SLIDE 1
New developments for geno2pheno
- Prof. Nico Pfeifer
Uni Tübingen
Excellence cluster: “Machine Learning: New Perspectives for Science” (Tübingen)
3
New developments for geno2pheno Prof. Nico Pfeifer Uni Tbingen - - PowerPoint PPT Presentation
New developments for geno2pheno Prof. Nico Pfeifer Uni Tbingen - - PowerPoint PPT Presentation
New developments for geno2pheno Prof. Nico Pfeifer Uni Tbingen Excellence cluster: Machine Learning: New Perspectives for Science (Tbingen) 3 Biomedical Data Science Biomedical Data Science Coding skills are necessary for working
SLIDE 2
SLIDE 3
Biomedical Data Science
SLIDE 4
Biomedical Data Science
5
Coding Skills
Coding skills are necessary for working with massive amounts of electronic data that must be acquired, cleaned, and manipulated.
SLIDE 5
As transformative as electricity was?
SLIDE 6
Nature (2017) doi:10.1038/nature21056
~130,000 labeled images to train the model
SLIDE 7
Machine Learning
- What can Machine Learning methods
solve? –“Everything “
SLIDE 8
Machine Learning
- What can Machine Learning methods
solve? –Andrew Ng: “Everything a human can do in a second” (low hanging fruits)
- Automation of certain processes
– Classification: Spam / no spam – Scene understanding: object recognition in pictures/videos – Text translation: English – German – Predict wait times: Pizza delivery
SLIDE 9
Machine Learning in Medicine
- What could Machine Learning methods
solve (harder cases)? –Predict therapy success
- Complex data, but simple phenotype (e.g.,
therapy success / failure)
- Find predictors in complex data to predict
success or failure early during course of the disease
- Provide interpretable results for treatment
decision support
SLIDE 10
Treatment Decision Support System (TDSS) for HIV- infected patients
- T. Lengauer, N. Pfeifer, R. Kaiser, Personalized HIV therapy to
control drug resistance. Drug Discov. Today Technol. 11, 57–64 (2014).
- Prime example of
personalized medicine:
- Many different drugs
available
- Drugs are given in
combination (>2)
- Drug resistance mainly
depends on genotype of virus (small genome)
- Large data collections have
been established to enable training the machine learning algorithms (AREVIR, EuResist, ...)
SLIDE 11
If we have NGS data
12
https://ngs.geno2pheno.org/
SLIDE 12
Treatment decision support systems for NGS data from HIV/HCV
13
Döring M., …, Rolf Kaise, Thomas Lengauer, and Pfeifer N. Nucleic Acids Research 2018, gky349
SLIDE 13
Treatment decision support systems for NGS data
14
https://ngs.geno2pheno.org/
SLIDE 14
Server for HIV resistance prediction against bNAbs
15
geno2pheno[bNAbs]
SLIDE 15
Based on
16 http://journals.plos.org/ploscompbiol/issue?id=10.1371/issue.pcbi.v13.i11#Cover
- A. Hake and N. Pfeifer,
PLOS Computational Biology 2017, 13(10):e1005789
SLIDE 16
Estimated tuberculosis cases and deaths, 2017: Estimated mortality of TB cases (all forms, excluding HIV) per 100,000 population
17
SLIDE 17
Notified MDR/RR-TB Cases absolute numbers
18
SLIDE 18
- Project runs for 2 years
- brings together 12 partners from 10 countries
that are led by EuResist Network (Italy).
- Centre of Excellence for Health, Immunity and
Infections (CHIP, Denmark) is responsible for scientific coordination of the project.
- CARE operates under a specific Horizon 2020
call aiming to foster research between the EU and the Russian Federation.
19
SLIDE 19
20
Kick off meeting in Rome 1/2019
SLIDE 20
Kick off meeting in Rome 1/2019
21
SLIDE 21
Geno2pheno[TB]
- Start with existing data / existing rules
– Collaboration with Borstel (Leibniz Center for Medicine and Biosciences)
- Build customized version for Eastern
Europe
– Transfer Learning with general model and newly generated data from St. Petersburg and Chisinau
22
SLIDE 22
Future of geno2pheno
- [ngs-freq]: Predict treatment success from NGS data
directly
– Use linkage information between positions for more accurate models – Get rid of “expert” cut-offs – Differentiate functional from non-functional viral variants
- [ngs-freq]: Include host information
– Some HLA footprints overlap with positions predictive for treatment success
- [bNAbs]: Optimize for clinical use case
- Prediction methods for other diseases: TB, HCV, …
26
SLIDE 23
Treatment decision support systems
- More and more data can be measured
(e.g., different omics layers)
- Data is available for more and more
diseases (see medical informatics initiative)
- We need efficient machine learning / data
mining methods to learn association between treatment and patient outcome
- Results need to be communicated in a
transparent and interpretable form
SLIDE 24
Thanks to everybody, especially:
Thomas Lengauer Rolf Kaiser Francesca Incardona Florian Klein Till Schoofs Jan Heyckendorf Matthias Döring Anna Hake Joachim Büch Georg Friedrich Marius Herr