FUNCTIONALLY PROFILING METAGENOMES AND... - Eric Franzosa - Late-Breaking Research - ISMB 2016

Summary

In the script, Dr. Orlando discusses advancements in profiling microbial communities from sequencing data, focusing on taxonomic and functional profiling. The tool 'MetaPhlan 2' is highlighted for taxonomic profiling, while 'HUMAnN2' is introduced for functional profiling, using a tiered read mapping strategy to improve accuracy and speed. HUMAnN2 enables species-level functional profiling, offering insights into both the composition and activity of microbial communities, with applications demonstrated in the Human Microbiome Project and other cohorts.

Takeaways

• 🌟 The script discusses advancements in profiling microbial communities from shotgun sequencing data, focusing on taxonomic and functional profiling.
• 🔍 Taxonomic profiling identifies species and higher-level clades present in a microbial community, while functional profiling identifies gene families and pathway composition.
• 🚀 MetaPhlan 2 is a tool developed for taxonomic profiling that searches reads against a pre-selected set of marker genes unique to each clade or species for efficiency and accuracy.
• 🛠 Humann2 introduces a 'tiered read mapping' strategy to improve the accuracy, speed, and resolution of functional profiling by leveraging species identification for more targeted searches.
• 🔑 The first tier of Humann2's method rapidly identifies species in a community by mapping reads against species-specific marker genes.
• 🌐 The second tier builds a custom database for the sample by concatenating the pan-genomes of detected species, allowing for a detailed search of the remaining reads against these genomes.
• 📚 Humann2's tiered search is more specific and sensitive, leading to higher accuracy and faster processing times compared to traditional comprehensive searches.
• 📈 The method was tested using synthetic metagenomes created from sets of bacterial genomes, demonstrating high sensitivity and precision even for low-abundance species.
• 🌿 Humann2 has been applied to real-world data, profiling hundreds of human metagenomes, and has shown to explain a majority of reads during the initial accelerated search tier.
• 🔬 The tool allows for the identification of metabolic pathways that are signatures for particular body areas, providing insights into the taxonomic resolution of functions within communities.
• 📊 Humann2 also enables the distinction between functional potential (DNA level) and functional activity (RNA level), showing that the two are not always the same.

Q & A

What are the two primary questions in microbial community profiling from sequencing data?

The two primary questions are 'who is there', which refers to taxonomic profiling to identify species and higher-level clades, and 'what are those species doing', which is functional profiling to identify gene families and pathway composition.

What is the computational challenge in profiling microbial communities from sequencing data?

The computational challenge lies in the need to search short reads from shotgun sequencing against a vast database of microbial reference genomes, which is computationally intensive and prone to errors due to potential spurious mappings.

What is MetaPhlAn 2 and how does it address the taxonomic profiling challenge?

MetaPhlAn 2 is a tool developed for taxonomic profiling that searches reads against a pre-selected set of marker genes unique to each clade or species, rather than the entire database, making the profiling process more efficient and accurate.

What is the concept behind the tool HUMAn2 and its tiered read mapping strategy?

HUMAn2 implements a tiered read mapping strategy, which starts with rapidly identifying species in a community by mapping reads against species-specific marker genes, followed by a custom database search of pan genomes of detected species, and finally a comprehensive search for any remaining reads.

How does HUMAn2 improve the accuracy and speed of functional profiling?

HUMAn2 improves accuracy by ensuring reads are mapped to the correct species-specific genes, reducing spurious mappings. It improves speed by using a reduced database for the initial search, thus processing the metagenome much faster than a comprehensive search against an exhaustive database.

What is the advantage of HUMAn2's tiered search over traditional comprehensive searches?

HUMAn2's tiered search is more specific and sensitive, placing reads in the correct locations, and thus increasing overall accuracy. It also processes the metagenome faster due to the use of a reduced pan genome database.

How does HUMAn2 provide species-level resolution in functional profiling?

HUMAn2's tiered search allows for the reconstruction of functions on a species-by-species basis within the community, providing insights into which species are performing specific functions, even for low-abundance species.

What synthetic metagenomes or meta transcriptomes are used to evaluate HUMAn2's performance?

Synthetic metagenomes or meta transcriptomes are created by taking sets of bacterial genomes and pulling synthetic sequencing reads from them, allowing for the evaluation of HUMAn2's accuracy and performance against expected profiles.

What are the different patterns of species contribution to a conserved pathway observed in the human gut microbiome using HUMAn2?

Different patterns include a complex attribution where multiple species contribute to a conserved pathway in varying mixtures across individuals, a per-person dominant attribution where one species dominates the contribution to the pathway for each individual, and a universal dominant pattern where one species consistently provides the pathway across the population.

How does HUMAn2 distinguish between functional potential and functional activity in a community?

HUMAn2 can analyze both DNA and RNA data, allowing it to distinguish between the functional potential (number of encoded pathways) and functional activity (actual expression of those pathways) within a community, showing that the two are not always the same.

How can one access and learn more about HUMAn2 and its related tools?

HUMAn2 can be found by searching its name online, and it is installable via source pip as a Python package and via Homebrew. There is a detailed user manual and an active user group on Google Groups for further support and information.

Outlines

00:00

🌱 Advances in Microbial Community Profiling

Orlando discusses the lab's work on profiling microbial communities from shotgun sequencing data, focusing on taxonomic and functional profiling. Taxonomic profiling identifies species and clades, while functional profiling examines gene families and pathways within the community. The challenge lies in the computational intensity and potential for error due to the vast size of data and databases. The lab developed MetaPhlan 2 to address taxonomic profiling by searching reads against a reduced, pre-selected set of marker genes unique to each clade or species, resulting in faster and more accurate taxonomic composition profiling.

05:00

🔍 Humann2: Tiered Read Mapping for Enhanced Functional Profiling

The script introduces Humann2, a method that improves the speed and accuracy of functional profiling by implementing tiered read mapping. The process begins with a rapid identification of species using a small, specific dataset of marker genes. This is followed by a custom database search using the pan genomes of detected species, excluding genomes with no evidence of presence. The final tier involves a comprehensive search for unmapped reads. Humann2's approach is shown to be more accurate and faster, with the ability to reconstruct functions on a species-by-species basis, providing both community totals and species-specific insights.

10:00

🧬 Synthetic Metagenomes and Real-World Data Analysis

To evaluate Humann2's effectiveness, synthetic metagenomes were created using bacterial genomes and sequencing reads. The method was tested on a challenging dataset with a wide abundance range and congeneric species. Humann2 demonstrated higher specificity and sensitivity compared to traditional methods, resulting in improved accuracy and a significant speedup. The method was also applied to real-world data from the Human Microbiome Project, showing similar performance with faster processing times and the ability to identify signature metabolic pathways conserved in specific body areas.

15:01

🌐 Humann2's Broader Impact on Microbiome Research

Humann2 has been used to analyze hundreds of human metagenomes, revealing distinct metabolic pathway signatures for different body areas. The method allows for species-level resolution of function, showing different patterns of species contribution to conserved pathways across individuals. It also distinguishes between functional potential and activity within a community. Humann2 is available for use and is part of a suite of tools developed by the Huttenhower lab for microbiome analysis, which will be further discussed in upcoming presentations.

📚 Availability and Further Resources for Humann2

Humann2 is accessible for those interested, available through a simple Google search and installable via pip for Python or Homebrew. The lab provides a detailed user manual and an active user group for support. The presentation also highlights the broader ecosystem of tools for microbiome analysis developed by the Huttenhower lab, with a technology track presentation planned for the following day. The team acknowledges the contributions of collaborators and the Human Microbiome Project for the valuable data used in their research.

Keywords

💡Microbial Communities

Microbial communities refer to diverse groups of microorganisms that live in a specific environment and interact with each other. In the context of the video, the focus is on profiling these communities using shotgun sequencing data to understand their composition and function. The script discusses how to identify the species within these communities and what roles they play, highlighting the importance of taxonomic and functional profiling.

💡Shotgun Sequencing

Shotgun sequencing is a method used in genomics to sequence DNA by breaking it into small, random pieces and then determining the sequence of each piece. The script mentions this technique as the starting point for profiling microbial communities, where short reads are generated and then compared against databases to identify the species and functions present.

💡Taxonomic Profiling

Taxonomic profiling is the process of identifying and categorizing the different species within a microbial community. The script explains that this involves answering the question 'who is there' by identifying the species and higher-level clades present in the community, which is a fundamental step in understanding the community's composition.

💡Functional Profiling

Functional profiling is the identification of the functions performed by the species within a microbial community, such as the gene families and pathway composition. The script emphasizes this aspect as the main focus of the talk, explaining how it differs from taxonomic profiling and is crucial for understanding what the species in a community are doing.

💡Metaphlan 2

Metaphlan 2 is a tool mentioned in the script for taxonomic profiling. It is designed to alleviate some of the computational challenges associated with searching reads against a large database by using a pre-selected set of marker genes that are unique for each clade or species. This allows for efficient and accurate profiling of the community's taxonomic composition.

💡Tiered Read Mapping

Tiered read mapping is a strategy implemented in the tool HUMANn2, as discussed in the script. It involves searching reads through different tiers or levels of databases, starting with a small, specific dataset of species-specific marker genes, followed by a custom database built from the pan-genomes of detected species, and finally, a more traditional comprehensive search. This approach aims to improve the accuracy, speed, and resolution of functional profiling.

💡Pan-genome

A pan-genome refers to the complete set of genes found within all the isolates of a given species. In the script, the concept is used to describe the custom database created for each sample after the initial taxonomic prescreen step, which includes the genomes or pan-genomes of the species believed to be present in the community.

💡HUMANn2

HUMANn2 is a method developed for improving the accuracy and speed of functional profiling in microbial communities. The script describes it as implementing tiered read mapping and being able to reconstruct functions on a species-by-species basis within the community, providing both community totals of different functions and the ability to attribute these functions to specific species.

💡Synthetic Metagenomes

Synthetic metagenomes are artificial datasets created for testing and validation purposes. In the script, they are used to evaluate the performance of HUMANn2 by simulating a complex microbial community with a wide range of species abundances and genetic similarities, allowing the researchers to assess the accuracy and sensitivity of the method.

💡Metabolic Pathways

Metabolic pathways are a series of chemical reactions in a cell that are linked together to produce a certain product or carry out a specific function. The script discusses how HUMANn2 can identify and quantify these pathways within a microbial community, providing insights into the functional capabilities of the community and how different species contribute to these pathways.

Highlights

Orlando discusses advancements in profiling microbial communities from shotgun sequencing data.

Two primary questions in microbial community profiling: taxonomic profiling and functional profiling.

Taxonomic profiling identifies species and clades present in a microbial community.

Functional profiling focuses on gene families and pathway composition within the community.

Metaphlan 2 is a tool for taxonomic profiling using pre-selected marker genes.

Metaphlan 2's efficiency comes from reduced database size and specificity of marker genes.

Humann2 introduces a tiered read mapping strategy for functional profiling.

Tiered read mapping starts with rapid species identification using marker genes.

Second tier involves building a custom database from detected species' pan genomes.

Remaining reads are searched against a comprehensive protein database in the final tier.

Humann2 provides functional profiles stratified by contributing organism and unclassified abundance.

Humann2 can quantify gene families and collapse their abundance into pathways.

Synthetic metagenomes are used to validate Humann2's accuracy and performance.

Humann2 outperforms traditional methods in precision, sensitivity, and speed.

Humann2 enables species-level resolution of functional profiles in microbial communities.

The method has been applied to hundreds of human metagenomes from the Human Microbiome Project.

Humann2 can identify metabolic pathway signatures specific to different body areas.

Species-level profiling reveals different patterns of contribution to conserved pathways.

RNA data analysis with Humann2 shows a distinction between functional potential and activity.

Humann2 is available as a Python package and part of a broader suite of microbiome analysis tools.

The Huttlehauer lab offers a technology track presentation for an overview of their tools.

Transcripts

mv orlando talking about some of the

work

in our lab profiling microbial

communities from shotgun sequencing data

and we'll be talking about some recent

advancements in that area today

so when we're talking about the

microbial community there's two primary

questions or types of questions we're

trying to answer from sequencing data

one is this question of who is there

which is the issue of taxonomic

profiling of identifying the species and

higher level clades that are present in

that microbial community

and the second question which will be

the closer focus for today

is a question of what those species are

doing which is functional profiling

identifying the gene families

and pathway composition of that

community and both of these because

they're starting from sequence

data are classic bioinformatics problems

in in sequence search so that's where we

begin our story um so to actually do

this to actually do either of these

types of profiling

we're interested in searching short

reads a shotgun sequence metagenome or

meta transcriptome

against a vast database of microbial

reference genomes

and as you can imagine as the size of

these data metagenomes

increases as the size of this database

increases with new isolate genomes being

sequenced every year

this is a very computationally intensive

problem and there's also a lot of

opportunity for error here if we're

spuriously mapping

reeds where they don't belong and so

this is really where we find ourselves

in trying to solve these profiling

questions

previously we developed a technique for

taxonomic profiling to alleviate some of

those issues

and more specifically what we're doing

there in this tool called metaphon 2

is instead of searching reads against

the entire database

is to search them against a pre-selected

set of marker genes that are unique for

each clade or species

across the community so for example here

we have isolated this a gene here that's

well conserved within the yellow species

we always see it in isolates of that

species

but it's not seen anywhere else so if

that gene recruits a reed if a reed maps

to that gene in the community

it's sort of like a little name tag

telling us that the yellow species

is there and we can use this technique

to very efficiently and accurately

profile the taxonomic composition of a

community this is a reduced database so

it gives us a nice speed

uh bonus and we've also pre-selected

these genes to be very specific so we

know when they recruit reads

that were um that they're being assigned

at the correct place

the issue in texas and functional

profiling is we're not interested in

just a subset of genes but rather we're

interested in all the genes and pathways

in the community

but in today's talk what i would like to

to go over is how we can leverage this

idea

of being able to rapidly and accurately

identify the species in a community

in order to improve the accuracy speed

and resolution of functional profiling

so the method that we developed for that

is called human2 and it implements a

strategy called

tiered read mapping and i'll go through

what that means now

so the idea here is that we're starting

from a shotgun sequenced

metagenome or meta transcriptome from a

microbial community and then we're going

to search these reads through a set of

tiers of different searches against

different databases

the first search tier is what i just

described in the previous slide

we're going to attempt to rapidly

identify the species in this community

by mapping these reeds against the small

and highly specific data set

of species-specific marker genes and so

you can see here in this example

that genes are being recruited to this

marker gene from the blue species

and the orange species here indicating

that they are likely to be present in

this community

but not to this green species indicating

that it's likely absent from that

community

and this is our first search tier the

second search tier which is really the

meat of this whole process then

is to build a custom database for this

sample by concatenating the pan genomes

of species that we detected in this

taxonomic prescreen step

so now we're going to do is do a

detailed search of all the remaining

reeds

against the genomes or pan genomes of

species we believe to be present in the

community

we're throwing away this genome here

we're not including this genome here

because we had no

evidence that it was actually present in

the community and although i'm just

showing one here being excluded

in reality this process is excluding

many many pan genomes that we won't

search through in this second tier of

the search

in the last tier of the search anything

that doesn't map in this process will

then

let flow into a more traditional

comprehensive search strategy so we'll

try to explain as much as we can

as quickly and as specifically as

possible and then what's left over will

take a

a traditional approach and just search

comprehensively by translated search

against a protein database at the very

end of this some reads will still map

nowhere they don't map to any reference

and these are set aside for possible

assembly downstream and outside of human

too

so the end result of this tiered search

are functional profiles of a metagenome

or meta transcriptome looking something

like this

where we have a particular function in

this case a gene family

and it's stratified by both contributing

organism as well as unclassified

abundance that we couldn't assign to a

particular species

that adds up to a total here this is for

gene families

once we've actually quantified the gene

families in the community

for genes like imp dehydrogenase that

participate in a metabolic pathway we

can collapse the abundance of multiple

genes

into a smaller subset of pathways which

is

more tractable to work with downstream

and we end up with similar looking data

where we get four each pathway

an abundance at the community level as

well as stratified by the species that

contributed to that pathway

and a measure of pathway coverage which

is a measure of our confidence that the

pathway is actually complete within this

particular sample

so these are what the outputs look like

for a typical run of human 2 on a

metagenome or metatranscriptome

to evaluate that this actually works and

it's behaving the way that we expect it

to

we were able to create synthetic

metagenomes or meta transcriptomes

by taking sets of bacterial genomes and

pulling synthetic sequencing reads from

them

and so in the example i'll go over here

we have a selection of 20 bacterial

species that are the most commonly

occurring species in the human gut

microbiome

and what we've done is to sample wreaths

i think the color there just shifted if

there's anything we're able to do

with that on the projector if not you

can adjust your eyes

um so we have sampled these reeds in a

staggered composition here

such that the most abundant species in

this

synthetic metagenome is about a thousand

times more abundant than the least

abundant species and so this makes for a

challenging

problem here and that the species have a

really broad range of

abundance brilliance also you can see

challenging us here the fact that we

have

congeneric species multiple species

within the same genus

and so there's a lot of homology we

expect among these congeneric species

which can make

mapping reuse two specific species

within that gene is more difficult

so once we have this synthetic

metagenome we can create an expected

profile of what genes and pathways we

expect to observe

and then analyze this meta genome using

different methods and see how well they

do

when we actually analyze this using a

traditional method a traditional

comprehensive search

we actually see a lot of error due to

spurious mapping we have

reads that are mapping where they're not

supposed to across those broad databases

which hurts our precision

as well as reads that we're supposed to

map to a gene and wound up mapping

somewhere else which hurts our

sensitivity

in contrast humantu's tiered search is

both more specific

and more sensitive in terms of putting

reads in the right place which gives us

a nice boost in overall accuracy

in addition because the tiered search is

trying to explain as much as possible

using that reduced pan genome database

it tends to process the metagenome a lot

faster than the comprehensive search

you're spending more time working with a

small database than you are the large

database with the tiered search

and so in this particular synthetic

example it was about a 7x speedup

but the last thing which is really one

of the key advantages of human 2 here is

that in addition to getting

us to community totals of different

functions which both methods can do

human 2's tiered search is able to

reconstruct functions on a species by

species basis within the community

and what we can see here is our

sensitivity for those functions across

species is very very high

down to about 1x coverage once we get

below 1x coverage

you're actually not sampling the entire

genome anymore and so you're well the

gold standard says you should be able to

find everything

because breeds weren't necessarily

sampled from every gene we don't see

them

however it's critical to note that our

precision remains very high even for

these low abundant species meaning that

their genomes are in this database

they're in this custom database

but they're not recruiting reads that

they're not supposed to so we were very

happy with the overall performance of

the method here

in terms of performance on real world

data we've used human 2 to profile

hundreds of human metagenomes from the

human microbiome project

and there we tend to see similar

performance that we're able to move very

quickly through these meta genomes

in the pan genome search stage about an

order to two orders of magnitude faster

than in the translated search

now that doesn't work that well for us

if we don't end up explaining a lot of

reads in pan genome search

and indeed what we find is that up to

about 60 percent of reeds in a typical

metagenome are explained by the pangeno

mapping during that fast step

with about 15 percent of additional

reads explained when we then

take the rest of the reeds and push them

through pan genomes translated search

so we're really explaining the majority

of what we can explain during this

accelerated

initial tier of the search

so that's that's some benchmarking and

accuracy stats for you

moving into some actual science we've

then looked at the profiles that human 2

produced

from those hmp metagenomes and we've

isolated metabolic pathways that we call

signatures for particular body areas

meaning that they're really well

conserved within a particular body area

and tend to be absent from other body

areas so in this example at the top

romnos degradation we see that it's

quite abundant and conserved at the in

gut metagenomes across individuals and

fairly rare at these three other

microbiome body sites so this is sort of

a grand overview

if we zoom in on that top example here

to see how this looks sample by sample

we can see that indeed human 2 is

showing us a lot of very consistent

abundance for this pathway

of about six parts per thousand across

the gut meta genomes and then it drops

off very quickly thereafter that we

don't really see much abundance for

those pathways at the other

sites in addition this is a little

tricky to see here but this light gray

down here is the

unclassified amount of the uh the

pathway that's

identified during the translated search

the rest of this in darker gray

outside that box is actually being

assigned a particular species so in this

particular example

not only are we seeing this pathway

consistently across gut meta genomes but

we're able to assign it

assign the majority of the copies of

that pathway to particular species

and we actually take those signature

pathways and dig into that species level

attribution

we see some interesting patterns so for

example if we zoom in on what's going on

in that little box there for the gut

meta genomes and take a look at the

species attribution

we can see different patterns of how

species contribute to a conserved

pathway or how a pathway is conserved

across different individuals

in the case of this gut pathway romney's

degradation

the attribution is actually relatively

complex that although the total is

fairly well conserved across individuals

we see very different mixtures of

species contributing the pathway from

one person to another

most individuals have a handful of

species that are contributing

and those aren't necessarily the same

from one person to the next

suggesting that the overall abundance of

this pathway seems to be more conserved

than its taxonomic attribution

i'll contrast that with another

mechanism where we can observe a per

person

dominant attribution of the pathway so

for example this is peptidoglycan

biosynthesis from the vaginal microbiome

here again we see a relatively constant

abundance across individuals

but a very different pattern of

attribution that each individual is

dominated by about

one species mostly in the the genus

lactobacillus

and so while they all wind up with about

the same abundance of the pathway it

tends to be contributed by just

one species per person that may differ

between people

a last mechanism is a universal dominant

pattern of attribution and

an example of this is trellis

degradation on the skin

this is a pathway that's provided by

propianobacterium acnes it's not really

seen

at other body sites and on the skin this

pathway is fairly consistently

and completely provided by just

propionibacterium

acnes across the population so unlike

the previous two examples

where different species could contribute

the pathway in different mixtures

here on the skin we're really just

seeing this one very common skin bug

prop acne is contributing this

signature pathway for the skin across

individuals

and so humantu's tier search combined

with the species level profiling

allows us to see this sort of taxonomic

resolution to function

that we haven't been able to do in

previous approaches

as a final biological example we'll move

outside of the human microbiome project

to another cohort that

two of my colleagues jason and gallop

work on

this is a cohort of health professionals

within the boston area where we have

both

metagenomes and metatranscriptomes of

the human gut

and jason and elite have profiled these

samples using human two

and we look at the dna level

contributions for for pathways in this

uh

group we see a lot of things that are

similar to that first mechanism i showed

a

complex attribution where a particular

pathway is relatively conserved in the

gut but can be contributed by multiple

organisms per person and those

potentially differ between people

so this is looking at the the dna level

we see that this is also a fairly

conserved pattern between individuals of

this mixture of bugs

when we look at the rna data however we

see a very different picture

that there's more of a gradient of some

people having a complex attribution

pattern in the rna pool

whereas in others the rna pool is

completely dominated by a single species

tequila bacterium prismiciae

and so here what we're seeing from human

too is this ability to distinguish

the functional potential of a community

in this case

numbers or relative copy numbers of

encoded pathways across bugs

with functional activity the actual

relative expression of that

that pathway within a community and see

that the two are not always the same

so in summary human too implements this

tiered approach which is a new approach

to functional profiling that aims to

explain as many reads as possible with

progressively

broader and less specific databases this

approach is more accurate and a lot

faster than the traditional approach of

just doing a comprehensive search

against an exhaustive database

and lastly we get stratification of our

results by species for free in this

process which allows us to access both

those questions of who's there

and what they're doing at the same time

which is increasingly what we want to

know not just what a community is doing

but what species are actually performing

those functions in the community

human 2 is available now if this is

something that's of interest to you it's

the first hit if you google humond 2.

it's installable via source pip as a

python package and

via homebrew we have a fairly detailed

user manual as well as a very active

user group on google groups where you

and i can converse by email if you'd

like

and so i'd encourage you to try it out

human 2 is part of a broader menagerie

of tools that the huttonhauer lab has

put together for analyzing microbiomes

both in terms of profiling data

profiling metagenomes from raw

sequencing as well as doing downstream

statistical analysis on those results if

you'd like to learn more about that

overall system

we'll be doing a technology track

presentation tomorrow evening that will

give more of a broader survey of these

tools than i've done now

and also if you stay tuned my colleague

ali will be presenting one of our

statistical methods for the analysis of

paired high-dimensional data sets

so a big thanks to the lab especially

the human 2 team highlighted there in

green for

having a lot of fun with this project

our collaborators on human 2 and

also the entire human microbiome project

for providing a lot of excellent data

for us to work with

thank you