Professor Anshul Kundaje
Machine Learning Approaches to Decode the Human Genome
The Human Genome Project
In 2003, after 10 years of work, costing about 3 billion dollars, we got a first draft of the genome
Genome = long string of A, C, T, G nucleotides
Exact sequence encodes cellular function, identity, etc.
~3 billion nucleotides
These letters represent some average kind of genome for mankind.
Population sequencing to identify disease-associated genetic variants
Fast forwarding to today, there's been a revolution in genome sequencing technologies.
Sequencing costs dropped from 3 billion dollars to close to 1000 dollars in the span of a few years.
Sequencing machines have reduced in size
Oxford Nanopore technology = fits in your palm; effectively a USB device
Live sequences and streams data directly into computer
One-time use
Useful for remote areas
Number of genomes being sequenced rising faster than Moore's Law
Advantage of cheap sequencing:
Can conduct population sequencing
With the consensus human genome, at specific positions of the genome, individuals can differ. This is genetic variance.
Genetic variation is mostly not harmful, just makes us different
Genome-wide association study: can detect statistically significant associations with disease of interest for individual genomes
Being heavily used today
Can identify loci, but not sure how to act on this information
Not sure what loci does or how loci translates into some molecule that we could perhaps counteract with a drug or whatnot
This is the area of functional genomics: take large sequence genome databases and annotate what every position in genome does/does not do
To answer basic questions in functional genomics, we need to think about the anatomy of the genome.
A sequence of a gene encodes a protein. Genes, proteins are building blocks of the cells for it to perform various activities.
By expressing different proteins at different times, can produce different cells with different function
Proteins are the workhorses
Unique expression of proteins defines cell function
Cell is a complex control system
Regulate which proteins are expressed by controlling genes (activate/repress)
Among 3 billion base pairs, ~99% base pairs are non-coding (potentially do something else); only ~1% encode proteins
Control elements (on/off) are scattered throughout the 99% - we don't know where they are or what they do.
How to decide if specific sequence in genome is control element or not?
The cell does this with epigenetic modifications; there are a number of proteins that can be attached
Depends on epigenetic proteins that interact with a loci; can perform sequencing to generate genome-wide profiles of epigenetic modifications
There is an epigenetic language/code - beyond A, C, G, T - that dynamically decides what portion of sequence is active or not
One genome sequence is identical in all cell types/tissues in body
How does one genome encode for so many cell types?
Each cellular state corresponds to a specific repertoire of proteins that are expressed expression = specific combination of epigenetic control elements expressed
If you want to decode how a genome encodes function, need to think about in terms of cell state.
NIH created the epigenetic cube; x-axis = 100s of cell types/tissues, y-axis = genomic coordinates (3 billion positions), z-axis = 100s of biochemical measurements
This is where machine learning, probabilistic models, and deep learning can help understand this massive amount of data.
3 questions we ask are:
1) Identifying tissue-specific control elements
2) Learning sequence code of control elements
3) Interpreting disease-associated genetic variation (now we see a mutation in the human genome, we can predict what would happen if you broke that sequence)
First challenge: cube has massive amount of data; need to reduce to some compressed representation
No single biomarker just encodes function.
Evolution is smart; uses combinations of biochemical markers to encode a function
Combinatorial chromatin states
In an unsupervised approach, walk across the genome and identify all combinations that are frequently found
Hidden Markov Model
Walk across genome; identify latent/hidden states
Latent state emits an observed biochemical marker
Can train giant HMM that runs across all cell types to learn hidden states
This HMM captures state transitions and the combinatorial combinations.
Can then compress all the biomarker profiles into a vector of hidden states, and represent genome and any single cell type with one track.
A comprehensive functional annotation of the human genome
~20,000 genes, we know largely where they are
We discovered ~2 million novel putative control elements: only occupy ~8% of genome
Also discovered cell-types specific usage of elements
A very complex control systems enables the complexity of a human.
This is how we first started opening up the rest of the genome and understanding the other 99%.
We still believe that most of the genome is neural and not doing anything; hasn't been weeded out through evolution.
2 million control elements show highly modular activity defining tissue identity
What's defining cell type tissue and identity is intricate activation of elements.
Beautiful mapping of function of genes and what activates them.
>95% of complex disease-associated variants are not in genes instead likely disrupting control elements
Only 4.9% of known disease-associated variants break genes (breaks the coding sequence of genes)
Remaining 95% of disease-associated variants are outside; likely break control elements
Question: What cell type is relevant to study for a specific disease?
Ex: Alzheimer's Disease, neurodegenerative disease, which cell type should you study?
Neurons perhaps? Neurons degenerate
We find actually immune cells affect Alzheimer's.
Take AD variant positions and overlap with cell type Manhattan plots
Immune/blood cell types prove most prominent
Immune cells are responsible for clearing plaques
But genetic variants that disrupt immune cells prevent removal of plaques
May cause neurons to die -- exact mechanism is not known
Data-driven approach shows that large proportion of variants that affect AD act through immune elements rather than neuronal elements
Supported by mouse model experiments
Learning the DNA regulatory code of the genome
What are the DNA words and grammars that define tissue-specific control elements?
Transcription factors love to bind to DNA sequences with specific words. The basis of the regulatory code is these little landing pads for proteins.
Can we learn from sequences to determine their activation patterns?
Supervised machine learning for decoding DNA words in tissue-specific control elements
Classify a sequence as part of a module/tissue type or not (is it active?)
Used neural networks; we don't know a good way of encoding DNA sequences (people have tried for a while hand-engineering features)
Apply an agnostic approach
Every artificial neuron in a neural network can encode some sort of DNA word and this neuron can be activated every time it sees this word. The neurons are pattern detectors.
Advantage: do not need to specify features; neurons in neural network will identify features on its own
We use deep convolutional neural networks since we want to exploit locality across sequences.
Instead of just predicting one activation, we use multi-task deepCNN to jointly predict binding of multiple transcription factors
With this approach, we have prediction accuracy of mean auROC of 0.82 and mean auPRC of 0.65.
We're getting tissue specific activations of the genome.
More interested in features extracted by the neural network rather than predictions
Learn 1000s of known and novel DNA words defining the regulatory code of tissue-specific control elements
Next, can we interpret how the same DNA sequence gives rise to dynamic activation of control elements in different cell states?
Say we have a neural network that predicts activations for various regions. We want to understand why it does that. So we can that recursively backpropagate to individual nucleotides that contribute to the activation.
DeepLIFT: Kundaje lab.
Different DNA word grammars can activate same control element in different tissues.
We can see for different cell types, which words are active and combinatorially define activation.
So like natural language, DNA has words and grammars.
How can we use this to understand disease?
Let's look at coronary heart disease.
CNNs can predict and interpret effects of disease-associated genetic variants in the relevant tissue context.
We look at unstimulated coronary smooth muscle cells. Genome-wide association study says a certain C->T variant is super important. However, the neural network indicates it doesn't matter.
When we stimulate these cells, the neural network actually finds a certain word - that has the position of the C - that is important.
So we find that stimulated cells are the ones that are important to target, not unstimulated cells.
Hence, neural networks help us understand what may be important for disease and take action appropriately.
Future of personalized medicine
Personal genome sequence + personal functional genomic data + electronic medical records/clinical data/biometrics/literature mining -> domain-specific machine learning + AI -> rapid interpretation of personal genomes + data-driven personal diagnosis (cause rather than symptoms) + drug target identification and design + optimal treatment regimes
For example, in cancer, if we can track patients over time (with functional assays), we can better determine how to treat the disease.
Within a decade, most of the human population will probably be sequenced, leading to many exciting opportunities