Introduction
Hidden Markov models (HMMs) underlie many of the most important tasks
in computational biology, including multiple sequence alignment, genome
annotation, and increasingly, sequence database searching. Originally
developed for speech recognition algorithms, their application to the
field of molecular biology has increased dramatically since advances in
computational capacity have enabled full probabilistic analysis in place
of heuristic approximations. Pioneering this transition are two groups
lead by Anders Krogh and Sean Eddy, whose respective software packages
SAM and HMMER have underpinned
HMM-based bioinformatic analysis for over two decades.
Here, we present the aphid package for analysis with
profile hidden Markov models in the R environment (R Core Team 2017). The package contains
functions for developing, plotting, importing and exporting both
standard and profile HMMs, as well as implementations of the forward,
backward and Viterbi algorithms for computing full and optimal
conditional sequence probabilities. The package also features a multiple
sequence alignment tool that produces high quality alignments
via profile HMM training.
The ‘aphid’ package
###Dependencies The aphid package is designed to
work in conjunction with the “DNAbin” and “AAbin” object types produced
using the ape package (Paradis,
Claude, and Strimmer 2004; Paradis 2012). These object types, in
which sequences are represented in a bit-level coding scheme, are
preferred over standard character-type sequences for maximizing memory
and speed efficiency. While we recommend using ape
alongside aphid, it is not a requisite and as such is
listed in the “Suggests” rather than “Imports” section of the package
description. Indeed, any sequence of standard ASCII characters is
supported, making aphid suitable for other applications
outside of biological sequence analysis. However, it should be noted
that if DNA and/or amino acid sequences are input as character vectors,
the functions may not recognize the ambiguity codes and therefore are
not guaranteed to treat them appropriately.
To maximize speed, the low-level dynamic programming functions
(including the forward, backward, Viterbi, and maximum a
posteriori algorithms) are written in C++ linking to the Rcpp
package (Eddelbuettel and Francois 2011).
R versions of these functions are also maintained for the purposes of
debugging, experimentation and code interpretation. This package also
relies on the openssl package (Ooms 2016) for sequence and alignment
comparisons using the MD5 hash algorithm.
###Classes Two primary object classes, “HMM” (hidden Markov model)
and “PHMM” (profile hidden Markov model) are generated using the
aphid functions deriveHMM and derivePHMM, respectively.
These objects are lists consisting of emission and transition
probability matrices (elements named “E” and “A”), and non-mandatory
elements that may include vectors of background emission and transition
probabilities (“qe” and “qa”, respectively) and other model metadata
including “name”, “description”, “size” (the number of modules in the
model), and “alphabet” (the set of symbols/residues emitted by the
model). Objects of class “DPA” (dynamic programming array) are also
generated by the Viterbi and forward/backward functions. These are
predominantly created for the purposes of succinct console-printing.
###Functions HMMs and PHMMs are explained in more detail throughout
the following sections using aphid package functions to
demonstrate their utility. The examples are borrowed from Durbin et al
(1998), to which users are encouraged to
refer for a more in-depth explanation on the theory and application of
these models. Book chapter numbers are provided wherever possible for
ease of reference.
Hidden Markov Models
A hidden Markov model is a probabilistic data-generating mechanism
for a sequence or set of sequences. It is depicted by a network of
states each emitting symbols from a finite alphabet
according to a set of emission probabilities, whose values are
specific to each state. The states are traversed by an interconnecting
set of transition probabilities, that include the probability
of remaining in any given state and those of transitioning to each of
the other connected states.
An example of a simple HMM is given in Durbin et al (1998) chapter 3.2. An imaginary casino has two
dice, one fair and one weighted. The fair dice emits residues from the
alphabet {1, 2, 3, 4, 5, 6} with equal probabilities (1/6 for each
residue). The probability of rolling a “6” with the loaded dice is 0.5,
while that of each of the other five residues is 0.1. If the dealer has
the fair dice, he may secretly switch to the loaded dice with a
probability of 0.05 after each roll, leaving a 95% chance that he will
retain the fair dice. Alternatively, if he has the loaded dice, he will
switch back to the fair dice with a probability of 0.1, or more likely,
retain the loaded dice with a probability of 0.9.
This example can be represented by a simple two-state hidden Markov
model. The following code manually builds and plots the “HMM”
object.
library("aphid")
states <- c("Begin", "Fair", "Loaded")
residues <- paste(1:6)
### Define transition probability matrix A
A <- matrix(c(0, 0, 0, 0.99, 0.95, 0.1, 0.01, 0.05, 0.9), nrow = 3)
dimnames(A) <- list(from = states, to = states)
### Define emission probability matrix E
E <- matrix(c(rep(1/6, 6), rep(1/10, 5), 1/2), nrow = 2, byrow = TRUE)
dimnames(E) <- list(states = states[-1], residues = residues)
### Create the HMM object
x <- structure(list(A = A, E = E), class = "HMM")
### Plot the model
plot(x, textexp = 1.5)
### Optionally add the transition probabilities as text
text(x = 0.02, y = 0.5, labels = "0.95")
text(x = 0.51, y = 0.5, labels = "0.90")
text(x = 0.5, y = 0.9, labels = "0.05")
text(x = 0.5, y = 0.1, labels = "0.10")
Figure 1: A simple hidden Markov model for the dishonest casino
example. The plot.HMM method depicts the transition
probabilities as weighted lines, and emission probabilities as
horizontal grey bars. No begin/end state is modeled in this example;
however, this can be achieved by entering non-zero probabilities in the
first row and column of the transition matrix and passing “begin = TRUE”
to plot.HMM
.
For a sequence of observed rolls, we can establish the most likely
sequence of hidden states (including when the dice-switching most likely
occurred) using the Viterbi algorithm. In the example given in Durbin et
al (1998) chapter 3.2, the observed
sequence of 300 rolls is:
## 31511624644664424531132163116415213362514454363165
## 66265666666511664531326512456366646316366631623264
## 55236266666625151631222555441666566563564324364131
## 51346514635341112641462625335636616366646623253441
## 36616611632525624622552652522664353533362331216253
## 64414432335163243633665562566662632666612355245242
Some observable clusters of 6’s suggest that the loaded dice made an
appearance at some stage, but when did the dice-switching occur? In the
following code, the Viterbi algorithm is used to find the most likely
sequence of hidden states given the model.
data(casino)
### The actual path is stored in the names attribute of the sequence
actual <- c("F", "L")[match(names(casino), c("Fair", "Loaded"))]
### Find the predicted path
vit1 <- Viterbi(x, casino)
predicted <- c("F", "L")[vit1$path + 1]
### Note the path element of the output Viterbi object is an integer vector
### the addition of 1 to the path converts from C/C++ to R's indexing style
Comparing the predicted path with the actual hidden sequence, the
Viterbi algorithm wasn’t far off:
## Actual FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFLLLLL
## Predicted FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFLL
##
## Actual LLLLLLLLLLLLLLLLFFFFFFFFFFFFLLLLLLLLLLLLLLLLFFFLLL
## Predicted LLLLLLLLLLLLLLLLFFFFFFFFFFFFLLLLLLLLLLLLLLLLLLLLLL
##
## Actual LLLLLLLLLLLFFFFFFFFFFFFFFFFFLLLLLLLLLLLLLFFFFFFFFF
## Predicted LLLLLLLLLLLLFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF
##
## Actual FFFFFFFFFFFFFFFFFFFFFFFFFFFFLLLLLLLLLLFFFFFFFFFFFF
## Predicted FFFFFFFFFFFFFFFFFFFFFFFFFFFFFLLLLLLLLLLLLLFFFFFFFF
##
## Actual FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF
## Predicted FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF
##
## Actual FFFFFFFFFFFFFFFFFLLLLLLLLLLLLLLLLLLLLLLFFFFFFFFFFF
## Predicted FFFFFFFFFFFFFFFFFFFFLLLLLLLLLLLLLLLLLLLFFFFFFFFFFF
We can also calculate the full and posterior probabilities of the
sequence given the model using the forward
and/or
backward
algorithms:
casino.post <- posterior(x, casino)
plot(1:300, seq(0, 1, length.out = 300), type = "n", xlab = "Roll number",
ylab = "Posterior probability of dice being fair")
starts <- which(c("L", actual) == "F" & c(actual, "F") == "L")
ends <- which(c("F", actual) == "L" & c(actual, "L") == "F") - 1
for(i in 1:6) rect(starts[i], 0, ends[i], 1, col = "grey", border = NA)
lines(1:300, casino.post[1, ])
Figure 2: Posterior state probabilities for the 300 dice
rolls. The line shows the posterior probability that the dice
was fair at each roll, while the grey rectangles show the actual periods
for which the loaded dice was being used. See Durbin et al (1998) chapter 3.2 for more details.
Deriving HMMs from sequence data
The aphid package features the function
deriveHMM
for building an HMM from a set of training
sequences. The following code derives a simple HMM from our single
sequence of dice rolls with its known state path (stored as the ‘names’
attribute of the sequence).
y <- deriveHMM(list(casino), logspace = FALSE)
plot(y, textexp = 1.5)
### Optionally add the transition probabilities as text
text(x = 0.02, y = 0.5, labels = round(y$A["Fair", "Fair"], 2))
text(x = 0.51, y = 0.5, labels = round(y$A["Loaded", "Loaded"], 2))
text(x = 0.5, y = 0.9, labels = round(y$A["Fair", "Loaded"], 2))
text(x = 0.5, y = 0.1, labels = round(y$A["Loaded", "Fair"], 2))
Figure 3: A simple HMM derived from the sequence of 300 dice
rolls. As in Fig. 1, transition probabilities are shown as
weighted lines and emission probabilities as horizontal grey bars.
This appears to be fairly close to the actual model, despite the fact
that the training data consisted of just a single sequence. One would
typically derive an HMM from a list of many such sequences (hence why
the input argument is a list and not a vector) but this example is
simplified for clarity.
Profile Hidden Markov Models
A profile hidden Markov model is an extension of a standard HMM,
where the emission and transition probabilities are position
specific. That is, they can change at each point along the
sequence. These models typically have many more parameters than their
simpler HMM counterparts, but can be very powerful for sequence
analysis. The precursor to a profile HMM is normally a multiple sequence
alignment. Each column in the alignment will often (but not always) be
represented by one internal position or “module” in the model, with each
module consisting of three states:
- a silent delete state that does not emit residues.
- an insert state with emission probabilities reflecting the
background residue frequencies averaged over the entire alignment.
- a match state with emission probabilities reflecting the
residue frequencies in the alignment column.
Figure 4 shows the three state types listed above as circles diamonds
and rectangles, respectively. The states are linked by transition
probabilities shown as weighted lines in the graph.
Consider this small partial alignment of amino acid sequences from
Durbin et al (1998) chapter 5.3:
## [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10]
## HBA_HUMAN "V" "G" "A" "-" "-" "H" "A" "G" "E" "Y"
## HBB_HUMAN "V" "-" "-" "-" "-" "N" "V" "D" "E" "V"
## MYG_PHYCA "V" "E" "A" "-" "-" "D" "V" "A" "G" "H"
## GLB3_CHITP "V" "K" "G" "-" "-" "-" "-" "-" "-" "D"
## GLB5_PETMA "V" "Y" "S" "-" "-" "T" "Y" "E" "T" "S"
## LGB2_LUPLU "F" "N" "A" "-" "-" "N" "I" "P" "K" "H"
## GLB1_GLYDI "I" "A" "G" "A" "D" "N" "G" "A" "G" "V"
Position-specific patterns include a high probability of observing a
“V” at position 1 and an “A” or “G” at position 3. When tabulating the
frequencies it is also prudent to add pseudo-counts, since the absence
of a particular transition or emission type does not preclude the
possibility of it occurring in another (unobserved) sequence.
Pseudo-counts can be Laplacean (adds one of each emission and transition
type), background (adjusts the Laplacean pseudo-counts to reflect the
background frequencies derived from the entire alignment), or user
defined, which can include more complex pseudo-count schemes such as
Dirichlet mixtures (Durbin et al.
1998).
The default option for the derivePHMM function is “background”.
The following code derives a profile HMM from the globin data and
plots the model:
globins.PHMM <- derivePHMM(globins, residues = "AMINO", pseudocounts = "Laplace")
## Applying uniform sequence weights
Figure 4: Profile HMM derived from a partial globin sequence
alignment. Match states are shown as rectangles, insert states
as diamonds, and delete states as circles. Grey horizontal bars
represent the emission probabilities for each residue in the alphabet
(in this case the amino acid alphabet) at each position in the model.
Numbers in the delete states are simply model module numbers, while
those in the insert states are the probabilities of remaining in the
current insert state at the next emission cycle. Lines are weighted and
directed where necessary to reflect the transition probabilities between
states. The large “B” and “E” labels represent are the silent begin and
end states, respectively.
Note that there are only 8 internal modules (excluding the begin and
end states), while the alignment had 10 columns. The
derivePHMM
function decided (using the maximum a
posteriori algorithm) that there was not enough residue information
in columns 4 and 5 of the alignment to warrant assigning them internal
modules in the model. Instead, the last sequence in the alignment
(GLB1_GLYDI) was considered to have entered the insert state at position
3 where it remained for two emission cycles (emitting an “A” and a “D”)
before transitioning to the match state in module 4. We can show this by
calculating the optimal path of that sequence through the model, again
using the Viterbi algorithm:
path <- Viterbi(globins.PHMM, globins["GLB1_GLYDI", ])$path
path
## [1] 1 1 1 2 2 1 1 1 1 1
The “path” element of the Viterbi object is an integer vector with
elements taking values 0 (“delete”), 1 (“match”) or 2 (“insert”). The
path can be expressed more intuitively as characters instead of indices
as follows:
c("D", "M", "I")[path + 1]
## [1] "M" "M" "M" "I" "I" "M" "M" "M" "M" "M"
Note that the addition of 1 to each path element is simply to convert
from the C/C++ indexing style (which begins at 0) to R’s style.
Sequences do not need to be aligned to produce a profile HMM. The
function derivePHMM
can optionally take a list of unaligned
sequences, in which case the longest sequence is used as a ‘seed’ to
create a preliminary profile HMM, and the model is iteratively trained
with the sequence list using either the Baum Welch or Viterbi training
algorithm (see model training section below).
File I/O
Profile HMMs can be exported as text files in the HMMER v3 format (http://www.hmmer.org/)
using the function writePHMM
. For example, the small globin
profile HMM can be exported by running
writePHMM(globins.PHMM)
. Similarly, a HMMER v3 text file
can be parsed into R as an object of class “PHMM” with the function
readPHMM
.
Sequence Simulation
To simulate data with random variation, the aphid
package features the function generate
with methods for
both HMMs and PHMM objects. Sequences are generated recursively using
the transition and emission probabilities from within the model. There
are two compulsory arguments, a model (object class “HMM” or “PHMM”) and
the “size” argument, which specifies the maximum length of the sequence
(this prevent an overflow situation that can occur if insert-insert
transition probabilities are relatively high). For example, the
following code simulates a list of 10 random sequences from the small
globin profile HMM:
sim <- list(length = 10)
suppressWarnings(RNGversion("3.5.0"))
set.seed(9999)
for(i in 1:10) sim[[i]] <- generate(globins.PHMM, size = 20)
sim
## $length
## I I I M I I M M D D M M I M
## "Y" "A" "I" "S" "K" "V" "R" "Y" "-" "-" "D" "A" "D" "S"
##
## [[2]]
## M M M M M D M M
## "H" "E" "H" "A" "I" "-" "S" "E"
##
## [[3]]
## M M M I I M M D D M
## "Y" "K" "F" "G" "V" "V" "I" "-" "-" "Y"
##
## [[4]]
## M M I I I M M M M M M
## "C" "M" "E" "D" "S" "A" "S" "G" "M" "G" "K"
##
## [[5]]
## M D D D D D D M
## "V" "-" "-" "-" "-" "-" "-" "S"
##
## [[6]]
## M M I M D D D D D
## "F" "N" "I" "M" "-" "-" "-" "-" "-"
##
## [[7]]
## M M M I I I I M M M M M
## "G" "M" "N" "C" "G" "V" "A" "N" "N" "P" "H" "Y"
##
## [[8]]
## M I I M M M I M M M M
## "V" "H" "N" "A" "K" "A" "K" "I" "E" "A" "S"
##
## [[9]]
## M D D D D D D M
## "Y" "-" "-" "-" "-" "-" "-" "L"
##
## [[10]]
## M M M D D M M M
## "D" "L" "G" "-" "-" "I" "C" "H"
Note that the names attributes specify which state each residue was
emitted from, and gap symbols are emitted from delete states. If these
gaps are not required they can be removed as follows:
sim <- lapply(sim, function(s) s[names(s) != "D"])
Model Training
The aphid package offers the function
train
for optimizing model parameters using either the Baum
Welch or Viterbi training algorithm. Both are iterative refinement
algorithms; the former does not rely on a multiple sequence alignment
but is generally much slower than the latter. The Viterbi training
operation can be sped up further by specifying the “cores” argument for
parallel processing. The best choice of training algorithm will
generally depend on the nature of the problem and the computing
resources available. For more information see Durbin et al (1998) chapter 3.3 for standard HMMs and chapter
6.5 for profile HMMs.
The following code trains the small globin profile HMM with the
sequences simulated in the previous step using the Baum Welch
algorithm.
globins2.PHMM <- train(globins.PHMM, sim, method = "BaumWelch",
deltaLL = 0.01, seqweights = NULL)
## Iteration 1 log likelihood = -244.2442
## Iteration 2 log likelihood = -226.6468
## Iteration 3 log likelihood = -220.4567
## Iteration 4 log likelihood = -216.7471
## Iteration 5 log likelihood = -214.2664
## Iteration 6 log likelihood = -212.5519
## Iteration 7 log likelihood = -211.3709
## Iteration 8 log likelihood = -210.5676
## Iteration 9 log likelihood = -210.0423
## Iteration 10 log likelihood = -209.7105
## Iteration 11 log likelihood = -209.4912
## Iteration 12 log likelihood = -209.3357
## Iteration 13 log likelihood = -209.2219
## Iteration 14 log likelihood = -209.1379
## Iteration 15 log likelihood = -209.0752
## Iteration 16 log likelihood = -209.0272
## Iteration 17 log likelihood = -208.9891
## Iteration 18 log likelihood = -208.9575
## Iteration 19 log likelihood = -208.9305
## Iteration 20 log likelihood = -208.9068
## Iteration 21 log likelihood = -208.8859
## Iteration 22 log likelihood = -208.8672
## Iteration 23 log likelihood = -208.8506
## Iteration 24 log likelihood = -208.836
## Iteration 25 log likelihood = -208.8231
## Iteration 26 log likelihood = -208.8118
## Iteration 27 log likelihood = -208.8021
## Convergence threshold reached after 27 EM iterations
As shown in the feedback (which can be switched off by setting “quiet
= TRUE”), this operation took 7 expectation-maximization iterations to
converge to the specified delta log-likelihood threshold of 0.01.
Sequence Alignment
The aphid package can be used to produce
high-quality multiple sequence alignments using the iterative model
training method outlined above. The function align
takes as
its primary argument a list of sequences either as a “DNAbin” object, an
“AAbin” object, or a list of character sequences. An object of class
“PHMM” can be passed to the function as an optional secondary argument
(“model”), in which case the sequences are simply aligned to the model
to produce the alignment matrix. If “model” is NULL, a preliminary model
is first derived using the ‘seed’ sequence method outlined above, after
which the model is trained using either the Baum Welch or Viterbi
training algorithm (specified via the “method” argument). The
sequences are then aligned to the model in the usual fashion to produce
the alignment. Note that if only two sequences are present inthe input
list, the align
function will perform a pairwise alignment
without a profile HMM (Smith-Waterman or Needleman-Wunch alignment).
In this final example, we will deconstruct the original globin
alignment and re-align the sequences using the original PHMM as a
guide.
globins <- unalign(globins)
align(globins, model = globins.PHMM, seqweights = NULL, residues = "AMINO")
## 1 2 3 I I 4 5 6 7 8
## HBA_HUMAN "V" "G" "A" "-" "-" "H" "A" "G" "E" "Y"
## HBB_HUMAN "V" "-" "-" "-" "-" "N" "V" "D" "E" "V"
## MYG_PHYCA "V" "E" "A" "-" "-" "D" "V" "A" "G" "H"
## GLB3_CHITP "V" "K" "G" "-" "-" "-" "-" "-" "-" "D"
## GLB5_PETMA "V" "Y" "S" "-" "-" "T" "Y" "E" "T" "S"
## LGB2_LUPLU "F" "N" "A" "-" "-" "N" "I" "P" "K" "H"
## GLB1_GLYDI "I" "A" "G" "A" "D" "N" "G" "A" "G" "V"
Note that the column names show the progressive positions along the
model and where residues were predicted to have been emitted by insert
states (e.g. the 4th and 5th residues of sequence 7).
These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.