1. Predictive and interpretable Bayesian machine learning models for understanding microbiome dynamics
Georg K. Gerber, MD, PhD, MPH
Assistant Professor of Pathology, Harvard Medical School
Member of the Harvard-MIT Health Sciences & Technology Faculty
Co-Director, Massachusetts Host-Microbiome Center
Chief, Computational Pathology
Brigham and Women’s Hospital
JSM 2018, July 29, 2018

2. Disclosure of Financial Relationships
Kaleido Biosciences, Inc.
Strategic Advisory Board (stock options)
Sponsored Research Agreement (no sponsored research discussed in this talk)
ConsortiaRx, Inc.
Scientific co-founder and Scientific Advisory Board (stock)
Patent filed for food allergy microbiome therapeutic
Provisional patent for C. difficile bacteriotherapy and diagnostic
None of the work presented here was supported by commercial organizations
Funding: DARPA HR C-0094, BWH Precision Medicine Initiative, NIH/NHLBI T32 5T32HL007627

3. Microbiome therapeutics and diagnostics landscape
Exponential growth in R&D into clinical applications of the microbiome
bacteriotherapies (“bugs as drugs”)
prebiotics - dietary compounds to reshape microbiota
molecular diagnostics
“Pharmacology meets ecology”
Opportunity to rationally design microbiome therapeutics and predict their effects

4. Microbial Dynamical Systems INference Engine (MDSINE)
Bucci et al., Genome Biology :121

5. Generalized Lotka-Volterra equations (gLV)

6. MDSINE used novel computational methods to overcome inference challenges
New Bayesian adaptive penalized spline method for accurately estimating trajectories and gradients from noisy microbiome time-series data
Efficient Bayesian methods for accurately inferring dynamical systems (gLV) parameters using estimates from step 1
Prediction with estimates of confidence (Bayesian)

7. Some limitations of MDSINE
Needs to learn O(N2) microbe-microbe interaction parameters – not scalable to human data
Doesn’t propagate uncertainty completely through model – multi-step procedure that discards uncertainty from previous step
Doesn’t assume a constraint on overall carrying capacity at the dynamical systems inference stage

8. “MDSINE 2.0” Gibson and Gerber, ICML 2018.
Interaction modules: automatically discovers groups of microbes with redundant interaction structure; interaction parameters now scale as O((log N)2)
Stochastic dynamics
Fully propagates uncertainty through model
Handles “proportionality” of microbiome data in physically realistic manner
New dynamic auxiliary variable technique enables efficient inference with physically realistic dynamics

9. Stochastic dynamics

10. Complete model

11. Auxiliary variable

12. Synthetic data experiment

13. Synthetic data experiment (cont.)

14. Gnotobiotic mouse model of C. difficile infection
5 mice / 26 serial stool samples
GnotoComplex v1.0 = 13 human commensal bacterial selected for phylogenetic diversity/functional capabilities
16S rRNA on MiSeq for relative abundances of strains
16S rRNA qPCR (universal primers) to quant bacterial biomass

15. Gnotobiotic mice

16. Original MDSINE C. difficile model interaction network
Edge widths = Bayes factors (evidence favoring edge)

17. “MDSINE 2.0” C. difficile interaction network

18. Conclusions MDSINE (Genome Bio. 2016) “MDSINE 2.0” (ICML 2018)
Efficient method for inferring dynamical systems models from microbiome time-series data
Multi-step Bayesian method (trajectory learning -> gradient matching)
“MDSINE 2.0” (ICML 2018)
Fully integrated Bayesian model, fully propagates error
Learns interaction modules that compress the dynamical system model (enables scalability)
Stochastic dynamics
Handles “proportionality” of microbiome data in physically realistic manner
New dynamic auxiliary variable technique enables efficient inference with physically realistic dynamics
Ongoing work
Nonparametric models of dynamics
Interpretable supervised learning architectures
Incorporation of prior biological knowledge

19. Ongoing projects

20. Advance the field from descriptive associations to mechanistic effects of microbial communities on the host.
Director: Lynn Bry Co-Director: Georg Gerber
Funded by Massachusetts Life Sciences Center + NIH + Brigham and Women’s Hospital
Open to all (industry and academic): fee-for-service, research collaborations
Capabilities
Largest gnotobiotic facility in New England
Traditional isolators + high-throughput caging systems
Microbiology
Advanced anaerobic culture
Continuous chemostat system
Computational
Standard and customized analysis (specialty = longitudinal studies)
‘Omics
16S rRNA seq
Whole genome sequencing
Short-chain fatty acids and other metabolites
Micron-scale imaging (with the Forsyth)
CLIA lab for human microbiome trials

21. Why use dynamical systems models?
Interpretable models (growth rates, interaction strengths, etc.)
Forecasting:
dynamics with perturbations
dynamics with subsets of organisms present
Formal dynamical systems properties: stability, resilience
Directed interactions (mathematical causality)
But, challenges to inference: traditional methods use numerical integration and nonlinear optimization
slow, not scalable
poor results on noisy/limited temporal resolution data
non-probabilistic
too many potential interactions relative to data (p > n)