Brain-specific HIV Nef identified in multiple patients with neurological disease

Abstract

HIV-1 Nef is a flexible, multifunctional protein with several cellular targets that is required for pathogenicity of the virus. This protein maintains a high degree of genetic variation among intra- and inter-host isolates. HIV Nef is relevant to HIV-associated neurological diseases (HAND) in patients treated with combined anti-retroviral therapy because of the protein’s role in promoting survival and migration of infected brain macrophages. In this study, we analyzed 2,020 HIV Nef sequences derived from 22 different tissues and 31 subjects using a novel computational approach. This approach combines statistical regression and evolved neural networks (ENNs) to classify brain sequences based on the physical and chemical characteristics of functional Nef domains. Based on training, testing, and validation data, the method successfully classified brain Nef sequences at 84.5% and provided informative features for further examination. These included physicochemical features associated with the Src-homology-3 binding domain, the Nef loop (including the AP-2 Binding region), and a cytokine binding domain. Non-brain sequences from patients with HIV-associated neurological disease were frequently classified as brain, suggesting that the approach could indicate neurological-risk using blood-derived virus or for the development of biomarkers for use in assay systems aimed at drug efficacy studies for the treatment of HIV-associated neurological diseases.

Introduction

In the HIV-infected population, current combined antiretroviral therapy (cART) can reduce plasma viral loads (pVL) to undetectable levels and restore T-cell counts; however, the spectrum of HIV-associated neurocognitive disorders (HAND) remains a common co-morbidity. Multiple studies have suggested that the HIV Nef protein plays a role in the development of HAND by acting directly and indirectly on brain cells (Khan et al. 2016; Ranki et al. 1995; Acharjee et al. 2014; Saribas, Khalili, and Sariyer 2015; Ghiglione and Turk 2011; Gray et al. 2011b). Nef increases monocyte and macrophage migration in animal models when injected into the brain (Mordelet et al. 2004; Overholser et al. 2003), and modulates cell-signaling pathways (Herbein et al. 2010). Nef contributes to an inflammatory environment by stimulating the production of inflammatory cytokines by macrophages (such as MIP-1α, MIP-1β, TNF-α, IL-1β, and IL-6) (Herbein et al. 2010), and corresponds to activation of the inflammatory factor NF-κB (Overholser et al. 2003; Olivetta et al. 2003). Nef is directly toxic to astrocytes and neurons in vitro, possibly through indirect activity of IP-10 (van Marle et al. 2004), and/or by stimulating the neurotoxin quinolinic acid (Smith et al. 2001). Nef is found in astrocytes of patients with HIV-associated dementia (HAD) and HIV encephalitis (HIVE) (Ranki et al. 1995), and astrocytic expression of Nef impairs spatial and recognition memory and increases neuronal loss (Chompre et al. 2013).

Previously, we compared twelve HIV Nef 3-dimensional (3D) protein structures, generated through sequencing and structural minimization, from AIDS patients with or without severe HIV-associated dementia (HAD) (Lamers, Poon, and McGrath 2011). We found several structural differences within Nef that could alter SH3 binding, change the orientation of the internal core domain, and make available an additional cysteine residue in the variable loop domain, which could alter membrane- or protein-binding potential. We proposed that these changes could play a functional role in the development of HAD. Others have also identified Nef structural-functional relationships (Geyer, Fackler, and Peterlin 2001; Arold and Baur 2001). Interestingly, in another study, we found that host-specific nef DNA and RNA emerged prior to env, and that early (<90 days) convergent evolution in nef, but not env, was evident in brains of simian immunodeficiency virus SIVMac251-infected macaques (Lamers et al. 2015; Strickland et al. 2011). These findings led us to hypothesize that certain HIV nef isolates are better adapted to immune cells in the brain environment (Lewis et al. 2012; Jia et al. 2012; Olivieri et al. 2010; Gray et al. 2011a), which are presumably of myeloid lineage (Williams and Burdo 2012; Williams et al. 2001).

In the current study, we used a newly developed evolved neural network approach to determine if it was possible to classify brain and non-brain Nef sequences, and more specifically, to identify brain-specific Nef (Lamers et al. 2017; Lamers, Fogel, et al. 2016). The approach elucidated structure-function properties of brain Nef that were not discernable by viewing sequence information alone. The study resulted in a comprehensive investigation into the physiochemical properties of brain-adapted Nef.

Materials and Methods

Sequences and alignment

Sequences were downloaded from either GenBank or the HIV Database at Los Alamos (www.lanl.HIV.gov), and included 2,020 unique Nef sequences derived from the anatomical tissues of 31 subjects (Fig I; Supplemental Table I). All 31 subjects in the study had multiple sequences and 26 subjects had sequences from multiple tissues. For 23 subjects, the primary pathology of the subjects at death was known (other than HIV-1 infection) (Fig IA). No cause of death was associated with sequence data for 8 subjects. Sequences from 4 subjects were derived from a longitudinally studied cohort of donors who were cART adherent and who had an undetectable plasma viral load at death (C02*, C04*, C05* and C09*) (Lamers, Rose, et al. 2016). Most brain sequences available for the study were derived from HAD or HAND patients; however, there were 25 sequences generated from brain tissue of patient (AZ) who died from a brain hemorrhage due to severe atherosclerosis, and two sequences were derived from brain tissue of a patient with lung cancer (C05*). Only two subjects included in the study (CB1 and CB3) had sequences from anatomical tissues paired with peripheral blood mononuclear cells (PBMC)-derived sequences at death. Overall, 512 brain sequences and 1,359 non-brain sequences derived from other lymphoid tissues were available for study (Fig IB). We did not categorize a subset of sequences as “brain” or “non-brain” that were obtained from CSF, spinal cord, and meninges because they can share similarity with either brain or lymphoid tissues (Lamers et al. 2011; Salemi et al. 2009; Gonzalez-Perez et al. 2012). These sequences were not used for neural network training or testing but were used as a validation set (as further described below) (Fig IB). Nine molecular clone sequences were included in the study (ADA, CAM1, HCB2, JRCSF, JRFL, MANC and 3 sequences derived from BAL). All amino acid sequences were aligned using Geneious software (ver. 10) (www.geneious.com) followed by manual editing to correct for any obvious alignment errors.

Fig I. Number of nef sequences in study derived from anatomical tissues from subjects used in study.

Panel A Subject IDs are on the left and are sorted by primary disease pathology (if known) at death. The number of sequences for each subject are on the y-axis. Asterisks indicate subjects who died while on cART with no measurable plasma viral load. Bars are colored by specific tissue of origin for each sequence. Panel B. Sequences in Panel A are sorted by major categories used for training and testing (brain and non-brain) and validation (CSF, meninges, PBMC and spinal cord) of the ENN.

Feature generation and correlations

72 different features describing various physicochemical characteristics of amino acids were identified from the available literature and resources such as ProtScale and ProtParam (www.expasy.org) (Wilkins et al. 1999) (Table I). These features were grouped into six major classes: amino acid size, shape or structure (n=24), polarity (n=6), composition (n=5), hydrophobicity (n=26), and miscellaneous other features such as those associated with HPLC and pKa (n=8). For each translated sequence, all features were calculated for the complete Nef protein and 10 specific functional domains in separate analyses. The chosen domains were already known as important for proper or altered Nef function (Fig II). For example the myristoylation domain, known for its high degree of variation, is required for the association of Nef with cellular lipid membranes (Bentham, Mazaleyrat, and Harris 2006). The conserved polyproline motif within the MHC1-DM domain is responsible for a number of functions, including MHC-1 down-modulation and positions associated with SH3 binding (Kuo et al. 2012; Jung et al. 2011). Other structural motifs were also considered in the study, including two alpha helix motifs in the center of the structure and another region near the 3′ end; a flexible loop domain, flanked by two beta-sheets that is associated with AP2 binding; and a region associated with altered cytokine immune-system signaling responses (Percario et al. 2015). This effort generated a total set of 72 features calculated for 11 total regions, or 792 total features.

Table I.

Physicochemical and structural features used for the study.

Class	Features
A. Size, Shape, Structure	Molecular Weight (Gasteiger et al. 2005)	Bulkiness (Zimmerman, Eliezer, and Simha 1968)	Antiparallel Beta Strand (Lifson and Sander 1979)
	Beta Turn (Deleage and Roux 1987)	Beta-sheet Levitt (Levitt 1978)	Alpha Chou and Fasman (Chou and Fasman 1978)
	Coil (Deleage and Roux 1987)	Beta Strand (Lifson and Sander 1979)	Parallel Beta (Lifson and Sander 1979)
	Surface Area (Darby and Creighton 1993)	Volume (Darby and Creighton 1993)	Surface Exposure (Grantham 1974)
	Average Flexibility(Grantham 1974)	Mol. Fraction of Buried Res.(Janin 1979)	Beta Chou and Fasman (Chou and Fasman 1978)
	% Accessible Residues (Janin 1979)	2D Propensity (Lamers et al. 2008)	Transmembrane (Zhao and London 2006)
	Avg. Area Buried (Rose et al. 1985)	Membership Class (Grantham 1974)	Mass Membership Class (Lamers et al. 2008)
	Recognition Factors (Hofmann and Hadge 1987)	Alpha Helix Levitt (Levitt 1978)	Alpha Helix (Deleage and Roux 1987)
B. Polarity	Polarity (Grantham 1974)	Charge1	Charge Polarity (Zimmerman, Eliezer, and Simha 1968)
	Charge Scale (Grantham 1974)	Grantham (Grantham 1974)
C. Composition	Amino Acid Composition (McCaldon and Argos 1988)	Length of V3	Amino Acid Composition Swiss Prot (Gasteiger et al. 2005)
	Relative Mutability (Dayhoff, Schwartz, and Orcutt 1978)	Total Sequence Glycosylation (Van Baelen et al. 2007)
D. Hydrophobicity	Sweet et al. (Sweet and Eisenberg 1983)	Kyte and Doolittle (Kyte and Doolittle 1982)	Hydrophobicity (Grantham 1974)
	Abraham and Leo (Abraham and Leo 1987)	Bull and Breese (Bull and Breese 1974)	Guy (Guy 1985)
	Roseman (Roseman 1988)	Wolfen et al.(Wolfenden et al. 1981)	Wilson et al. (Wilson et al. 1981)
	Eisenberg et al. (Eisenberg et al. 1984)	Hopp and Woods (Hopp and Woods 1981)	Manvalan et al. (Manavalan and Ponnuswamy 1978)
	Fauchere et al. (Fauchere and Pliska 1983)	Janin (Janin 1979)	Rao and Argos (Mohana Rao and Argos 1986)
	Tanford (Tanford 1962)	Welling et al. (Welling et al. 1985)	Chothia (Chothia 1976)
	Cowan and Whittaker (Cowan and Whittaker 1990)	Parker (Parker, Guo, and Hodges 1986)	Browne (Browne, Bennett, and Solomon 1982)
	Meek (Meek 1980)	Aboderin(Aboderin 1971)	Rose et al. (Rose et al. 1985)
	Black and Mould (Black and Mould 1991)	Miyazawa et al. (Miyazawa and Jernigan 1996)
E. Local Features	Glycosylation at Positions 6–8 (Van Baelen et al. 2007) Charge at Position 30 (Milich, Margolin, and Swanstrom 1993)	Glycosylation at Positions 5,7,9 (Van Baelen et al. 2007)	Charge at Position 12 (Milich, Margolin, and Swanstrom 1993)
F. HPLC and Other	Retention at pH 2.1 (Meek 1980)	HPLC/TFA (Browne, Bennett, and Solomon 1982)	Refractivity (Jones 1975)
	HP Scale (Grantham 1974)	pKa alpha carboxylate (Anaspec 2013)	Exchange (Lamers et al. 2008)
	pKa Amine (Anaspec 2013)	pI at 25°C (Anaspec 2013)

¹“Charge” feature = [+1] for K and R; [−1] for D and E.

Fig II. Variation in Nef sequence alignment and location of functional domains studied.

Amino acid sequence variation is presented as a “SeqLogo,” which shows the major amino acids along the alignments; the size of the letter coincides with its relative abundance at each position. Below the SeqLogo is an “identity” panel, where the height is equal to coverage at each position in the alignment and color represent conservation of amino acids with dark green = 100%, light green = 30–99%, red = less than 30%. Functional domains used for ZAPP development are indicated. Asterisks represent individual amino acids associated with SH3 binding.

A standard regression analysis was used to determine which of the 792 features individually maximized class separation between Nef sequences derived from brain and non-brain tissues. 62 of the 792 features had slight class separation (R²>0.8) and are described in Supplemental Table I for each region.

These 62 features were then used as input in the development of evolved neural networks using ZAPP as described below (Fogel 2008; Porto, Fogel, and Fogel 1995; Fogel, Fogel, and Porto 1990). The set of all sequences with known brain or non-brain origin was divided randomly into training (n=1251 sequences) and testing (n=626 sequences). As noted previously, a validation set was generated using sequences derived from CSF, spinal cord, and meninges. The process of generating training and testing sets was repeated three times with random assortments of the sequence data to avoid sampling bias.

ZAPP utilizes an evolutionary algorithm to optimize a population of neural networks using a set of input features on training and testing data (Lamers, Fogel, et al. 2016; Lamers et al. 2017; Fogel et al. 2015; Fogel et al. 2014; Porto, Fogel, and Fogel 1995). A population of neural networks was generated and scored with respect to minimization of mean squared error (MSE) on the output (in this case, the assignment of a sequence to either the brain or non-brain class). At each generation of the evolutionary optimization, models with lowest MSE were saved as “parent” neural networks for the generation of “offspring” neural networks with some variance in the weights associated with connections, or aspects of the neural network architecture itself. For neural network development, MSE is minimized on the training data, while performance on the testing data is measured every 50 generations to avoid over-fitting. The best neural network is then pulled from the population and assayed for performance on the training, testing, and validation sets. For the experiments conducted here, a population of fully-connected, feed-forward neural networks was used, with sigmoid functions for each node, and a fixed architecture. However, given the feature set had a maximum of 62 features, we forced the neural networks to use reduced sets of inputs, where the number of inputs was fixed at either 19 or 20 at a time, with the evolution allowed to randomly choose which of the 62 features to use at input. Evolutionary optimization was used to simultaneously optimize both the weights of the neural network and the features used as input. This approach allowed for the exploration of small feature subsets and their interdependency. In addition, the use of three different divisions of the training and testing data allowed us to determine if the same sub-selected features were obtained for all three replicates or if features were largely specific to each division of the data, which is important because a useful biomarker for brain vs. non-brain Nef would rely on features that were found to be common over as many divisions of the data as possible.

All neural networks were optimized using 50 parents and 50 offspring models in the evolving population, with tournament selection using 4 opponents. The number of generations of evolutionary optimization varied for each neural network architecture and dataset and was chosen as the maximum amount of training generations without increase in testing MSE. Once a best neural network was achieved, a binary discriminating threshold was determined to maximize classification accuracy of the two classes brain and non-brain. This threshold was fixed prior to use with any testing or validation samples. In this way, the neural network output (ZAPP score) could be compared to the threshold for the purpose of binning sequences into one or the other class. The thresholds used for each dataset are provided in Table II.

Table II.

Thresholds generated on the output of the best neural network on the training samples only, used for classification of all subsequent sequences in the testing and validation sets.

Neural Network Architecture (number of input-hidden-output nodes)	DS1	DS2	DS3
62-10-1	0.82	0.87	0.77
20-10-1	0.72	0.76	0.79
10-10-1	0.75	0.77	0.78
10-5-1	0.73	0.72	0.76
5-3-1	0.78	0.78	0.74

VESPA, a web-based tool used to identify signature patterns in aligned sequence data (www.lanl.gov), was applied to two aligned sequence populations: those classified as brain and those classified as non-brain by the best neural network. VESPA calls a signature position when the most common character in the query data (in our case brain sequences) differs from the most common character in the background data (in our case non-brain sequences). and provides a frequency score.

Sequence Accession Numbers

AB221005, AB253432, AY314054-AY314063, AY713409, DQ357219-DQ357221, DQ358012-DQ358047, EF656973-EF656985, EF656994, EF656995, EF657017-EF657029, EF657035-EF657045, EF657063-EF657121, U63632, GQ868779-GQ869380, HM002302-HM002466, HQ174334-HQ174415, JQ990945, KU645011- KU645017, KU645030- KU645194, KU709356- KU709831, MF511264- MF511677, MF579824-MF579855, U23487, U66543-U66556.

Results

Amino acid variation along the Nef sequence population used in the study is shown in “SeqLogo” format in Fig II. Only two positions in the alignment (shown as dark green in the identity plot) were conserved at 100% (66A, 227P). The performance of each of five separate neural network architectures was evaluated on both training and testing data for dataset 1 (DS1), dataset 2 (DS2), dataset 3 (DS3)(Supplemental Tables II & III). These results indicate that neural networks with as few as 10 inputs have reasonable accuracy in separating brain from non-brain Nef sequences. The 10-10-1 neural networks retained reasonable mean performance on training and testing while using reduced sets of features and using 10 hidden nodes rather than 5 for improved generality. Among the three best 10-10-1 neural networks from the three datasets, only one feature, AA Comp SwissProt 2 in the SH3 binding domain, was common to all three, nine features were found in two of the three, and another nine features were unique to a single dataset (Table III). The most shared features were largely from the SH3 binding domain, Nef loop, cytokines binding, Alpha CD, and AP2 binding domain, suggesting that these domains have importance in the classification of brain vs. non-brain associated Nef. The 10-10-1 neural network from DS2 had the highest predictive accuracy on training (85.4%) and testing (84.3%) and was then used to process all of the sequences from the validation set. The resulting output scores were used, along with the same discriminating threshold used for the analysis of training and testing data, to assign validation sequences to the class brain or non-brain (Figs III and IV). Note that the distribution of the training and testing sequences was roughly similar across patients and over the final scores.

Table III.

Features identified through three independent optimizations of 10-10-1 neural networks. Model 2 had the highest predictive accuracy on training (85.4%) and testing (84.3%) and was used to process all remaining sequences with unknown etiology as either brain or non-brain.

Model 1		Model 2		Model 3
Location	Feature	Location	Feature	Location	Feature
SH3 Binding	AA Comp SwissProt 2	SH3 Binding	AA Comp Swiss Prot 2	SH3 Binding	AA Comp Swiss Prot 2
		AP2 Binding	Grantham	AP2Binding	Grantham
Nef Loop	ChargeConversion			NefLoop	Charge Conversion
Nef Loop	Polarity 2	Nef Loop	Polarity 2
SH3 Binding	Parker	SH3 Binding	Parker
Cytokines Binding	Cowan			Cytokines Binding	Wilson
		Cytokines Binding	Welling	Cytokines Binding	Welling
		Alpha CD	Beta Sheet Levett	Alpha CD	Beta Sheet Levett
		AP2 Binding	HPScale	AP2 Binding	HP Scale
Alpha CD	Transmembrane			Alpha CD	Transmembrane
AP2 Binding	Tanford
Cytokines Binding	Aboderin
AP2 Binding	Kyle Doolittle
SH3 Binding	Antiparallel Beta
		AlphaCD	Janin
		Alpha Helix A	HP Scale
		AP2 Binding	Browne
				Alpha Helix A	Kyte Doolittle
				Nef Loop	Wolfenden

Fig. III. ZAPP scores for all subjects.

Subject ID, subject pathology and compartments (brain, non-brain, and validation sequence populations from meninges, spinal cord, PBMC, CSF) are listed on left-hand columns. An asterisk indicates subjects who died with no detectable plasma viral load. ZAPP score is on the x-axis, with a calculated discriminatory threshold for Brain and Non-Brain designated as a vertical line at 0.77. Scores for sequence used for testing, training, validation are shown and colored according to legend.

Fig. IV. ZAPP scores for sequences for selected cloned HIV-1 Nef sequences.

Sequence names are listed on left-hand columns along with the tissue of origin (if known). ZAPP score is shown on x-axis, with a calculated discriminatory threshold for Brain and Non-Brain designated as a vertical line at 0.77. Scores for sequence used for testing, training, validation are colored according to the legend in Fig. 3.

These results indicated that there were underlying physicochemical differences between Nef derived from patients with HAD as compared to other populations, particularly cancer patients (Fig III). 26 out of 27 of brain sequences from the two patients without neurological disease (C05, and AZ) were classified as non-brain. On the other hand, and 309 out of 451 non-brain sequences from patients with HAD/HAND were classified as brain, suggesting that brain-associated features persist in Nef derived from non-brain body tissues of patients with neurological disease. The sequences from one HAND patient in the study who was on cART until death (C02*) scored similarly to patients with HAD. The 55 sequences from HAD subjects MACS1 and MACS2 had an intermediary neural network score and were somewhat differentiated from cancer subjects, whose neural network scores were routinely low (non-brain); however, just 6 of these sequences were derived from actual brain tissues. Only one lymphoma subject (PATU3) had a single sequence that scored as brain; however, all additional sequences from this patient had low scores.

Sequences from seven molecular clones available from the NCBI AIDS reagent resource were included in the study (Fig IV). Through random allocation, two were used in training, one was used in testing, and six of these isolates were used in the validation set. Notably. the only molecular clone sequence (JR-FL) validated as brain by ZAPP was derived from brain tissue. JR-FL is derived from the same patient as the sequence JR-CSF, which was obtained from cerebral spinal fluid, and classified by ZAPP as non-brain.

The most important features for classification were the composition of SH3 binding positions, the Welling hydrophobicity scale applied to the cytokines binding domain, and the polarity of the Nef Loop and AP-1 binding domains (Table III). Each of these features have important biological functions associated to the function of Nef in the brain, specifically to astrocyte and macrophage infection. These biological implications are considered in greater detail in the discussion.

VESPA (www.lanl.gov) was applied to two aligned sequence populations: brain and non-brain as classified by the best 10-10-1 neural network from DS2. The program identified 13 positions in the brain sequence alignment that differed from the most common amino acids in each position in the non-brain sequence population (Fig V). Of the 13 total positions, six occurred in the highly variable myristoylation domain; however, most of these were had relatively low frequency scores (<22%). Note that the neural networks also did not identify the myristoylation domain important for brain vs non-brain separation in any of the best evolved networks. This result is similar to reports by others (Johnson et al. 2016). Five other positions had a difference in frequency score between 35–53% and occurred in regions of Nef that were identified by the neural networks as important for brain classification (corresponding to Fig 2 alignment positions 110, 128, 129, 160, 190) (Fig V). We sampled 7 brain and non-brain sequences from our alignment to view these 13 positions directly in sequences with known tissue origin; overall, it appeared unlikely that this signature alone could be used to correctly classify brain sequences (Fig V).

Fig. V. Signature analysis of predicted brain and predicted non-brain sequence populations.

Panel A. A sequence alignment containing all DS2-predicted brain sequences was used to query against the background alignment containing all DS2-predicted non-brain sequences. A signature position (top row) is called when the most common character in the query data differs from the most common character in the background data. Signature amino acids for the brain alignment are colored green and for the non-brain alignment are colored light orange. Positions where the query differed from the background more than 30% are indicated with red boxes. In the middle and lower panels, a selection of brain and non-brain signature positions derived from Nef sequences with known tissue origin are shown. The ZAPP scores for each particular sequences are shown in the second column. The signature positions are colored according to a brain signature (green) or a non-brain signature (salmon). Positions where the amino acid did not match either signature are not colored. Panel B. 3D Nef JRCSF in “cartoon” format showing selected structural and functional regions. Identified signature positions are shown as “spheres”. Alpha Helices A and C, which were identified as useful for brain classification, are indicated.

Discussion

HIV-associated neurocognitive changes have been reported since early in the AIDS epidemic (Snider et al. 1983). These CNS symptoms were initially called “HIV Associated Dementia” or HAD (Navia, Jordan, and Price 1986; Antinori et al. 2007). HAD was associated with one or more histopathological abnormalities on autopsy that were unique to HIV patients, such as HIV encephalitis (HIVE), HIV leukoencephalopathy, and diffuse poliodystrophy (Budka 1991; Budka et al. 1991; Budka et al. 1987). HIVE was defined as multiple disseminated foci of microglia, macrophages, and multinucleated giant cells. HIV leukoencephalopathy was defined as diffuse damage to white matter that included myelin loss, reactive astrogliosis, macrophages, and multinucleated giant cells with little or no inflammatory infiltrates. Diffuse poliodystrophy was defined as reactive astrogliosis and microglial activation involving the cerebral gray matter (Budka et al. 1991; Pumarola-Sune et al. 1987). Overall, the postmortem finding most strongly associated with HAD was the number of activated macrophages present within affected areas of the brain (Glass et al. 1995). In addition, dendritic loss (Masliah et al. 1997), neuronal loss (Bell 1998), and brain HIV viral load (Bhaskaran et al. 2008) were associated with HAD.

After the introduction of cART, the incidence of HAD declined dramatically (Bhaskaran et al. 2008); however, milder forms of HIV-associated neurological disorders became prevalent, found in up to 50% of HIV-infected individuals (Heaton et al. 2010) and led to a new nosology called “HIV-Associated Neurocognitive Disorders” (HAND), which includes a diverse spectrum of pathologies, from mild neurocognitive disorders to HAD (Antinori et al. 2007). cART-treated patients with HAND differ from pre-cART patients in many ways. They are less likely to have detectable HIV RNA or have high levels of certain inflammatory biomarkers in their CSF (McArthur et al. 2004). At autopsy, HAND may or may not manifest florid neuropathological changes (Everall et al. 2009; Koedel et al. 1999; Lamers, Rose, et al. 2016). Currently there is a re-evaluation of the pathogenic mechanisms of HAND, including interest in persistent brain viral reservoirs or HIV proteins that can act independently of intact HIV that may drive persistent CNS inflammation and neurotoxicity, all of which are still associated with roles for activated macrophages {Koedel, 1999 #1711;Annunziata, 2003 #1712;Bergonzini, 2009 #1713;Fiala, 1996 #1714;Neumann, 1995 #387;van Marle, 2004 #246;Khan, 2016 #1648;Lenassi, 2010 #1679}.

HIV-1 Nef is a multifunctional protein required for pathogenicity of the virus (Kestler et al. 1991). The protein maintains a high degree of genetic variation among intra- and inter-host isolates. Increased attention has focused on the role of Nef in HAND pathogenesis, as it is expressed abundantly from astrocytes (Sami Saribas et al. 2017) and has been identified in blood and brain exosomes during cART (Raymond et al. 2016; Khan et al. 2016). This could constitute a potential mode of intercellular communication (van Niel et al. 2006) as Nef could prime permissive cells or activate immune responses in myeloid cells (Sami Saribas et al. 2017). Nef also dysregulates autophagy, an adaptive response to stress, in several cell types (Saribas, Khalili, and Sariyer 2015; Gupta et al. 2017). As the actions of Nef continue to grow, we were interested if brain-derived Nef shared characteristics that could allow for its identification and would provide further support for the protein’s role in HAND pathogenesis.

In this study we used statistical analysis combined with ENNs (Lamers et al. 2017; Lamers, Fogel, et al. 2016) to identify physicochemical alterations in Nef sequence domains that could be used to classify Nef sequences into either brain or non-brain classes. ENNs are nonlinear functions (neural networks) that map input variables/features to an output, through a hidden layer of nodes and processing functions. Evolutionary computation is used to search the space of nonlinear functions that has highest accuracy on training examples, while retaining useful generalization on unseen testing data. This approach allows for model optimization and simultaneous sub-selection of the features used as input to the model. In that sense, the approach provides a unique understanding of the features that are best suited for classification. This is in direct contrast to standard statistical methods that only use correlation analysis to search for “best” features one at a time. The best ENN discovered through this process had an 85.4% true positive and true negative classification accuracy on the training data set and 84.3% accuracy on the testing data set. Signature pattern analysis (Fig V) confirmed that changes associated with brain or non-brain data sets were more complex than simply the identity of the amino acids in positions across the Nef genome. The analysis suggests that subtle adjustments of amino acid physicochemical features within multiple functional domains helps define a Nef sequence associated with HAND tissues. These slight adjustments may only be identified through modeling approaches that allow for multiple features to be used simultaneously through a nonlinear processor. Despite the finding of brain-specific Nef sequences across numerous subjects, it remains difficult to translate this information into functional viral life cycle without further assessment of such variants in biological assays.

In all three ENN experiments with 10-10-1 neural networks, the composition of the Src homology-3 (SH3) binding domain was important in the classification of brain isolates. The SH3 binding capacity by Nef is required for many other HIV protein interactions to take place (Tokarev and Guatelli 2011; Arold et al. 1997; Alvarado et al. 2014). Nef interacts with SH3 domain proteins primarily via a conserved PxxP motif in its core domain; however, other positions in Nef are also associated with SH3 binding (Fig II) (Grzesiek et al. 1996; Lee et al. 1996). One of the alignment positions involved in SH3 binding, 110G, had the a highly conserved brain amino acid signature, identified in 80% of brain sequences and in only 2% of non-brain sequences. One well-known binding partner of HIV Nef is the monocyte-specific kinase Hck (Grzesiek et al. 1996), and therefore a specific SH3-binding configuration would be consistent with infection of brain immune cells, e.g. monocyte-derived macrophages (Cornall et al. 2013).

Polarity in the Nef Loop domain and the AP-2 binding domain (a subset of amino acids in the Nef loop) was used to classify brain viruses in two of the 10-10-1 neural networks. AP-2 is a cell-specific transcription factor family consisting of five closely related proteins (α, β, γ, δ and ε) that regulate the expression of specific target genes (Damberg 2005; Eckert et al. 2005). AP-2α and β are abundant isoforms in the brain and represent binding sites for putative transcriptional factors present in the Apolipoprotein E (ApoE) promoter. ApoE has been shown to co-localize with neuropathological lesions in Alzheimer’s disease. The Nef loop has other known functions, such as trafficking and Nef internalization (Zinkernagel 1976). The CD4 binding domain is concentrated in the N-terminal flexible loop and the core domain of Nef; recruitment of the partners AP and COP I is mediated by the C-terminal loop (Geyer, Fackler, and Peterlin 2001).

The Welling scale applied to the cytokine-binding domain was found to be important the classification of brain vs. non-brain Nef sequences. This unique scale was developed based on the percentage of amino acids in known antigenic determinants of a variety of proteins compared with the percentage of the amino acids in the average composition of a protein (Welling et al. 1985). Increased antigenicity of HIV brain Nefs could associate with macrophages that produce inflammatory cytokines, which are important in driving adaptive immune responses (Koppensteiner, Brack-Werner, and Schindler 2012).

Other structural features identified through our approach (Fig VB) were within the first alpha-helix region A and two other smaller regions in the unstructured C-terminal Nef. Alpha-helix A is in the hydrophobic pocket in the core of the protein; proper positioning of the helix with a tryptophan residue at position 15 (Fig II) is thought to be necessary for positioning of Nef at membrane surfaces (Johnson et al. 2016). In our own Nef structural minimization experiments (unpublished data), we have observed that the smaller defined alpha-helices in C-terminal regions have varied potential for the formation of secondary structure which is typical for unstructured regions of the protein (Barnham et al. 1997).

A variety of subjects, tissues, and pathologies were evaluated in the study. Of the 13 patients diagnosed prior to death with neurological disease, 11 had sequences that were classified as brain. Furthermore, nef sequences from the non-brain tissues of these patients were also frequently classified as brain-like, suggesting that sequences derived from anatomical tissues outside the brain could be used a biomarker for HAND potential. An extensive phylogenetic study of the HIV env gene by Rife et al., using six CD8-depleted and twelve naturally progressing SIV-infected macaques, demonstrated early entry of virus into the brain, followed by a period of adaptation in the periphery that lead to neurotropism later in infection (Rife et al. 2016). The higher scores for non-brain sequences derived from end-stage HAD patients are consistent with these findings, although evolution in env and nef is not always comparable (Lamers et al. 2015). Notably, 13 sequences derived from PBMCs from three patients with HAD (BI, CB1, and CB3) all classified as brain. A future study could apply ZAPP to pre-cART PBMC samples combined with post-mortem tissues, in order to substantiate the hypothesis that Nef derived from PBMCs, which scores in the brain class by ZAPP, is predictive of HIV-associated neurological disease.

Relatively low ZAPP scores were calculated for cancer patients, suggesting that a cancer-associated Nef sequence population may exist. Therefore, a similar ZAPP workflow could be designed to identify a Nef biomarker for cancer potential in HIV-infected patients. This finding could relate to the alternative macrophage activation states in these disease processes (Lamers et al. 2012). During lymphoma pathogenesis, M2-activated macrophages produce anti-inflammatory cytokines and Nef stimulates MDMs to form conduits that selectively transfer Nef to B-cells, thus allowing the protein to bypass B-cell immunoglobulin receptors (Lamers, Fogel, et al. 2010; Xu et al. 2009; Moir and Fauci 2010). During HAD pathogenesis, M1-macrophages produce inflammatory cytokines that trigger the recruitment of further macrophages into the brain (Lamers et al. 2012).

Most subjects in the study had an unknown course of cART; however, at least four were treated until death, at which time no measurable plasma viral load was detected. These tissues were derived from donors enrolled in the National Neuralological AIDS Bank (NNAB), which clinically treats and monitors patients with neurological disease or a severe life-limiting HIV-associated comorbidity. As we reported previously (Lamers et al. 2017; Rose et al. 2016), brain tissues from these subjects were assayed for HIV using a digital droplet PCR (ddPCR) approach as well as env-nef single genome sequencing of HIV RNA and DNA. While ddPCR confirmed the presence of HIV in the brain tissues for all four of these subjects, the larger segment of HIV DNA or RNA (env-nef) was only abundantly amplifiable from subject C02*, who was diagnosed with HAND and schizophrenia over a year prior to death (Lamers et al. 2017). Three different tissue biopsies from subject C05*, whose primary pathology at death was noted as metastatic lung cancer, yielded only two env-nef brain DNA sequences. However, all of these subjects had noted brain pathology at death: C02 with neoplastic infiltrate; C04, C05, and C09 with Alzheimer’s type-2 gliosis. Interestingly C02, C09 and C05 all had sequences (brain or not) that classified as brain-like, thus again suggesting that it could be possible to use our approach to indicate HAND-risk in patients on cART using Nef sequences derived from non-brain tissues such as blood.

The best neural network did not show perfect separation between brain and non-brain classes, which should be expected as the sequences are sampled from a very complex human system, and we might not have identified all of the features required to attain perfect separation. Furthermore, previous phylogenetic studies have shown that HIV/SIV from brain tissues can move across the blood-brain barrier and infect cells in the periphery (Lamers, Salemi, et al. 2010; Rife et al. 2016). Additional re-training and testing of such nonlinear classifiers using a larger number of brain sequences, especially from diverse donors, could increase the accuracy of the system. However, with an accuracy of 84.5%, the best neural network already has high potential to identify isolates with brain-like characteristics that could be used in vivo or in vitro to better understand HIV-associated neurological comorbidities.

While it can be difficult to generate an abundance of long HIV sequences (env-nef) from most cART+ subjects, recent studies have shown that defective proviruses from patients on cART can express HIV RNA species with unusual exon combinations that usually contain Nef, even when the full HIV genome cannot be detected (Imamichi et al. 2016), thus, the Nef protein remains an attractive target for biomarker development in cART+ patients in comparison with other HIV proteins. In order to increase our understanding of the HIV Nef protein’s role in HAND, additional studies are needed to identify if Nef proteins alone are frequently within brain tissues derived from patients on cART, even when quantitative PCR, which usually targets gag RNA, are negative.

HIV Nef is a fascinating protein due to its flexible structure, multiple functions and diverse cellular targets. Furthermore, HIV Nef is relevant to neurological disease in the cART-era when considering the protein’s role in promoting survival and migration of infected brain macrophages. In this study, we show that the novel ZAPP approach has strong potential to classify brain-tropic Nef sequences, especially with the generation of additional data for training and testing.