NAME
AI::TensorFlow::Libtensorflow::Manual::Notebook::InferenceUsingTFHubEnformerGeneExprPredModel - Using TensorFlow to do gene expression prediction using a pre-trained model
SYNOPSIS
The following tutorial is based on the Enformer usage notebook. It uses a pre-trained model based on a transformer architecture trained as described in Avsec et al (2021) and introduced in Avsec's DeepMind blog post Predicting gene expression with AI.
Running the code requires an Internet connection to download the model (from Google servers) and datasets (from GitHub, UCSC, and NIH).
Some of this code is identical to that of InferenceUsingTFHubMobileNetV2Model
notebook. Please look there for explanation for that code. As stated there, this will later be wrapped up into a high-level library to hide the details behind an API.
NOTE: If running this model, please be aware that
the Docker image takes 3 GB or more of disk space;
the model and data takes 5 GB or more of disk space.
meaning that you will need a total of 8 GB of disk space. You may need at least 4 GB of free memory to run the model.
COLOPHON
The following document is either a POD file which can additionally be run as a Perl script or a Jupyter Notebook which can be run in IPerl (viewable online at nbviewer). If you are reading this as POD, there should be a generated list of Perl dependencies in the CPANFILE section.
You will also need the executables gunzip
, bgzip
, and samtools
. Furthermore,
Bio::DB::HTS
requireslibhts
andPDL::Graphics::Gnuplot
requiresgnuplot
.
If you are running the code, you may optionally install the tensorflow
Python package in order to access the saved_model_cli
command, but this is only used for informational purposes.
TUTORIAL
Load the library
First, we need to load the AI::TensorFlow::Libtensorflow
library and more helpers. We then create an AI::TensorFlow::Libtensorflow::Status
object and helper function to make sure that the calls to the libtensorflow
C library are working properly.
use strict;
use warnings;
use utf8;
use constant IN_IPERL => !! $ENV{PERL_IPERL_RUNNING};
no if IN_IPERL, warnings => 'redefine'; # fewer messages when re-running cells
use feature qw(say);
use Syntax::Construct qw( // );
use lib::projectroot qw(lib);
BEGIN {
if( IN_IPERL ) {
$ENV{TF_CPP_MIN_LOG_LEVEL} = 3;
}
require AI::TensorFlow::Libtensorflow;
}
use URI ();
use HTTP::Tiny ();
use Path::Tiny qw(path);
use File::Which ();
use List::Util ();
use Data::Printer ( output => 'stderr', return_value => 'void', filters => ['PDL'] );
use Data::Printer::Filter::PDL ();
use Text::Table::Tiny qw(generate_table);
my $s = AI::TensorFlow::Libtensorflow::Status->New;
sub AssertOK {
die "Status $_[0]: " . $_[0]->Message
unless $_[0]->GetCode == AI::TensorFlow::Libtensorflow::Status::OK;
return;
}
AssertOK($s);
And create helpers for converting between PDL
ndarrays and TFTensor
ndarrays.
use PDL;
use AI::TensorFlow::Libtensorflow::DataType qw(FLOAT);
use FFI::Platypus::Memory qw(memcpy);
use FFI::Platypus::Buffer qw(scalar_to_pointer);
sub FloatPDLTOTFTensor {
my ($p) = @_;
return AI::TensorFlow::Libtensorflow::Tensor->New(
FLOAT, [ reverse $p->dims ], $p->get_dataref, sub { undef $p }
);
}
sub FloatTFTensorToPDL {
my ($t) = @_;
my $pdl = zeros(float,reverse( map $t->Dim($_), 0..$t->NumDims-1 ) );
memcpy scalar_to_pointer( ${$pdl->get_dataref} ),
scalar_to_pointer( ${$t->Data} ),
$t->ByteSize;
$pdl->upd_data;
$pdl;
}
Download model and data
Enformer model from
> Avsec Ž, Agarwal V, Visentin D, Ledsam JR, Grabska-Barwinska A, Taylor KR, Assael Y, Jumper J, Kohli P, Kelley DR. Effective gene expression prediction from sequence by integrating long-range interactions. I<Nat Methods>. 2021 Oct;B<18(10)>:1196-1203. doi: L<10.1038/s41592-021-01252-x|https://doi.org/10.1038/s41592-021-01252-x>. Epub 2021 Oct 4. PMID: L<34608324|https://pubmed.ncbi.nlm.nih.gov/34608324>; PMCID: L<PMC8490152|https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8490152/>.
Human target dataset from
> Kelley DR. Cross-species regulatory sequence activity prediction. I<PLoS Comput Biol>. 2020 Jul 20;B<16(7)>:e1008050. doi: L<10.1371/journal.pcbi.1008050|https://doi.org/10.1371/journal.pcbi.1008050>. PMID: L<32687525|https://pubmed.ncbi.nlm.nih.gov/32687525>; PMCID: L<PMC7392335|https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7392335/>.
UCSC hg38 genome. More info at http://hgdownload.cse.ucsc.edu/goldenPath/hg38/bigZips/; Genome Reference Consortium Human Build 38 from Genome Reference Consortium.
> Schneider VA, Graves-Lindsay T, Howe K, Bouk N, Chen HC, Kitts PA, Murphy TD, Pruitt KD, Thibaud-Nissen F, Albracht D, Fulton RS, Kremitzki M, Magrini V, Markovic C, McGrath S, Steinberg KM, Auger K, Chow W, Collins J, Harden G, Hubbard T, Pelan S, Simpson JT, Threadgold G, Torrance J, Wood JM, Clarke L, Koren S, Boitano M, Peluso P, Li H, Chin CS, Phillippy AM, Durbin R, Wilson RK, Flicek P, Eichler EE, Church DM. Evaluation of GRCh38 and de novo haploid genome assemblies demonstrates the enduring quality of the reference assembly. I<Genome Res.> 2017 May;B<27(5)>:849-864. doi: L<10.1101/gr.213611.116|https://doi.org/10.1101/gr.213611.116>. Epub 2017 Apr 10. PMID: L<28396521|https://pubmed.ncbi.nlm.nih.gov/28396521>; PMCID: L<PMC5411779|https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5411779/>.
ClinVar file
> Landrum MJ, Lee JM, Benson M, Brown GR, Chao C, Chitipiralla S, Gu B, Hart J, Hoffman D, Jang W, Karapetyan K, Katz K, Liu C, Maddipatla Z, Malheiro A, McDaniel K, Ovetsky M, Riley G, Zhou G, Holmes JB, Kattman BL, Maglott DR. ClinVar: improving access to variant interpretations and supporting evidence. I<Nucleic Acids Res.> 2018 Jan 4;B<46(D1)>:D1062-D1067. doi: L<10.1093/nar/gkx1153|https://doi.org/10.1093/nar/gkx1153>. PMID: L<29165669|https://pubmed.ncbi.nlm.nih.gov/29165669>; PMCID: L<PMC5753237|https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5753237/>.
# Model handle
my $model_uri = URI->new( 'https://tfhub.dev/deepmind/enformer/1' );
$model_uri->query_form( 'tf-hub-format' => 'compressed' );
my $model_base = substr( $model_uri->path, 1 ) =~ s,/,_,gr;
my $model_archive_path = "${model_base}.tar.gz";
my $model_sequence_length = 393_216; # bp
# Human targets from Basenji2 dataset
my $targets_uri = URI->new('https://raw.githubusercontent.com/calico/basenji/master/manuscripts/cross2020/targets_human.txt');
my $targets_path = 'targets_human.txt';
# Human reference genome
my $hg_uri = URI->new("http://hgdownload.cse.ucsc.edu/goldenPath/hg38/bigZips/hg38.fa.gz");
my $hg_gz_path = "hg38.fa.gz";
# From http://hgdownload.cse.ucsc.edu/goldenPath/hg38/bigZips/md5sum.txt
my $hg_md5_digest = "1c9dcaddfa41027f17cd8f7a82c7293b";
my $clinvar_uri = URI->new('https://ftp.ncbi.nlm.nih.gov/pub/clinvar/vcf_GRCh38/clinvar.vcf.gz');
my $clinvar_path = 'clinvar.vcf.gz';
my $http = HTTP::Tiny->new;
for my $download ( [ $model_uri => $model_archive_path ],
[ $targets_uri => $targets_path ],
[ $hg_uri => $hg_gz_path ],
[ $clinvar_uri => $clinvar_path ],) {
my ($uri, $path) = @$download;
say "Downloading $uri to $path";
next if -e $path;
$http->mirror( $uri, $path );
}
STREAM (STDOUT):
Downloading https://tfhub.dev/deepmind/enformer/1?tf-hub-format=compressed to deepmind_enformer_1.tar.gz
Downloading https://raw.githubusercontent.com/calico/basenji/master/manuscripts/cross2020/targets_human.txt to targets_human.txt
Downloading http://hgdownload.cse.ucsc.edu/goldenPath/hg38/bigZips/hg38.fa.gz to hg38.fa.gz
Downloading https://ftp.ncbi.nlm.nih.gov/pub/clinvar/vcf_GRCh38/clinvar.vcf.gz to clinvar.vcf.gz
Now we
extract the saved model so that we can load it and
check the MD5 sum of the reference genome to make sure it was downloaded correctly.
use Archive::Extract;
$Archive::Extract::DEBUG = 1;
$Archive::Extract::PREFER_BIN = 1; # for the larger model, prefer bin
if( ! -e $model_base ) {
my $ae = Archive::Extract->new( archive => $model_archive_path );
die "Could not extract archive" unless $ae->extract( to => $model_base );
}
use Digest::file qw(digest_file_hex);
if( digest_file_hex( $hg_gz_path, "MD5" ) eq $hg_md5_digest ) {
say "MD5 sum for $hg_gz_path OK";
} else {
die "Digest for $hg_gz_path failed";
}
STREAM (STDOUT):
MD5 sum for hg38.fa.gz OK
RESULT:
1
In order to quickly seek for sequences in the reference genome FASTA, we
convert the gzip'd file into a block gzip'd file and
index that
.bgz
file usingfaidx
fromsamtools
.
(my $hg_uncompressed_path = $hg_gz_path) =~ s/\.gz$//;
my $hg_bgz_path = "${hg_uncompressed_path}.bgz";
use IPC::Run;
if( ! -e $hg_bgz_path ) {
IPC::Run::run(
[ qw(gunzip -c) ], '<', $hg_gz_path,
'|',
[ qw(bgzip -c) ], '>', $hg_bgz_path
);
}
use Bio::Tools::Run::Samtools;
my $hg_bgz_fai_path = "${hg_bgz_path}.fai";
if( ! -e $hg_bgz_fai_path ) {
my $faidx_tool = Bio::Tools::Run::Samtools->new( -command => 'faidx' );
$faidx_tool->run( -fas => $hg_bgz_path )
or die "Could not index FASTA file $hg_bgz_path: " . $faidx_tool->error_string;
}
Model input and output specification
Now we create a helper to call saved_model_cli
and called saved_model_cli scan
to ensure that the model is I/O-free for security reasons.
sub saved_model_cli {
my (@rest) = @_;
if( File::Which::which('saved_model_cli')) {
system(qw(saved_model_cli), @rest ) == 0
or die "Could not run saved_model_cli";
} else {
warn "saved_model_cli(): Install the tensorflow Python package to get the `saved_model_cli` command.\n";
return -1;
}
}
say "Checking with saved_model_cli scan:";
saved_model_cli( qw(scan),
qw(--dir) => $model_base,
);
STREAM (STDOUT):
Checking with saved_model_cli scan:
MetaGraph with tag set ['serve'] does not contain the default denylisted ops: {'ReadFile', 'PrintV2', 'WriteFile'}
RESULT:
1
We need to see what the inputs and outputs of this model are so saved_model_cli show
should show us that:
saved_model_cli( qw(show),
qw(--dir) => $model_base,
qw(--all),
);
STREAM (STDOUT):
MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:
signature_def['__saved_model_init_op']:
The given SavedModel SignatureDef contains the following input(s):
The given SavedModel SignatureDef contains the following output(s):
outputs['__saved_model_init_op'] tensor_info:
dtype: DT_INVALID
shape: unknown_rank
name: NoOp
Method name is:
Concrete Functions:
B<RESULT>:
1
It appears that it does not! What we can do is load the model using tensorflow
in Python and then save it to a new path. Now when we run saved_model_cli show
on this new model path, it shows the correct inputs and outputs.
my $new_model_base = "${model_base}_new";
system( qw(python3), qw(-c) => <<EOF, $model_base, $new_model_base ) unless -e $new_model_base;
import sys
import tensorflow as tf
in_path, out_path = sys.argv[1:3]
imported_model = tf.saved_model.load(in_path).model
tf.saved_model.save( imported_model , out_path )
EOF
saved_model_cli( qw(show),
qw(--dir) => $new_model_base,
qw(--all),
);
STREAM (STDOUT):
MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:
signature_def['__saved_model_init_op']:
The given SavedModel SignatureDef contains the following input(s):
The given SavedModel SignatureDef contains the following output(s):
outputs['__saved_model_init_op'] tensor_info:
dtype: DT_INVALID
shape: unknown_rank
name: NoOp
Method name is:
signature_def['serving_default']:
The given SavedModel SignatureDef contains the following input(s):
inputs['args_0'] tensor_info:
dtype: DT_FLOAT
shape: (-1, 393216, 4)
name: serving_default_args_0:0
The given SavedModel SignatureDef contains the following output(s):
outputs['human'] tensor_info:
dtype: DT_FLOAT
shape: (-1, 896, 5313)
name: StatefulPartitionedCall:0
outputs['mouse'] tensor_info:
dtype: DT_FLOAT
shape: (-1, 896, 1643)
name: StatefulPartitionedCall:1
Method name is: tensorflow/serving/predict
Concrete Functions:
Function Name: 'predict_on_batch'
Option #1
Callable with:
Argument #1
args_0: TensorSpec(shape=(None, 393216, 4), dtype=tf.float32, name='args_0')
RESULT:
1
We want to use the serve
tag-set and
the input
args_0
which has the nameserving_default_args_0:0
andthe output
human
which has the nameStatefulPartitionedCall:0
.
all of which are DT_FLOAT
.
Make note of the shapes that those take. Per the model description at TensorFlow Hub:
The input sequence length is 393,216 with the prediction corresponding to 128 base pair windows for the center 114,688 base pairs. The input sequence is one hot encoded using the order of indices corresponding to 'ACGT' with N values being all zeros.
The input shape (-1, 393216, 4)
thus represents dimensions [batch size] x [sequence length] x [one-hot encoding of ACGT]
.
The output shape (-1, 896, 5313)
represents dimensions [batch size] x [ predictions along 114,688 base pairs / 128 base pair windows ] x [ human target by index ]
. We can confirm this by doing some calculations:
my $model_central_base_pairs_length = 114_688; # bp
my $model_central_base_pair_window_size = 128; # bp / prediction
say "Number of predictions: ", $model_central_base_pairs_length / $model_central_base_pair_window_size;
STREAM (STDOUT):
Number of predictions: 896
RESULT:
1
and by looking at the targets file:
use Data::Frame;
my $df = Data::Frame->from_csv( $targets_path, sep => "\t" )
->transform({
file => sub {
my ($col, $df) = @_;
# clean up the paths in 'file' column
[map { join "/", (split('/', $_))[7..8] } $col->list];
}
});
say "Number of targets: ", $df->nrow;
say "";
say "First 5:";
say $df->head(5);
STREAM (STDOUT):
Number of targets: 5313
First 5:
------------------------------------------------------------------------------------------------------------------------------------------------
index genome identifier file clip scale sum_stat description
------------------------------------------------------------------------------------------------------------------------------------------------
0 0 0 ENCFF833POA encode/ENCSR000EIJ 32 2 mean DNASE:cerebellum male adult (27 years) and male adult (35 years)
1 1 0 ENCFF110QGM encode/ENCSR000EIK 32 2 mean DNASE:frontal cortex male adult (27 years) and male adult (35 years)
2 2 0 ENCFF880MKD encode/ENCSR000EIL 32 2 mean DNASE:chorion
3 3 0 ENCFF463ZLQ encode/ENCSR000EIP 32 2 mean DNASE:Ishikawa treated with 0.02% dimethyl sulfoxide for 1 hour
4 4 0 ENCFF890OGQ encode/ENCSR000EIS 32 2 mean DNASE:GM03348
------------------------------------------------------------------------------------------------------------------------------------------------
RESULT:
1
Load the model
Let's now load the model in Perl and get the inputs and outputs into a data structure by name.
my $opt = AI::TensorFlow::Libtensorflow::SessionOptions->New;
my @tags = ( 'serve' );
my $graph = AI::TensorFlow::Libtensorflow::Graph->New;
my $session = AI::TensorFlow::Libtensorflow::Session->LoadFromSavedModel(
$opt, undef, $new_model_base, \@tags, $graph, undef, $s
);
AssertOK($s);
my %puts = (
## Inputs
inputs_args_0 =>
AI::TensorFlow::Libtensorflow::Output->New({
oper => $graph->OperationByName('serving_default_args_0'),
index => 0,
}),
## Outputs
outputs_human =>
AI::TensorFlow::Libtensorflow::Output->New({
oper => $graph->OperationByName('StatefulPartitionedCall'),
index => 0,
}),
outputs_mouse =>
AI::TensorFlow::Libtensorflow::Output->New({
oper => $graph->OperationByName('StatefulPartitionedCall'),
index => 1,
}),
);
p %puts;
STREAM (STDERR):
{
inputs_args_0 AI::TensorFlow::Libtensorflow::Output {
index 0,
oper AI::TensorFlow::Libtensorflow::Operation {
Name "serving_default_args_0",
NumInputs 0,
NumOutputs 1,
OpType "Placeholder"
}
},
outputs_human AI::TensorFlow::Libtensorflow::Output {
index 0,
oper AI::TensorFlow::Libtensorflow::Operation {
Name "StatefulPartitionedCall",
NumInputs 274,
NumOutputs 2,
OpType "StatefulPartitionedCall"
}
},
outputs_mouse AI::TensorFlow::Libtensorflow::Output {
index 1,
oper AI::TensorFlow::Libtensorflow::Operation {
Name "StatefulPartitionedCall",
NumInputs 274,
NumOutputs 2,
OpType "StatefulPartitionedCall"
}
}
}
We need a helper to simplify running the session and getting just the predictions that we want.
my $predict_on_batch = sub {
my ($session, $t) = @_;
my @outputs_t;
$session->Run(
undef,
[$puts{inputs_args_0}], [$t],
[$puts{outputs_human}], \@outputs_t,
undef,
undef,
$s
);
AssertOK($s);
return $outputs_t[0];
};
undef;
Encoding the data
The model specifies that the way to get a sequence of DNA bases into a TFTensor
is to use one-hot encoding in the order ACGT
.
This means that the bases are represented as vectors of length 4:
| base | vector encoding | |------|-----------------| | A | [1 0 0 0]
| | C | [0 1 0 0]
| | G | [0 0 1 0]
| | T | [0 0 0 1]
| | N | [0 0 0 0]
|
We can achieve this encoding by creating a lookup table with a PDL ndarray. This could be done by creating a byte PDL ndarray of dimensions [ 256 4 ]
to directly look up the the numeric value of characters 0-255, but here we'll go with a smaller [ 5 4 ]
ndarray and transliterate the numeric values of NACGT
to 0-4.
use PDL;
our $SHOW_ENCODER = 1;
sub one_hot_dna {
my ($seq) = @_;
my $from_alphabet = "NACGT";
my $to_alphabet = pack "C*", 0..length($from_alphabet)-1;
# sequences from UCSC genome have both uppercase and lowercase bases
my $from_alphabet_tr = $from_alphabet . lc $from_alphabet;
my $to_alphabet_tr = $to_alphabet x 2;
my $p = zeros(byte, bytes::length($seq));
my $p_dataref = $p->get_dataref;
${ $p_dataref } = $seq;
eval "tr/$from_alphabet_tr/$to_alphabet_tr/" for ${ $p_dataref };
$p->upd_data;
my $encoder = append(float(0), identity(float(length($from_alphabet)-1)) );
say "Encoder is\n", $encoder->info, $encoder if $SHOW_ENCODER;
my $encoded = $encoder->index( $p->dummy(0) );
return $encoded;
}
####
{
say "Testing one-hot encoding:\n";
my $onehot_test_seq = "ACGTNtgcan";
my $test_encoded = one_hot_dna( $onehot_test_seq );
$SHOW_ENCODER = 0;
say "One-hot encoding of sequence '$onehot_test_seq' is:";
say $test_encoded->info, $test_encoded;
}
STREAM (STDOUT):
Testing one-hot encoding:
Encoder is
PDL: Float D [5,4]
[
[0 1 0 0 0]
[0 0 1 0 0]
[0 0 0 1 0]
[0 0 0 0 1]
]
One-hot encoding of sequence 'ACGTNtgcan' is:
PDL: Float D [4,10]
[
[1 0 0 0]
[0 1 0 0]
[0 0 1 0]
[0 0 0 1]
[0 0 0 0]
[0 0 0 1]
[0 0 1 0]
[0 1 0 0]
[1 0 0 0]
[0 0 0 0]
]
RESULT:
1
Note that in the above, the PDL ndarray's
first dimension is 4 which matches the last dimension of the input
TFTensor
;second dimension is the sequence length which matches the penultimate dimension of the input
TFTensor
.
Now we need a way to deal with the sequence interval. We're going to use 1-based coordinates as BioPerl does. In fact, we'll extend a BioPerl class.
package Interval {
use Bio::Location::Simple ();
use parent qw(Bio::Location::Simple);
sub center {
my $self = shift;
my $center = int( ($self->start + $self->end ) / 2 );
my $delta = ($self->start + $self->end ) % 2;
return $center + $delta;
}
sub resize {
my ($self, $width) = @_;
my $new_interval = $self->clone;
my $center = $self->center;
my $half = int( ($width-1) / 2 );
my $offset = ($width-1) % 2;
$new_interval->start( $center - $half - $offset );
$new_interval->end( $center + $half );
return $new_interval;
}
use overload '""' => \&_op_stringify;
sub _op_stringify { sprintf "%s:%s", $_[0]->seq_id // "(no sequence)", $_[0]->to_FTstring }
}
#####
{
say "Testing interval resizing:\n";
sub _debug_resize {
my ($interval, $to, $msg) = @_;
my $resized_interval = $interval->resize($to);
die "Wrong interval size for $interval --($to)--> $resized_interval"
unless $resized_interval->length == $to;
say sprintf "Interval: %s -> %s, length %2d : %s",
$interval,
$resized_interval, $resized_interval->length,
$msg;
}
for my $interval_spec ( [4, 8], [5, 8], [5, 9], [6, 9]) {
my ($start, $end) = @$interval_spec;
my $test_interval = Interval->new( -seq_id => 'chr11', -start => $start, -end => $end );
say sprintf "Testing interval %s with length %d", $test_interval, $test_interval->length;
say "-----";
for(0..5) {
my $base = $test_interval->length;
my $to = $base + $_;
_debug_resize $test_interval, $to, "$base -> $to (+ $_)";
}
say "";
}
}
undef;
STREAM (STDOUT):
Testing interval resizing:
Testing interval chr11:4..8 with length 5
-----
Interval: chr11:4..8 -> chr11:4..8, length 5 : 5 -> 5 (+ 0)
Interval: chr11:4..8 -> chr11:3..8, length 6 : 5 -> 6 (+ 1)
Interval: chr11:4..8 -> chr11:3..9, length 7 : 5 -> 7 (+ 2)
Interval: chr11:4..8 -> chr11:2..9, length 8 : 5 -> 8 (+ 3)
Interval: chr11:4..8 -> chr11:2..10, length 9 : 5 -> 9 (+ 4)
Interval: chr11:4..8 -> chr11:1..10, length 10 : 5 -> 10 (+ 5)
Testing interval chr11:5..8 with length 4
-----
Interval: chr11:5..8 -> chr11:5..8, length 4 : 4 -> 4 (+ 0)
Interval: chr11:5..8 -> chr11:5..9, length 5 : 4 -> 5 (+ 1)
Interval: chr11:5..8 -> chr11:4..9, length 6 : 4 -> 6 (+ 2)
Interval: chr11:5..8 -> chr11:4..10, length 7 : 4 -> 7 (+ 3)
Interval: chr11:5..8 -> chr11:3..10, length 8 : 4 -> 8 (+ 4)
Interval: chr11:5..8 -> chr11:3..11, length 9 : 4 -> 9 (+ 5)
Testing interval chr11:5..9 with length 5
-----
Interval: chr11:5..9 -> chr11:5..9, length 5 : 5 -> 5 (+ 0)
Interval: chr11:5..9 -> chr11:4..9, length 6 : 5 -> 6 (+ 1)
Interval: chr11:5..9 -> chr11:4..10, length 7 : 5 -> 7 (+ 2)
Interval: chr11:5..9 -> chr11:3..10, length 8 : 5 -> 8 (+ 3)
Interval: chr11:5..9 -> chr11:3..11, length 9 : 5 -> 9 (+ 4)
Interval: chr11:5..9 -> chr11:2..11, length 10 : 5 -> 10 (+ 5)
Testing interval chr11:6..9 with length 4
-----
Interval: chr11:6..9 -> chr11:6..9, length 4 : 4 -> 4 (+ 0)
Interval: chr11:6..9 -> chr11:6..10, length 5 : 4 -> 5 (+ 1)
Interval: chr11:6..9 -> chr11:5..10, length 6 : 4 -> 6 (+ 2)
Interval: chr11:6..9 -> chr11:5..11, length 7 : 4 -> 7 (+ 3)
Interval: chr11:6..9 -> chr11:4..11, length 8 : 4 -> 8 (+ 4)
Interval: chr11:6..9 -> chr11:4..12, length 9 : 4 -> 9 (+ 5)
use Bio::DB::HTS::Faidx;
my $hg_db = Bio::DB::HTS::Faidx->new( $hg_bgz_path );
sub extract_sequence {
my ($db, $interval) = @_;
my $chrom_length = $db->length($interval->seq_id);
my $trimmed_interval = $interval->clone;
$trimmed_interval->start( List::Util::max( $interval->start, 1 ) );
$trimmed_interval->end( List::Util::min( $interval->end , $chrom_length ) );
# Bio::DB::HTS::Faidx is 0-based for both start and end points
my $seq = $db->get_sequence2_no_length(
$trimmed_interval->seq_id,
$trimmed_interval->start - 1,
$trimmed_interval->end - 1,
);
my $pad_upstream = 'N' x List::Util::max( -($interval->start-1), 0 );
my $pad_downstream = 'N' x List::Util::max( $interval->end - $chrom_length, 0 );
return join '', $pad_upstream, $seq, $pad_downstream;
}
sub seq_info {
my ($seq, $n) = @_;
$n ||= 10;
if( length $seq > $n ) {
sprintf "%s...%s (length %d)", uc substr($seq, 0, $n), uc substr($seq, -$n), length $seq;
} else {
sprintf "%s (length %d)", uc $seq, length $seq;
}
}
####
{
say "Testing sequence extraction:";
say "1 base: ", seq_info
extract_sequence( $hg_db,
Interval->new( -seq_id => 'chr11',
-start => 35_082_742 + 1,
-end => 35_082_742 + 1 ) );
say "3 bases: ", seq_info
extract_sequence( $hg_db,
Interval->new( -seq_id => 'chr11',
-start => 1,
-end => 1 )->resize(3) );
say "5 bases: ", seq_info
extract_sequence( $hg_db,
Interval->new( -seq_id => 'chr11',
-start => $hg_db->length('chr11'),
-end => $hg_db->length('chr11') )->resize(5) );
say "chr11 is of length ", $hg_db->length('chr11');
say "chr11 bases: ", seq_info
extract_sequence( $hg_db,
Interval->new( -seq_id => 'chr11',
-start => 1,
-end => $hg_db->length('chr11') )->resize( $hg_db->length('chr11') ) );
}
STREAM (STDOUT):
Testing sequence extraction:
1 base: G (length 1)
3 bases: NNN (length 3)
5 bases: NNNNN (length 5)
chr11 is of length 135086622
chr11 bases: NNNNNNNNNN...NNNNNNNNNN (length 135086622)
RESULT:
1
Now we can use the same target interval that is used in the example notebook which recreates part of figure 1 from the Enformer paper.
my $target_interval = Interval->new( -seq_id => 'chr11',
-start => 35_082_742 + 1, # BioPerl is 1-based
-end => 35_197_430 );
say "Target interval: $target_interval with length @{[ $target_interval->length ]}";
die "Target interval is not $model_central_base_pairs_length bp long"
unless $target_interval->length == $model_central_base_pairs_length;
say "Target sequence is ", seq_info extract_sequence( $hg_db, $target_interval );
say "";
my $resized_interval = $target_interval->resize( $model_sequence_length );
say "Resized interval: $resized_interval with length @{[ $resized_interval->length ]}";
die "resize() is not working properly!" unless $resized_interval->length == $model_sequence_length;
my $seq = extract_sequence( $hg_db, $resized_interval );
say "Resized sequence is ", seq_info($seq);
STREAM (STDOUT):
Target interval: chr11:35082743..35197430 with length 114688
Target sequence is GGTGGCAGCC...ATCTCCTTTT (length 114688)
Resized interval: chr11:34943479..35336694 with length 393216
Resized sequence is ACTAGTTCTA...GGCCCAAATC (length 393216)
RESULT:
1
To prepare the input we have to one-hot encode this resized sequence and give it a dummy dimension at the end to indicate that it is is a batch with a single sequence. Then we can turn the PDL ndarray into a TFTensor
and pass it to our prediction function.
my $sequence_one_hot = one_hot_dna( $seq )->dummy(-1);
say $sequence_one_hot->info; undef;
STREAM (STDOUT):
PDL: Float D [4,393216,1]
use Devel::Timer;
my $t = Devel::Timer->new;
$t->mark('prediction of sequence');
my $predictions = $predict_on_batch->( $session, FloatPDLTOTFTensor( $sequence_one_hot ) );
$t->mark('End of prediction of sequence');
p $predictions;
$t->mark('END');
$t->report();
STREAM (STDERR):
AI::TensorFlow::Libtensorflow::Tensor {
Type FLOAT
Dims [ 1 896 5313 ]
NumDims 3
ElementCount 4760448
}
Devel::Timer Report -- Total time: 14.5641 secs
Interval Time Percent
----------------------------------------------
01 -> 02 14.5634 100.00% prediction of sequence -> End of prediction of sequence
02 -> 03 0.0007 0.00% End of prediction of sequence -> END
00 -> 01 0.0000 0.00% INIT -> prediction of sequence
Now we turn the TFTensor
output into a PDL ndarray.
my $predictions_p = FloatTFTensorToPDL($predictions)->slice(',,(0)');
say $predictions_p->info; undef;
STREAM (STDOUT):
PDL: Float D [5313,896]
Plot predicted tracks
These predictions can be plotted
my @tracks = (
[ 'DNASE:CD14-positive monocyte female' => 41 => $predictions_p->slice('(41)') ],
[ 'DNASE:keratinocyte female' => 42 => $predictions_p->slice('(42)') ],
[ 'CHIP:H3K27ac:keratinocyte female' => 706 => $predictions_p->slice('(706)')],
[ 'CAGE:Keratinocyte - epidermal' => 4799 => log10(1 + $predictions_p->slice('(4799)')) ],
);
use PDL::Graphics::Gnuplot;
my $plot_output_path = 'enformer-target-interval-tracks.png';
my $gp = gpwin('pngcairo', font => ",10", output => $plot_output_path, size => [10,2. * @tracks], aa => 2 );
$gp->multiplot( layout => [1, scalar @tracks], title => $target_interval );
$gp->options(
offsets => [ graph => "0.01, 0, 0, 0" ],
lmargin => "at screen 0.05",
);
my $x = zeroes($predictions_p->dim(1))->xlinvals($target_interval->start, $target_interval->end);
my @tics_opts = (mirror => 0, out => 1);
for my $i (0..$#tracks) {
my ($title, $id, $y) = @{$tracks[$i]};
$gp->plot( {
title => $title,
border => [2],
ytics => { @tics_opts, locations => [ ceil(($y->max-$y->min)/2)->sclr ] },
( $i == $#tracks
? ( xtics => { format => '%.3f', @tics_opts } )
: ( xtics => 0 ) ),
( $i == $#tracks ? ( xlabel => 'location ({/Symbol \264}10^7 bases)' ) : () ),
},
with => 'filledcurves',
#'lc' => '#086eb5',
# $x scaled by 1e7; filled curve between $y and the x-axis
$x / 1e7, $y, pdl(0)
);
}
$gp->end_multi;
$gp->close;
if( IN_IPERL ) {
IPerl->png( bytestream => path($plot_output_path)->slurp_raw );
}
DISPLAY:
Parts of the original notebook that fall outside the scope
In the orignal notebook, there are several more steps that have not been ported here:
"Compute contribution scores":
This task requires implementing
@tf.function
to compile gradients."Predict the effect of a genetic variant" and "Score multiple variants":
The first task is possible, but the second task requires loading a pre-processing pipeline for scikit-learn and unfortunately this pipeline is stored as a pickle file that is valid for an older version of scikit-learn (version 0.23.2) and as such its behaviour is tied to that version.
# Some code that could be used for working with variants.
1 if <<'COMMENT';
use Bio::DB::HTS::VCF;
my $clinvar_tbi_path = "${clinvar_path}.tbi";
unless( -f $clinvar_tbi_path ) {
system( qw(tabix), $clinvar_path );
}
my $v = Bio::DB::HTS::VCF->new( filename => $clinvar_path );
$v->num_variants
COMMENT
undef;
RESOURCE USAGE
use Filesys::DiskUsage qw/du/;
my $total = du( { 'human-readable' => 1, dereference => 1 },
$model_archive_path, $model_base, $new_model_base,
$targets_path,
$hg_gz_path,
$hg_bgz_path, $hg_bgz_fai_path,
$clinvar_path,
$plot_output_path,
);
say "Disk space usage: $total"; undef;
STREAM (STDOUT):
Disk space usage: 4.66G
CPANFILE
requires 'AI::TensorFlow::Libtensorflow';
requires 'AI::TensorFlow::Libtensorflow::DataType';
requires 'Archive::Extract';
requires 'Bio::DB::HTS::Faidx';
requires 'Bio::Location::Simple';
requires 'Bio::Tools::Run::Samtools';
requires 'Data::Frame';
requires 'Data::Printer';
requires 'Data::Printer::Filter::PDL';
requires 'Devel::Timer';
requires 'Digest::file';
requires 'FFI::Platypus::Buffer';
requires 'FFI::Platypus::Memory';
requires 'File::Which';
requires 'Filesys::DiskUsage';
requires 'HTTP::Tiny';
requires 'IPC::Run';
requires 'List::Util';
requires 'PDL';
requires 'PDL::Graphics::Gnuplot';
requires 'Path::Tiny';
requires 'Syntax::Construct';
requires 'Text::Table::Tiny';
requires 'URI';
requires 'constant';
requires 'feature';
requires 'lib::projectroot';
requires 'overload';
requires 'parent';
requires 'strict';
requires 'utf8';
requires 'warnings';
AUTHOR
Zakariyya Mughal <zmughal@cpan.org>
COPYRIGHT AND LICENSE
This software is Copyright (c) 2022-2023 by Auto-Parallel Technologies, Inc.
This is free software, licensed under:
The Apache License, Version 2.0, January 2004