Model-based species identification using DNA barcodes Bogdan Paşaniuc CSE Department, University of Connecticut Joint work with Ion Măndoiu and Sotirios Kentros

Outline Existing approaches to species identification Proposed statistical model based methods Experimental Results Ongoing Work and Conclusions

Background on DNA barcoding Recently proposed tool for species identification Use short DNA region as “fingerprint” for the species Region of choice: cytochrome c oxidase subunit 1 mitochondrial gene ("COI", 648 base pairs long). Key assumption: inter-species variability higher than intra-species variability

Species identification problem Given:  Database DB containing barcodes from known species  New barcode x Find:  a high confidence assignment to a species in the DB  UNKNOWN, if confidence not high enough Use additional evidence/methods to resolve UNKNOWN assignments and possible discovery of new species

Existing approaches and limitations Neighbor Joining tree for new + known barcodes [Meyers&Paulay05]  One barcode per species  Runtime does not scale well with #species (quadratic or worse) Likelihood ratio test for species membership using MCMC [Matz&Nielsen06]  Impractical runtime even for moderate #species Distance-based [BOLD-IDS, TaxI(Steinke et al.05)]  Unclear statistical significance

BOLD BOLD: The Barcode of Life Data Systems [Ratnasingham&Hebert07]   Currently: 28,129 species, 251,429 barcodes Identification System: BOLD-IDS  Distance-based (NJ tree for visualization)  Employs a threshold (less than 1% divergence) to get a tight match to a barcode in the DB

BOLD-IDS [Ekrem et al.07]: “ … identifications by the BOLD facility must be cautiously evaluated as the system at present may return high probabilities of placements that obviously are erroneous”

Outline Existing approaches to species identification Proposed statistical model based methods Experimental Results Ongoing Work and Conclusions

Bayesian approach to species identification Assign barcode x=x 1 x 2 x 3 …x n to species SP i that maximizes P(SP i |x) over all species SP i P(SP i |x) computed using Bayes’ theorem: P(SP|x) = P(x|SP)*P(SP)/P(x)  Uniform prior P(SP)  P(x) constant for fixed x  Need model for P(x|SP) We explored three scalable models: position weight matrices, Markov chains, hidden Markov models  Similar to models used successfully in other sequence analysis problems such as DNA motif finding and protein families

Positional weight matrix (PWM) Assumption: independence of loci  P(x|SP) = P(x 1 |SP)*P(x 2 |SP)*…*P(x n |SP) For each locus, P(x i |SP) is estimated as the probability of seeing each nucleotide at that locus in DB sequences from species SP

Inhomogeneous Markov Chain (IMC) Takes into account dependencies between consecutive loci  start A C T G A C T G A C T G A C T G … locus 1locus 2locus 3locus 4

Hidden Markov Model (HMM) Same structure as the IMC  Each state emits the associated DNA base with high probability; but can also emit the other bases with probability equal to mutation rate Barcode x generated along path p with probability equal to product of emission & transitions along p P(x|HMM) = sum of probabilities over all paths  Efficiently computed by forward algorithm

Accuracy on BOLD dataset 37 species with at least 100 barcodes from BOLD  10-50% barcodes removed and used for test IMC yields better accuracy in all cases 10%20%30%40%50% PWM90.08%90.01%90.02%89.68%89.69% IMC99.97%99.93%99.90%99.91%99.89% HMM99.57% 99.66%99.70%99.76%

Score normalization DB barcodes have non uniform lengths and cover different regions of the COI gene  Membership probabilities not always comparable Normalization scheme:  Species models constructed only over positions covered in DB  Scores normalized using background IMC constructed from all sequences in DB

Computing the confidence of assignment x assigned to species SP with score s p-value: probability that a barcode generated under background model Ḿ has a score s’  s Methods for p-value estimation:  Random sampling  Generate random sequences and count how many exceed the score  Exact computation (for PWMs):  Dynamic programming [Rahmann03]  Branch and bound [Zhang et. Al 07]  Shiffted FFTs [Nagarajan et al. 05]

Exact computation for PWMs [Rahmann03] Computes the entire distribution Scores rounded by a granularity factor Score is a sum of n independent variables (score contribution of each position)  Probability of a rand. seq. of length i having a score of computed from the contribution of first i-1 positions and current position

Exact computation for IMCs Defineas the prob. of a random seq of length i having score and last letter Basic recurrence:

IMC exact p-value computation Initially The probability of a random barcode having score Runtime, where R is the difference between max and min score for any i.

Outline Existing approaches to species identification Proposed statistical model based methods Experimental Results Ongoing Work and Conclusions

Experimental setup (1) Compared methods  IMC  Species with highest score  If score < species specific threshold  UNKNOWN  Distance-based (BOLD-IDS like)  Species containing barcode showing less divergence  If divergence > threshold (default 1%)  UNKNOWN Basic questions  What is the effect of training set size (#barcodes per species) on accuracy?  What is the effect of the #species on accuracy?

Experimental setup (2) Two scenarios:  Complete DB: all new barcodes belong to species in DB  Incomplete DB: some new barcodes belong to species not in DB

Accuracy measures True positive rate = TP/(TP+FP)  Barcodes belonging to species present in the DB  TP = #barcodes assigned to correct species  FP = #barcodes assigned to incorrect species  Barcodes belonging to species not present in DB  TP = #barcodes assigned to unknowns  FP = #barcodes assigned to species in the DB

Effect of #barcodes/species Datasets containing all BOLD species with at least 5/25 barcodes  BOLD5: 1508 sp, barcodes  BOLD25: 270 sp, barcodes DB composed of randomly picked 5-20 barcodes from all species in BOLD25 Test barcodes  Complete database scenario  All remaining barcodes from BOLD25  Incomplete database scenario  All barcodes from BOLD5 not in DB

Effect of #barcodes/species, complete DB

Effect of #barcodes/species, incomplete DB

Effect of #species Datasets containing all BOLD species with at least 5/10 barcodes  BOLD5: 1508 sp, barcodes  BOLD10: 690 sp, barcodes DB composed of randomly picked 100 to 690 species from BOLD10  10 barcodes per species Test barcodes  Complete database scenario  All remaining barcodes from picked species  Incomplete database scenario  All barcodes from BOLD5 not in DB

Effect of #species, complete DB

Effect of #species, incomplete DB

Outline Existing approaches to species identification Proposed statistical model based methods Experimental Results Ongoing Work and Conclusions

Conclusions & Ongoing work IMC provides a scalable method for species identification  High accuracy, with useful tradeoff between TP rate and unknown rate  Efficiently computable p-values Comprehensive comparison of identification algorithms to be submitted to 2 nd International Barcode Conference  Broad coverage of methods  tree-based, distance-based, character-based, model-based  Assessment of further effects besides #species and #barcodes/species  Barcode length  Barcode quality  Number of regions  Runtime scalability (up to millions of species)  Diverse datasets (BOLD, cowries, flu viruses, simulated data, etc.)