferable knowledge from related microbial features.

8 of 15

MIOSTONE provides accurate predictions of the disease status

8

Data:

Task:

Crohn's disease (CD) and ulcerative colitis (UC)

AUPRC / AUROC

Evaluation:

IBD (inflammatory bowel disease)

Disease status prediction

174 samples / 5287 taxa

5-fold cross-validation

AUPRC

AUROC

RF

SVM

MLP

Popphy-CNN

TaxoNN

MIOSTONE

We evaluated MIOSTONE's predictive performance on the inflammatory bowel disease (IBD) datasets, which include Crohn's disease (CD) and ulcerative colitis (UC).

The dataset consists of 174 samples, with 5287 taxa, which is a typical high-dimensional low-sample-size dataset.

Using metrics such as AUPRC (Area Under the Precision-Recall Curve) and AUROC (Area Under the Receiver Operating Characteristic Curve), we compared MIOSTONE's performance with other machine learning models, including random forests, support vector machines (SVMs), and other neural network-based approaches.

For scientific rigor, we also conducted a paired t-test to assess the statistical significance of the performance differences between MIOSTONE and other models.

The results demonstrate that MIOSTONE outperforms other models in terms of predictive performance, highlighting its effectiveness in modeling microbiome-disease associations.

9 of 15

MIOSTONE improves predictive performance in sample-limited tasks through knowledge transfer

9

Data:

Direct prediction on IBD (Zero-shot)

Task:

A larger dataset for UC/CD

AUPRC / AUROC

Evaluation:

HMP2 (Human Microbiome Project)

IBD

Fine-tuning on IBD

Training on IBD from scratch

AUPRC

AUROC

Zero-shot

Training from scratch

Fine-tuning

10 of 15

MIOSTONE identifies microbiome-disease associations with high interpretability

10

Data:

Task:

Literature evidence

Evaluation:

IBD

Biomarker discovery

The Prevotella genus (Kabeerdoss et al, Indian Journal of Medical research, 2015 )

11 of 15

Conclusion

MIOSTONE serves as an effective predictive model, as it not only accurately predicts microbiome-trait

associations across extensive real datasets but also offers interpretability for scientific discovery

11

Facilitating in silico investigations into the biological mechanisms underlying microbiome-trait associations

Impact:

Key idea:

Reducing feature noises by aggregating taxa in a data driven way
Mitigating curse of dimensionality by a taxonomy-encoded architecture
Applicable to small datasets via knowledge transfer

12 of 15

Questions?

13 of 15

13

14 of 15

14

15 of 15

15