Concurrent influenza vaccination reduces anti-FVIII antibody responses in murine hemophilia A

The enigmatic development of factor VIII (FVIII)–neutralizing antibodies, known as inhibitors, remains the most serious treatment-related complication of hemophilia A (HA). Immune tolerance induction is currently the only proven strategy for eradicating antibodies toward FVIII but is costly and varies in efficacy.^1,2  Prevention of inhibitor development would therefore be an ideal solution. However, why 25% to 30% of severely affected patients develop inhibitors is poorly understood.^3,4  Clinical and epidemiological studies attribute this to a complex combination of genetic and treatment-related factors.

FVIII immunity likely begins with receptor-mediated endocytosis and degradation of FVIII by professional antigen-presenting cells (APCs) and the subsequent presentation of FVIII peptides on major histocompatibility complex (MHC) class II molecules.^5,6  In the immature steady state, APCs are unable to prime FVIII-specific effector T-cell responses because of insufficient expression of costimulatory molecules such as CD80 and CD86. However, in a proinflammatory milieu, APCs mature and upregulate expression of these costimulators that enable priming of FVIII-specific effector CD4⁺ T cells and trigger a cascade of events ultimately leading to the differentiation of FVIII-specific B cells into antibody-producing plasma cells. Pathogen-associated molecular patterns (PAMPs) and danger-associated molecular patterns (DAMPs) can act as maturation and immunologic “danger signals” that enhance the quality and magnitude of the immune response against FVIII by activating APCs through pattern recognition receptors.⁷  A previous report has shown that rhFVIII, alone and in complex with von Willebrand factor (VWF), does not convey danger signals to dendritic cells (DCs).⁸  It is therefore reasonable to believe that the presence of “danger signals” influences the risk of inhibitor development in pediatric patients. We have recently shown in 2 mouse models of HA that surgery-associated DAMPs increase inflammatory cytokines interleukin-1 and interleukin-6, and also upregulate CD80 on APCs, but do not influence the magnitude or incidence of the anti-FVIII immune response.⁹ 

Vaccination for infants occurs at around the time of initial FVIII replacement therapy and, as a result, may act as a danger signal.¹⁰  This concern is founded on the exposure to viral antigens and PAMPs, and in some cases adjuvants, that may enhance antibody responses. Indeed, studies in mice have implicated an amplification of the anti-FVIII immune response by lipopolysaccharide, a Toll-like receptor 2/4 (TLR2/4) agonist.¹¹  No studies have reported a role of vaccination in inhibitor development, save for a single case report in which subcutaneous immunotherapy treating a pollen allergy reverted FVIII tolerance in a severe HA patient who had already undergone successful immune tolerance induction.¹² 

From a contrasting perspective, there is some concern that simultaneous exposure to multiple vaccines may reduce their overall effectiveness through interference or immune overload, rendering patients susceptible to other infections.¹⁰  In the context of FVIII, this antigenic competition may divert the immune system’s resources away from FVIII and result in a more anergic or tolerogenic response. Indeed, animal models have shown decreased antibody titers against FVIII when concurrently treated with a FVIII/VWF complex vs FVIII alone.¹³  However, there is conflicting evidence in the literature regarding concurrent vaccination. Recent studies have suggested that multiple concurrent vaccinations do not influence the efficacy of protection against each antigen, nor the susceptibility to other infections.^14-16  Ultimately, the role of vaccination on FVIII immunogenicity or tolerance remains unclear and is the subject of this report.

Three commonly used vaccines were selected based, in part, on their recommended age of inoculation: the live-attenuated measles-mumps-rubella (MMR) vaccine and the seasonal inactivated subunit influenza vaccine with and without the adjuvant MF59.¹⁰  Here, we assessed the effects of vaccination on recombinant human FVIII (rhFVIII) immunogenicity in a humanized HA mouse model with a variable tolerance to intravenously administered rhFVIII.¹⁷ 

The 8- to 12-week-old male humanized F8 E17KO HLA-DRB1*1501 HA mice on a mixed S129/C57Bl/6 background were gifts from the Baxalta Corporation.¹⁷  These mice possess a complete knockout of the murine MHC class II locus and instead express a functional chimeric human-mouse HLA-DRB1*1501 allele. The 8- to 12-week-old male and female C57Bl/6 F8 E16KO HA mice were used for in vitro chemotaxis experiments.¹⁸  All mouse experiments were reviewed and approved by the Queen’s University Animal Care Committee.

Mice were vaccinated either subcutaneously or intravenously with the MMR vaccine (Priorix; GlaxoSmithKline), and intramuscular or intravenously with the influenza vaccine in the presence and absence of the adjuvant MF59 (Fluad [2013-2014 season] and Agriflu [2012-2013 season], respectively; Novartis).

For MMR experiments, naïve HA mice received 4 weekly intravenous infusions of 2 IU rhFVIII (0.2 μg or 80 IU/kg, Advate; Baxter BioScience), followed by 4 twice-weekly infusions of 6 IU rhFVIII (0.6 μg or 240 IU/kg). For influenza experiments, mice received 7 twice-weekly infusions of 6 IU rhFVIII. Experiments described represent aggregated cohorts of at least 3 independent experiments.

Mice were anesthetized and blood was collected using capillary tubes via the retro-orbital plexus. Sodium citrate (3.2%) was added to the samples at one-tenth total volume. Plasma was isolated by centrifugation. Terminal samples were collected by cardiac puncture.

Total FVIII-specific immunoglobulin G (IgG) titers were quantified by enzyme-linked immunosorbent assay (ELISA) as previously described.¹⁹  Total FVIII-specific IgG was detected using an horseradish peroxidase–conjugated goat anti-mouse IgG (Southern Biotech); positive results were defined as >0.1 optical density. FVIII-specific IgG isotypes were quantitated using the SBA Clonotyping System horseradish peroxidase (Southern Biotech). Wells coated with IgG capture antibodies were used with mouse reference serum (Bethyl Laboratories) to generate a standard curve.

FVIII inhibitors were assessed by a 1-stage FVIII clotting assay using an automated coagulometer STACompact (Stago). Samples were incubated at 56°C for 30 min to eliminate background FVIII activity in our experimental system and prepared as described previously.²⁰  Dilutions were performed using FVIII-deficient human plasma (Affinity Biologicals). Positive inhibitor samples were defined as >0.5 BU.²¹ 

Spleens and inguinal lymph nodes (LNs) were isolated from F8 E16KO HA mice. Splenocytes and LN lymphocytes were stimulated at 1 × 10⁷ cells per mL for 18 hours with FVIII (1 μg/mL) or influenza vaccine (1 μg/mL), respectively. Total CD4⁺ T lymphocytes were isolated using the EasySep Mouse CD4⁺ T-cell Isolation Kit (StemCell Technologies). Purity ranged from 88% to 95%. Naïve splenocytes or CD4⁺ T cells were labeled with carboxyfluorescein succinimidyl ester (CFSE) (CellTrace CFSE Cell Proliferation Kit; Life Technologies) and suspended in a 1.5 μg/mL rat tail collagen matrix. Three-dimensional migration was imaged by confocal microscopy using μ-Slide Chemotaxis 3D (ibidi) slides. Positive controls used were 10 μg/mL CCL2 (R&D Systems) and 20 ng/mL CCL5 (eBioscience). Cell migration was automatically tracked and plotted using the MosaicSuite and Chemotaxis Tool Fiji plugins, respectively.

The live-attenuated MMR vaccine is typically administered as a series of 2 subcutaneous doses.¹⁰  In this study, we attempted to mimic this regimen with an additional goal of assessing the ability of the MMR vaccine to break FVIII-specific tolerance, as well as expose humanized E17KO mice to low- and high-intensity FVIII treatments (Figure 1A). The vaccine was administered at 10 times the weight-scaled standard human dose; preliminary studies using a standard dose were inconclusive (data not shown). Plasma levels of anti-FVIII IgG and inhibitory Bethesda units were measured at the indicated time points. Our results show that subcutaneous vaccination of mice with the MMR vaccine 24 hours before the first FVIII exposure does not influence the proportion of mice that develop FVIII-specific IgG (Figure 1B). Similarly, a second reexposure to the vaccine followed by a more intensive FVIII treatment regimen exhibited no effect on FVIII IgG or inhibitors (Figure 1B-C). We also quantified the titers of FVIII-specific IgG and inhibitors in FVIII responders, which showed negligible differences between vaccinated and nonvaccinated mice (Figure 1D-E).

Given this lack of effect on FVIII immunogenicity, we then asked whether the MMR vaccine inherently conveys danger signals to the immune system. To address this, we immunized mice intravenously with MMR to deliver both the vaccine and FVIII for immune processing in the spleen.²²  Surprisingly, we observed no influence of the initial MMR vaccination, nor the reexposure, on the incidence of FVIII-specific IgG and inhibitors (Figure 1B-C). Among FVIII responders, the difference in titers of FVIII-specific IgG and inhibitors was again insignificant between vehicle and MMR-exposed groups. (Figure 1D-E).

To address the stark contrast between these data and our hypothesis, we considered the efficacy of the MMR vaccine in mice. Measles-specific IgG was undetectable in mice vaccinated either subcutaneously or intravenously, indicating unsuccessful immunization (data not shown), which may be required to observe an influence on FVIII immunogenicity. Furthermore, analysis of DC costimulatory molecule expression demonstrated that there are no accessible danger signals in the MMR vaccine that may modulate FVIII antigen presentation in mice (supplemental Figure 1, available on the Blood Web site).

We next investigated the potential effects of the influenza vaccine, a surface antigen vaccine composed of seasonal-specific hemagglutinin and neuraminidase. To assess its role on FVIII immunogenicity, humanized E17KO HA mice were immunized intramuscularly with a standard human dose (4.8 μg/kg), either 24 hours before, at the same time, or 24 hours after the first of 7 twice-weekly FVIII infusions (Figure 2A). Four weeks after the initial FVIII infusion, we observed significant decreases in the incidence of FVIII-specific antibodies regardless of when the mice were vaccinated (30%, 40%, 14% vs control: 80%; Figure 2B) We also observed similar decreases in the incidence of FVIII inhibitors (30%, 43%, 11% vs 80%; Figure 2C). Further serological analyses revealed that all mice exhibited high titers of anti-influenza IgG, suggesting successful immunization (Figure 2D). Titers of anti-FVIII antibodies and FVIII inhibitors were also measured in FVIII responders, and there were no changes in either outcome in response to influenza vaccination (Figure 2D-E). Furthermore, there was no correlation between the titer of influenza-specific antibodies and FVIII-specific antibodies or inhibitors (data not shown), indicating that influenza does not augment the anti-FVIII immune response. To examine the potential influence of influenza vaccination on the FVIII immune response profile, we analyzed anti-FVIII IgG subclasses IgG1, 2a, and 2b. We observed an increase in FVIII-specific IgG2a antibodies attributed to concurrent influenza and FVIII exposure (Figure 2F).

We then assessed the robustness of the apparent influenza-mediated reduction in anti-FVIII antibody responses. Following an initial intramuscular influenza immunization concurrent with the first of 7 FVIII infusions, mice were subjected to a 5-week period without FVIII exposure, after which they received 3 consecutive intravenous infusions of FVIII. Two weeks after this intensive reexposure, we observed an increase of 40% to 50% in the incidence of FVIII-specific IgG in the vaccination treatment arm (supplemental Figure 2). Moreover, this decrease in the FVIII immune response appears to be specific to the influenza vaccine, as mice immunized intramuscularly against a more inert antigen, ovalbumin, did not exhibit the same pattern of reduced FVIII immunoreactivity (supplemental Figure 3).

Intramuscular vaccination does not ensure its colocalization with FVIII. Therefore, we next evaluated the presence of a danger signal by inoculating mice intravenously with vaccine and FVIII at the same time points as described previously, with the assumption that the biodistribution of the 2 products is more likely to be similar. In this scenario, any potential danger signal would stimulate the same cells interacting with FVIII. When mice received FVIII at the same time as the intravenous vaccine, there was a 30% and 40% increase in the incidence of FVIII-specific IgG and inhibitors, respectively (Figure 3A-B). However, when vaccinated either 24 hours before or after the first FVIII infusion, the incidence of inhibitory antibodies was lower, although there was an increase in total anti-FVIII antibodies in mice vaccinated 24 hours before receiving FVIII. Intravenous immunization against influenza elicited similarly robust anti-influenza immune responses compared with intramuscular vaccination (Figure 3C). No differences in FVIII antibody or inhibitor titers were associated with influenza immunization compared with controls (Figure 3C-D). These data, albeit not statistically significant, suggest a trend of increased FVIII immunogenicity associated with danger signals in the influenza vaccine delivered to the same anatomic location as FVIII.

We next investigated FVIII immunogenicity in response to the influenza vaccine with the additional of an oil-in-water squalene adjuvant, MF59, which may confer immune-stimulatory properties against FVIII as well as influenza antigens. Intramuscular vaccination of HA mice together with an intravenous infusion of FVIII resulted in a decreased incidence of FVIII-specific IgG (45%, 25%, 42% vs 80%; Figure 4A). In addition, the frequency of FVIII inhibitor positive mice significantly decreased across all immunization time points (36%, 31%, 33% vs 80%; Figure 4B). All mice developed high titer anti-influenza antibodies, but there was no observed influence of the vaccine and adjuvant on the titer of total FVIII-specific IgG or inhibitors (Figure 4C-D). Upon more detailed analysis of IgG subclasses, we observed an increase of IgG1 antibodies at all immunization time points (Figure 4E).

Interestingly, mice concurrently exposed to FVIII and the adjuvanted influenza vaccine through intravenous injection did not exhibit an increased incidence of FVIII-specific antibodies or inhibitors (Figure 5A-B). Moreover, no significant changes were detected in the titers of FVIII antibodies or inhibitors (Figure 5C-D). Intravenous immunization of mice to influenza with MF59 resulted in anti-influenza responses of a similar magnitude to the intramuscular route.

FVIII immunity is thought to originate in the spleen, and the intramuscular influenza immunization likely initiates an immune response through the draining inguinal LNs.²²  We hypothesized that the decrease in FVIII antibodies and inhibitors observed through intramuscular influenza immunization is attributed to antigenic competition by means of lymphocyte trafficking to local LNs at the site of vaccination. E16KO HA mice were used to account for the variability of E17KO mice in inhibitor development, thus providing a more controlled model. To directly assess the chemotactic-inducing ability of FVIII and the influenza vaccines, we used a μ-Slide Chemotaxis 3D in which a central imaging lane, flanked by 2 separate reservoirs, was seeded with naïve splenocytes, labeled with CFSE in a 3-dimensional rat-tail collagen matrix. We then separately stimulated splenocytes with FVIII, and lymphocytes from the inguinal LNs with the influenza vaccine, with and without MF59. After 18 hours, the cell supernatant was transferred into the flanking reservoirs, establishing 2 converging chemokine gradients within the collagen matrix. The central lane was imaged overnight by confocal microscopy. Cell migration and directionality were analyzed on the basis of angular nonuniformity among end points using Rayleigh’s test. We addressed the susceptibility of this test to random short migration paths by analyzing the total distance traveled in the direction of each gradient using the Student t test. Although not statistically significant, we observed that unsorted naïve splenocytes exhibited greater chemotaxis toward media from cells stimulated with the influenza (Figure 6A) and particularly that with influenza + MF59 at 12 hours (Figure 6B). For comparison, media from unstimulated splenocytes and LN-lymphocytes resulted in the random movement of splenocytes (Figure 6C), and addition of monocyte and T-cell chemokines CCL2 and CCL5, respectively, induced direction-specific chemotaxis (Figure 6D). Similarly, there was a trend suggesting increased displacement of cells toward influenza-induced chemokines with both forms of the vaccine (Figure 6E-F).

Despite the lack of statistical power, which may be attributed to the heterogeneous nature of the cells, these data suggest that there is a cellular subset exhibiting a higher responsiveness to influenza-induced chemokines. We hypothesized that CD4⁺ T cells would be the migratory subset considering the rationale that resident APCs would contribute to the majority of the chemokines produced. Indeed, using the same experimental system, CD4⁺ T cells preferentially migrated toward media from LN cells stimulated with influenza (Figure 7A) and influenza + MF59 (Figure 7B) over a period of 12 hours. For comparison, media from unstimulated splenocytes and LN-lymphocytes resulted in the random movement of T cells (Figure 7C), and addition of the CCL5 induced direction-specific chemotaxis (Figure 7D). Furthermore, T cells were found to migrate a significantly greater distance toward the influenza-induced chemokines vs that of FVIII at 1 and 12 hours (P = .0008 and .01, respectively; Figure 7E). We observed a similar increase in migratory distance of T cells toward influenza + MF59-induced chemokines (P = .003 and .006, respectively; Figure 7F). These findings suggest that migration of CD4⁺ T cells to local LNs at the site of vaccination may explain antigenic competition between vaccine antigens and FVIII.

The immunogenic potential of APCs is influenced by maturation induced by inflammatory signals in the surrounding microenvironment such as PAMPs and DAMPs. It has been previously demonstrated that FVIII by itself, as well as in complex with VWF, does not induce APC maturation, and thus the immunogenic capability of human monocyte-derived DCs.⁸  This suggests that the microenvironment surrounding APCs in previously untreated HA patients is an essential regulator of whether immunity against FVIII develops. Indeed, a survey conducted across 42 centers showed a significant concern that administering FVIII in the presence of inflammatory stimuli, particularly surgery, would increase the risk of inhibitor development.²³  Results from the recent Concerted Action on Neutralizing Antibodies in Severe Hemophilia A (CANAL) and Research of Determinants of Inhibitor Development (RODIN) studies have respectively shown 3.7- and 2-fold increases in inhibitor risk associated with surgical procedures that are linked with moments of intensive high-dose FVIII treatment.^3,24  These findings suggest that proinflammatory signals, either foreign or endogenous, may potentiate the development of FVIII inhibitors.

We posed the question of whether vaccination would elicit an increased risk of FVIII antibody and inhibitor development with the hypothesis that vaccine-associated PAMPs would trigger APC activation. In a humanized murine model of HA, our data demonstrate that intramuscular immunization against both preparations of the influenza vaccine, with and without the adjuvant MF59, decreases the risk of developing FVIII-specific IgG antibodies and FVIII inhibitors. Furthermore, in FVIII-responding mice, we detected no differences in the titers of FVIII-specific IgG and their inhibitory activity suggesting that vaccination influences the initial decision of whether to mount an immune response but does not affect the magnitude of the immune response. However, mice immunized against influenza without an adjuvant showed increased levels of IgG2a antibodies suggesting that the normal antiviral T helper 1 (Th1) response expected exerts its influence on the immunoglobulin profile of the anti-FVIII response. As a single-stranded RNA virus, influenza has the potential to stimulate TLR7, and our data support the idea that TLR ligands, specifically TLR7 and TLR9, polarize anti-FVIII response toward Th1.^25,26  Furthermore, the adjuvanted-vaccine modulation of the anti-FVIII response with the associated increase in plasma levels of FVIII-specific IgG1 suggest that the addition of the M59 adjuvant alters the Th profile to a predominately Th2 response, a role that has previously been suggested.^27,28  However, these results may be strain specific as it has been suggested that the role of MF59 is to simply augment the default immune profile rather than biasing the profile.²⁹ 

The presence of danger signals in the vaccine preparation is irrelevant if they do not stimulate FVIII-presenting APCs. In mice intravenously immunized against influenza without an adjuvant, we observed a trend of increased FVIII immunogenicity, without a decrease in the immune response against influenza. Although not statistically significant, these findings confirm that the vaccine preparation contains danger signals, but that the effect on FVIII immunogenicity is dependent on the route of administration. Surprisingly, with the addition of the adjuvant, we observed no influence on the FVIII immune response. These findings may be attributed to a lack of interaction between FVIII and the adjuvant. In this scenario, the effect of the adjuvant on the antigen “depot” effect and lymphocyte recruitment may not be extended to FVIII.

In this study, we attribute the apparent decrease in FVIII immunogenicity to antigenic competition. The vaccine, administered intramuscularly, likely encounters the resident APCs in the muscle, followed by the subsequent migration of the APCs to the draining LN. Indeed, the MF59 preparation of the vaccine has been previously shown to provide antigen retention in the LN and induce lymphocyte recruitment.^30,31  In contrast, the intravenous infusion of FVIII likely elicits immunity in the spleen.²²  We hypothesized that concurrent vaccination and exposure to FVIII attracts immune cells toward the site of vaccination and thus decreases the effective immune response against FVIII. To directly address this hypothesis, we developed a novel in vitro antigenic competition assay allowing time-lapse 3-dimensional tracking of lymphocytes, toward media from either FVIII-stimulated splenocytes or influenza-stimulated LN cells. Following the documentation of unsorted splenocytes to migrate preferentially toward influenza-induced chemokines, we subsequently identified 1 candidate cell type as CD4⁺ T lymphocytes. Migration of helper CD4⁺ T cells away from the spleen may decrease the probability of MHC-bound FVIII peptides interacting with their corresponding antigen-specific T cell. Additionally, this trafficking may be attributed to the critical role of CD4⁺ T cells in the initiation and maintenance of influenza immunity.^32,33  Although the specific chemokines responsible for these observations were not identified, we present here, as proof of concept, the potential for vaccination to divert immune resources away from the site of anti-FVIII immune responses. It should be cautioned that the extent of this effect may be magnified in this mouse model in which APC–T cell interactions are restricted to a single human MHC class II molecule. Thus, theoretically, only a fraction of the potential FVIII peptides can be presented to T cells compared with wild-type mice.¹¹  The probability of an APC encountering an antigen-specific T cell is therefore less, and the inability of an APC to find an antigen-specific T cell may promote tolerance or anergy. Additional studies are needed to better elucidate this phenomenon, particularly in relation to the potential of FVIII-specific T-cell suppression in vivo.

We further assessed the role of a third pediatric vaccine, the MMR vaccine, and did not find any influence on the incidence or magnitude of the FVIII immune response. Mice in these studies did not mount an immune response against the vaccine delivered by either mode of administration. Incidentally, murine MMR immunization has not been reported in the literature, and mice are not known reservoirs for its viral constituents. Thus, mice may lack the necessary receptor-mediated processes required to elicit infection from the live-attenuated vaccine, and the viruses are simply degraded rather than presented to the immune system. Based on intravenous exposure of MMR and in vitro stimulation of DCs, it appears that the vaccine preparation does not contain accessible danger signals for mice.

Here, we have shown that concurrent intramuscular immunization with the influenza vaccine, with and without an adjuvant, reduces the immune response against rhFVIII in a humanized HA mouse model. A potential mechanism may be the redirection of immune resources and the recruitment of CD4⁺ T cells to the site of vaccination. Decreased anti-FVIII immunity in influenza-vaccinated mice was maintained upon intensive reexposure to rhFVIII after 5 weeks without FVIII administration. Furthermore, this phenomenon may have some specificity for the influenza vaccine given that mice immunized intramuscularly with ovalbumin did not exhibit a similar decrease in FVIII immunity. These data suggest that vaccination of young HA patients may not increase FVIII-specific antibody responses. Interestingly, an independent epidemiological study, by Hashemi et al, simultaneously observed that receiving vaccinations in close proximity to FVIII exposure in previously untreated severe HA patients did not increase inhibitor development.³⁴  Taken together with our findings, these data provide evidence that vaccination may not always increase the risk of inhibitor development and may in fact result in reduced antibody responses to FVIII. Caution should be used in extrapolating this information as this protective effect may be dependent on the properties of vaccine and the time of immunization relative to FVIII exposure. Further identification of specific T-cell subsets that migrate and remain at the site of FVIII immunity may provide insight into the critical linkage of innate and adaptive immunity required for the development of FVIII inhibitors. The inhibition of T-cell chemotaxis by fingolimod, a sphingosine-1-phosphate receptor modulator, and maraviroc, a CCR5 antagonist, have shown promising results in treating multiple sclerosis and graft-versus-host disease, respectively.^35,36  This diversion of immune resources may represent a novel avenue for FVIII immunotherapy that bypasses immunosuppression.

The authors thank Jeff Mewburn for confocal microscopy assistance and Maria Schuster for organizing breeding and shipment of humanized F8 E17KO HLA-DRB1*1501 HA mice. Research animals were provided in part by Baxalta.

This work was supported in part by an operating grant from the Canadian Institutes of Health Research (MOP-10912). D.L. is the recipient of a Canada Research Chair in Molecular Hemostasis.

Contribution: J.D.L. designed, performed, and interpreted the research and wrote the manuscript; P.C.M., K.N.S., B.M.R., C.H., and D.L. designed the study and edited the manuscript; and K.S. performed the animal experiments.

Conflict-of-interest disclosure: K.N.S. and B.M.R. are employees of Baxalta. The remaining authors declare no competing financial interests.

Correspondence: David Lillicrap, 88 Stuart St, Richardson Laboratory, Queen’s University, Kingston, ON K7L 3N6, Canada; e-mail: david.lillicrap@queensu.ca.