1. An Artificial Intelligence Approach to Precision Oncology
Alexandru Floares, MD-PhD
President of SAIA Institute & Artificial Intelligence Expert

2. Motivation: Precision Oncology Goals
We want highly accurate OMICS tests, 
for diagnosis, prognosis, & response to treatment prediction,
based on various omics - microarray, NGS - data.
Increasing measurements’ precision is a necessary but not sufficient condition for Precision Oncology (Medicine).
We also need to include ML & AI (Deep Learning) tools in our bioinformatics workflows. 

3. Bioinformatics Workflow Steps
Usually: quality control, preprocessing, batch effects removal, clustering, and differentially expressed genes (DEG). 
For most studies, DEG is the end result! Why?
Easy to obtain, but with low clinical impact.
For high translational impact we need Accurate, Robust (generalizing well) and Transparent predictive models. 
Machine Learning and Artificial Intelligence can be used to develop predictive models satisfying ART Criteria. 

4. ML Predictive Models
The highest performance needs advanced ML methods:
Parameter tuning: searching for the best parameters of the ML algorithms.
Ensemble Methods: combine multiple models which are voting for the final prediction.
Bad news: Needs programming and ML knowledge.
Good news: We automated the whole workflow. With a few clicks, one can obtain the highest performance (usually, > 95%),
Programming and ML knowledge is not needed.

5. AI vs. ML Predictive Models I
In our ML workflow, data preprocessing is needed but is automated and combine multiple methods.
The preprocessing methods depend on technology (e.g., array, NGS) and platform (e.g., Illumina, Agilent). However,
Our AI workflow (OmicsBrain™), does NOT needed preprocessing.
AI proved capable to learn omics data preprocessing.
This sounds crazy! but in Computer Vision it's known that AI could learn to preprocess images (much more sophisticated).
AI recently outperformed humans in image classification.  

6. AI vs. ML Predictive Models II
Both ML and AI approaches reached % accuracy. 
However, AI OmicsBrain™, learned to integrate microarray & NGS data on different platforms. 
Thus, it is independent on both technology and platform.
This allows us to mix various datasets in a large database for training.
OmicsBrain™ can use this knowledge to easily learn accurate & robust models from just a few (≤ 10) new unseen NGS or array cases.
Usually it's hard if not impossible to obtain the best model from such small studies!

7. OmicsBrain Learned Data Integr. & Preprocess
OmicsBrain Learned Data Integr. & Preprocess.: microarray circulating & NGS tissue miRNA
PCA Raw Data
PCA OmicsBrain Preprocessing

8. OmicsBrain learned to integrate microarray circulating miRNA with NGS tissue miRNA data (TCGA data).
The raw data Dx are initially mixed and not linearly separable.
OmicsBrain learned to make them linearly separable. 
Using both tissue and circulating miRNA, not from the same patients, we 'forced' AI to mirror better the cellular situation from liquid biopsy
It can used what it learned to model or predict both liquid biopsy data and cellular data.

9. OmicsBrain™ Transfer Learning:
OmicsBrain™ has the capability to transfer the knowledge learned from one problem to a similar but different one:
Eg., it learned to discriminate between various cancers and normal.
Than, it used this knowledge to easily learn the discrimination between other unseen cancers and normal.
Moreover, it only needs a few cases (e.g., ≤ 10) to reach %!
It generalizes well, because the main knowledge was extracted from a much larger dataset (> 5000).
No other ML approach can do this.

10. OmicsBrain™ Multi-Omics I
A key aspect of our AI approach is omics data encoding.
Ordinary ML accepts only vectorial data as input, where each sample is a vector of genes, either a row or a column.
We generalized to tensorial encoding: a vector is a 1-D tensor, a matrix is a 2-D tensor and so on.
This allows us not only to include multi-omics data but also various types of domain knowledge, e.g.,
Expression, methylation, mutations can be taken as inputs, together with GO, Pathways, etc.

11. OmicsBrain™ Multi-Omics II
Properly designed AI systems ca take as learning input such tensorial data structures.
While multiple omics data are thus mixed we can also ask questions like these:
For a highly relevant gene, which are the most important aspects, 
Expression, methylation or mutation? And/or
It's pathway, or GO terms, etc.

12. Transparency & Feature Selection
We introduced and followed the ART Criteria: Accuracy, Robustness and Transparency.
2 drawbacks of Deep Learning Neural Networks are:
No feature selection, e.g., all genes are used for prediction.
Black-Box, non-transparent, not-interpretable models.
OmicsBrain™ combines AI with ML for both feature selection and interpretable models developing. 
The final models are developed by NN but interpreted either as rules or decision trees.

13. Functional Redundancy
We believe that functional redundancy is a fundamental property of living systems, related to their robustness.
Thus, instead of trying to find, from the beginning, the minimal subset of relevant genes,
We are first finding all relevant genes, meaning that some of them are somehow equivalent.
This open the door for choosing multiple subsets from all relevant genes in a biomedically meaningful way:
E.g., actionable genes, genes from a certain pathway...

14. Interpretability and Individualization
Combining decision trees with AI, we are obtaining general, population rules.
Using other techniques, we can obtain individualized rules, allowing to ask questions like:
Why this patient was diagnosed as cancer or as normal?
Why this patient will progress or not?
Why this patient is responding or not responding to a certain drug?

15. Conclusion OmicsBrain™ is and end-to-end solution capable of:
Automatically preprocess your microarray, or NGS data from any platform (Illumina, Agilent, etc.),
Developing highly accurate, robust and interpretable molecular tests, at the population and individual level,
Using just a couple of cases, and without requiring programming or AI knowledge 
OmicsBrain™ will be available soon at

16. Thank You!