HIV- -1 tropism prediction 1 tropism prediction HIV Mattia CF - - PowerPoint PPT Presentation

Mattia CF Prosperi ahnven@yahoo.it University of “Roma TRE” Faculty of Computer Science Engineering Dept of Computer Science and Automation (DIA) via della vasca navale, 79 – 00149 – Rome, ITALY

HIV HIV-

• 1 tropism prediction

1 tropism prediction

Summary

• State of the art

– From charge rule to structural descriptors

• Roma TRE modelling

– Data collection

• Sequence manipulation
• Enhanced domain coding

– Univariable analysis and clustering – Model technologies

• Logistic regression and feature selection
• Validation and comparison with other models
• Interpretation of relevant features

State of the art

• Charge rule (De Jong, 1992)
• Neural Networks, Decision Trees, Support Vector Machines

(Resch, Pillai) on 200-300 examples

• Position Specific Scoring Matrices (Jensen)
• Support Vector Machines (Sing) on 1’100 examples with AUC

maximisation adding CD4+ cell count as additional input variable

• Support Vector Machines + Structural Analysis (Sander, 2007) with

AUC maximisation

• Neural Networks for dual-tropism prediction (Lamers, 2008)
• All models work on the sole V3 loop

State of the art (2)

• SVM + Structural Analysis (Sander, 2007) seems to

be the best performing model at present

– 91.56% accuracy – 0.93 AUC – Minor critics concerning sample collection (all different sequences, regardless patient, without accounting for real sequence population distribution) – Improvements gained with the structural analysis, over a reference SVM trained only on the V3 dummy variable encoding

Roma TRE approach

• Data: collection of samples from “Los Alamos” data base

– Only one sequence per patient (the longest available, no clones) except for sequences with different tropism – No problematic sequences – All subtypes – At least V3 loop, possibly all envelope gene – Clinical markers recorded

• Goal: prediction of CXCR4 usage probability (regardless CCR5 usage, dual

tropic strains are pooled into X4 strains)

Sequence manipulation

• Previous works used multiple alignment

(clustalw or muscle) either on nucleotide

• r amino-acids
• We used local pairwise alignment (Smith-

Waterman-Gotoh) with ambiguities and frameshifts correction/detection against HXB2 strain (which is X4)

– Minor differences with the output of other models

Domain coding

• Binary dummy variables for specific amino

acidic changes (plus ins-del and “any” substitution) in the V3 loop and in the envelope

• Phisico-chemical coding for position

changes

• Subtype
• Clinical markers (HIV RNA load, CD4 and

CD8 cell counts)

Univariable analysis

• CD4+ are significantly associated with tropism

(low CD4+ → X4)

• Subtype B, D isolates are prevalently X4
• Subtype A, C, 02_AG isolates are prevalently R5

Univariable analysis

• Highly significant positions in

the V3 loop

• 306 (11)
• 302 (7)
• 303 (8)
• 323 (28)
• 301 (6)
• 313 (18)
• 321 (26)
• 322 (27)
• 300 (5)
• 315 (20)
• 320 (25)
• 307 (12)
• 316 (21)
• 325 (30)
• 304 (9)
• …
• A few positions outside the

V3 loop are significant, but slightly over the Benj-Hoch adjusted threshold (adj.p<0.1)

• 440, 192, 169

Hierachical Clustering

• Threshold of 0.35: {318A, ins317}, {311I, 308S, 306del, 307del}, {322I, 320

hydrophilic, 326I}

• mutations positively associated with X4 viruses tend to behave more

independently (306S, 303I, 308K, 300Y and 307T)

Machine Learning

• Logistic Regression (LR)
• Feature selection via filter and embedded methods

(univariable analysis, AIC selection, CFS, ridge shrinkage)

• Comparison with other (non-linear) machine learning

techniques

– SVM (same settings as Sander, 2007) – Random Forests and Decision Trees (RF, DT) – Rule Bases (RIPPER, JRIP) – Instance Based Reasoning (IBR)

• Multiple 10-fold cross validation for model performance

assessment and model comparison

– Student’s t-test adjusted (Bengio and Nadeau) for sample overlap and multiple comparisons over 10 independent runs

Results

• Logistic

Regression

– High accuracy (92.76%) and AUC (0.93) – Enhanced domain coding performs significantly better that naïve variable encoding and sole V3 loop – Equally performing as the reference SVM

Results (2)

Conclusions

• Logistic Regression is a powerful and

interpretable tool for tropism prediction

– Importance of envelope region analysis – Importance of enhanced variable encoding – Importance of feature selection techniques – Importance of robust validation and comparison statistics – We have a linear model: from the comparison analysis, non-linear models seem not to improve performances

• The modelling technique is also suitable for

combination with structure-based methods