在手机上保存，获得更好体验
本字幕由TME AI技术生成
Welcome to the deep dive
Have you ever
Um
Have you ever tried to fix a greiny photo on your phone
Using one of those really aggressive smoothing filters
Oh Yeah
Where you just crank the slitter ter away
Exactly you crank it up
And sure the grain disappears
Yeah
But then you zoom in
And you realize you accidentally blurred out like the texture of the person's clothing or their eye
Ashes you basically lose the actual physical reality of the picture
Right It's
It's a huge trade off you get a smooth image
But it's fake
And um it becomes a massive crisis
When you realize the scientific community is doing the exact same thing to our genetic data
Which brings us to our source for today
We are the packing a dense
But honestly ground breaking twenty twenty five amics to single le methods
It's by goring charcharc really the stollar and pacter
Or it's titled monod model based discovery an integration through fitting sarcastic transcriptional dynamics to single cell sequencing data
是的是的
Okay
Let's come back to this
It's a mouthful of a title for sure
Yeah It is
But our mission today is to decode how this new python package which is called monod is basically forcing biologists to completely rethink the concept of noise
Instead of filtering out messy data
This approach uses the laws of physics to find hidden signals
Signals that are actually life saving
Especially
When we talk about cancer and cellular development for years
Biology is treated the variability and single cell data as a nuisance
Something to just to know scscrap away with math
对对对
Monod has no the noise is the signal to really get why this is such a big deal
We need to look at this datus quote
So right now scientists use single cell rna sequencing
Scrna sec
Right it's amazing technology
We can measure tens of thousands of genes
Across millions of individual cells all at once
But the data you get back from that is what it's massive
And it's incredibly complex
Yeah
And it's absolutely filled with noise exactly so a handle all that chaos
Scientists use these huristic machine learning algorithms
We're talking about standard tools like proportional fitting log normalization
Pcauma things anyone in data science would recognize absolutely
And the mamagooals is to remove technical noise and reduce the dimensionality of the data
So it's easy to cluster
Easy to look at on a graph
But here's the this sort of huh noment from the paper
The standard techniques act exactly like that aggressive beaubeaufilter
We just talk k they do the authors argue these algorithms cause seseere al noortion or even a total loss of biological signals
And they completely lack interpretability
Because they hyperfocus on the average expression of a gene
They smooth right over the actual physical reality of what the cell is doing
Think about this in your own life
It you're looking at a data set
It work right
But often do you look at the average
And totally ignore the outliers
It's supercommon
It's easy
It is easy um
But in biology ignoring those outliers means you are mising those analydoput a cell is not an average
No
A cell is a physical system
And that is where monot comes in monon takes a mechanistic biophysical approach
It doesn't treat genomomic data like an abstract math puzzle
It treats it as a physical manufacturing process
I really like the factory analogy for this
Let's say standard analysis is like standing outside a warehouse and a counting the finnistic products that get shipped out every day
Right the mature rna
The spanstarn a standard tools just count the final output
Yeah
But monon looks at the raw materials
Arriving at the loading outcess
But lelike like the glogy is raane
It looks at the finish products
It maps out the entire supply chain
Because reading the genome isn't just an instance software command
It's a physical process and a crucial concept from the paper
Here is bursting
Which is trtrscription isn't a steady stream
Exactly if a gene just turned on and produce a steady slow stream of rna
The data would look very predictable
But to regenomomic data is highly variable
The factory isn't running a slow steady convariable
It worworking is aotic frecy spts ts rn is produced in bursts
Quote opops soom of birth of rna in it closest
Yes
So monod physically measures two things birth size
Which is how much rna is made in one go and birth frequency
Which is how often those predicable
But to this
This accurate
It has to separate true biological corts from just random sequenening errors
Right it separate biological noise from technical noise for example
Monod accounts for something called linth biased capture links
Bias capture meaning
Meaning the sequcing machine itself is naturally just more likely to accidentally capture and copy a physically longer gene multiple times
It's a hardware bias
It has nothing to do with the cell monod mathematically filters out that hardware bias
So we can see the true biological bursting
What's fascinating here
That monnot allows us to test different bibiphysical models
It birsty transcription versus stedy a constitutive transcription to see which physical reality actually matches the data
We are looking at so before we gget into what this actually discovers in the lab
Let's do a constiker recap
We've estabshed the current tools act like a beauty filter
You're raising real physical data to give us clean averages
Monod takes away
The filter by tracking the raw and finish rna measures the actual physical burst of soular activity
It respects the noise exactly so is moving yyond the verages
What does this all mean
How does this physically change what scientists see in the lab
It changes everything for something called de data machanistic differential expression
Usually
Scientists just look for differentially expressed genes
Genes where the total every volume just goes
Uer er right did the factory output more boxes
Today that standard de but monod assie f ata
It looks for genes
Where the average stays the exact same
But the noise
The underlying burst size were the frequency radically changes
Let's look at their first case study the id u stress test the researchers introduced a chemical called idu
Which caused actual dna damage to the cells
Right
It's a setual stresser
Now when they looked at the average genexpression
It didn't really changed much
The cells look fine on paper
Exactly
But maana detected massive noise amplfication specifically in stress response genes like zfp one
Ten and yp one and those regulate apop toses es right sell that yes for those specific survival genes
The birth size massively increased
Which caused frequency dropped
So the cell is basically panicking king bursts
It's fele mmassive amounts of survival rna all at one and then going silent
But because the massive spikes and the deep olles mathematically average out to a normal baseline
The standard algorithms totally misthe fact that the cell was dying
They were completely blind to it
That is wild
And we see another example of this in the mouse brain data
They compared glue ommattergic neurons with the abbit ergic neurons too totally different types of brain cells right enthrgenes like nin and bock too
The average mature accounts were basically identical in both cell types
So monod revealed they have completely divergent noise profiles to maintain that same average one type of neuron has to use a completely different acsual bursting strategy than the other
It really highlight how averaging just destroyed the mechanism and mona goes beyond just bursting s
Let's talk about the germ cell development case study
Uh yes
So monaud can also calcuically the exact physical spespeof of r ana bleicing not just guessing it from a ratio
But calculate how actual connetics right because historically scientists just looacat the ratio of nasic to mature neurons
And guess how fast the splicing ing machinery was running
Which is a huge assumption
But monod mathematically derives the real physical splacing rate
And in this germ cell study looking at mouse embryos from day twelve today thirteen just really cancal window for development right monod found entirely new candidate genes
Genes like new step one and top two
I it prove that these genes physically speed up in spspicing rates
During that specific twenty for our window
The researchers could literally see the gears of the cell shifting into the higher speed
You cannot get that from a smooth data average
Okay
Let's do another quick summary
Because headadinto into highest stakes part of this paper
Monod's dedata finds hidnegativity
We saw sales panicing from dna damage
And we saw rna splicing physicicspespeeup up eeryyos all hidden behind completely normal looking average
Which it ings us to cancaner er resistance key study four pancreatic docktal add no cars monoma kdac
These are tumor cells that were treated with chemotherery and radiation
And they survived and obviously figuring out
How they survive is the holy grail of cancer research
Standanalalyis just gave a list of genes
That had a higher mean expression
Which tells you what is happening
But not how monod dug into the actual connecs
It looked at rb one
Which is a crucial tumor suppressor gene
It stop cells from dividing out of control and in these resistant tumors
Arb one was down regulated
But monod showed exactly how it was down regulated
Specifically through addrastic decrease in birth frequency
The tumor isn't dederoying in the machery y
It's just refusing to trigger the bursts and in their cdc forty two
Which actually help cells divide in resist chemo right for cdc forty two monod found a massive increase in birth frequency
So the tumor has a highly specific mechanical strategy
It loers the frequency of the tumor suppressand it crranks up up the frequency
The chemo resistance
嗯
If you are developing a drug
Knowing that you need to target the birth frequency mechanism rather than just total volume that completely changes your approach
It is the difference between knowing a machine is broken and knowing exactly which gear inside a slipping
Here's where it gets really interesting
Because the authors don't just say hey
Monnot is is
Because a prety the papapwith a pretty devastating mathematical proof against the standard algorithms
It is a brutal al titice of proportional fitting log one
Ppca and ump right
So let's break down
Why it goes back to the physics
Naand rna physically turns into mature orin a
It's a direct molecular precurser
Therefore there is a guaranteed causal mathematical correlation between the raw material and the finish product
They are physically linked
But when the authors ran simulated data
Where they knew the exact underlying physics through the standard preycessing algorithms
Because algorithms mathematically destroyed the biological correlation
Because things like the log gp trending information or dividing a genes count by the total rna of the cell
It completely sevvers the joint distribution
Exactly the standard dennoing tools litererally erised the evidence that the nacent and mature rna are related priprioritize a clean abstract cluster on a graphh over the physical laws of biology
So the very tools everyone uses to clean their data are actively are racing the mechanisms of disease
Let's bring all this home by respecting the noise
By building models based on actual physics like burst and splicing and degredation
Monod
Let's scientists see things
They've been blind to hitting cancer strategies
Hidden brain cell differences
And it proves that our favordata cleaning algorithms are actually destroying biological troops
So for you the the stener
The builbig take away
Here is about how we seek twledge right
Knowledge isn't about perfectly clean data
It is about understanding the underlying mechanisms that generate that data noise
And all if you aggressively filter out variance just to make a tart look look pretty you blind yourself to relaality
Which lease us with a final thought for you to moll over
If the our standard widely accepted scientific algorithms are actively are racing physical reality in genomics
What other fields of science or business or even your everyday life are losing the actual signal
Because we are just too obsessed with the clean data set what noise in your own life is actually the main signal
You should be listening to it's a fascinating question
Because we 're just joining on this deep dive
Keep questioning the averages
And we will catch you next time
