Fabrication of the nanoclip interface

The nanoclip interface for chronic peripheral nerve mapping and control used fabrication procedures and design features similar to those of our recent report26. Key steps in the fabrication of the device are overviewed in Supplementary Fig. 1 with device dimensions in Supplementary Fig. 2 and as follows: total thin-film device length, 40 mm; neck width, 250 μm electrode array width, 420 μm; electrode array thickness, 12 μm; longitudinal electrode pitch, 80 μm; transverse electrode pitch, 45 μm; gold interconnect line width, 10 μm; number of electrode sites, 6; overall dimensions of nanoclip, L × W × H = 300 × 300 × 400 μm. The key fabrication steps (illustrated with additional details in Supplementary Fig. 1) are as follows: (1) thin-film, multi-site electrode was custom fabricated using polyimide thin-film microfabrication techniques (HD Microsystems PI-2610; Smania, S.R.L.) and subsequently wire bonded to a printed circuit board for connection to test equipment; (2) the electrode was mounted on a thin optical glass substrate (24 × 60 mm, #0-thickness cover glass, Gold Seal) with 150-μm-thick double-sided acrylic tape (Tape #9713, 3M), and a drop (~0.5–2 μL) of liquid acrylic photoresist (IP-Dip, Nanoscribe, GmbH) was deposited over and beneath the recording array; (3) the nanoclip base was printed through the optical glass substrate using a two-photon-polymerization-based, dip-in resonant direct laser writing (rDWL) process;29 (4) the glass substrate and sample were inverted, and the top half of the nanoclip was printed with the rDWL process; (5) the photoresist was developed by submerging the glass substrate, electrode, and nanoclip in propylene glycol methyl ether acetate (PGMEA; Sigma Aldrich) for 20 min, and the whole device rinsed in methoxy-nonafluorobutane (Novec 7100; 3M) to remove trace PGMEA residue; (6) a three-electrode cell comprising the gold pads as working electrodes, an Ag|AgCl reference electrode, and a large surface area platinum counter electrode was used with an iridium tetrachloride-oxalic acid plating solution to deposit EIROF via cyclic voltammetry (CV), 20–60 cycles at 50 mV/s sweep rate between −0.05 and 0.575 V;54,55 (7) the electrodes were then rinsed with de-ionized water and characterized by CV and electrochemical impedance spectroscopy in 1x phosphate buffered saline (PBS) (see Supplementary Fig. 3; electrodes were plated to impedance, Z = 10–20 kΩ at 1 kHz and charge storage capacity, CSC = 25–40 mC cm−2). To confirm the integrity of fabrication, a subset of finished devices was sputter-coated with gold (3 cm distance to gold target; 10 s of sputtering at 0.05 mbar and 20 mA; Sputter Coater 108) and imaged via SEM micrographs (6 mm working distance with the secondary electron sensor, 3 kV accelerating voltage, 10 μm aperture; Zeiss Supra 55VP; see Fig. 1c). To investigate nonneuronal signals recorded in close proximity to devices implanted on the nXIIts, some nanoclips were fabricated with an additional nerve anchor printed on the base so that it could be attached to the nerve without the electrode pads in contact with the nerve. (See Fig. 5a for schematic of “off-nerve” nanoclips.) All mechanical design was performed with Solidworks (Dassault Systèmes). All tested and imaged devices were fabricated using a custom rDLW system29.

Mechanical modeling

All finite element mechanical modeling was performed in Solidworks Simulation (Dassault Systèms) with subsequent analysis using custom MATLAB scripts. The 3D model analyzed was the same as the 3D model from which the nanoclip was fabricated (Supplementary Fig. 2b) with the oval-shaped feet (which link to the nanoclip base through the thin-film electrode) removed (Supplementary Fig. 4a). For all simulations, the bottom edge of the model was fixed in space (indicated by the hatched base in Supplementary Fig. 4c, d). Estimates for IP-Dip mechanical properties listed in Supplementary Fig. 4b were collected from previously published studies56,57,58,59. Due to the dependence of some IP-Dip properties—and in particular Young’s modulus—on the degree of photoresist polymerization, we report the results of models using both upper- and lower-bound estimates. We assumed linearity of the material until the yield stress. For the simulation, we generated a solid mesh of 10,358 parabolic tetrahedral elements with a mean edge length of 16.5 ± 0.82 μm; the maximum element aspect ratio was 4.25 with 99.6% of elements < 3. We simulated the displacement and von Mises stresses of the trap door and hinge in response to graded forces applied uniformly and normal to the exterior surface of the door (Supplementary Fig. 4c, d, g). In Supplementary Fig. 4f, we report the maximum von Mises stress within the material, regardless of where in the structure (e.g., in which mesh element) it appeared. The full-open position of the door (Supplementary Fig. 4c–g) was defined as the first point of contact between the bottom edge of the door and the interior surface of the nanoclip. For the fatigue analysis, we simulated cyclic stretch of the trap door and hinge with sinusoidal force applied uniformly and normal to the exterior surface of the door at 10 Hz and max amplitude equal to that necessary to fully open doors in the displacement simulation (i.e., 1.25 and 7.5 μN).

Vertebrate animal subjects

Adult male zebra finches (Taeniopygia guttata; 90+ days after hatch, n = 37 birds) were obtained from the Boston University breeding facility and housed on a 13:11 h light/dark cycle in individual sound-attenuating chambers with food and water provided ad libitum. The care and experimental manipulation of the animals were performed in accordance with the guidelines of the National Institutes of Health and were reviewed and approved by the Boston University Institutional Animal Care and Use Committee. Because the behavioral effects of our interventions could not be prespecified prior to the experiments, we chose sample sizes that would allow for identification of outliers and for experimental reproducibility. No animals were excluded from experiments post hoc. The investigators were not blinded to allocation of animals during experiments and outcome assessment.

Surgical procedures

All surgical procedures were performed under isoflurane anesthesia (1–4% dissolved in oxygen), with peri-operative analgesia (1% Lidocaine, SC) and anti-inflammatory (1% Meloxicam, IM) regimens. All nerve implants reported in this study targeted the avian hypoglossal cranial nerve, the tracheosyringeal nerve (nXIIts). The nXIIts, which runs along the length of the songbird trachea and terminates at the syrinx, has a diameter of ~150 μm and is composed of both afferent and efferent fibers26. At all experiment end points, animals were given an overdose injection of sodium pentobarbital (250 mg kg−1 Euthasol, IC).

Acute preparation (n = 6 birds): an anesthesia mask was placed over the bird’s head and the animal placed in a supine position with a small pillow beneath the neck for support. Feathers were removed from the lower head, neck, and upper chest, and betadine antiseptic solution (5% povidone-iodine) and ethanol (70%) were successively applied to prepare the incision site. A 20–25 mm incision was made at the base of the neck, and the tissue blunt dissected to expose the trachea. Sutures were placed in the skin of the lateral edge of the incision and retracted from the body to expose the implant site. Connective tissue surrounding the nerve was blunt dissected away, and two sections of the nXIIts (each 3–4 mm and ~15 mm apart) were isolated from the trachea. A nanoclip interface was implanted at the rostral location. For the recording experiments reported in Fig. 3 (n = 3 birds), a bipolar silver hook stimulating electrode was placed at the caudal location; for the stimulating experiments reported in Fig. 6 (n = 3 birds), a second nanoclip interface was implanted at the caudal location. For all recording experiments, a platinum ground wire (0.003 in dia., Teflon coating; AM-Systems) was sutured to the inside of the skin and away from the neck muscles. Tissue dehydration during the procedure was minimized with generous application of PBS to the nerve and surrounding tissues. At the conclusion of the experiment, the animals were sacrificed and the devices recovered. Individual acute recording and stimulating experiments lasted 2–3 h in total.

Fictive singing preparation (n = 6 birds): the bird was prepared for surgery as above, with additional preparation of the skin at the abdomen at the caudal end of the sternum. A single 3–4 mm section of nXIIts was isolated from the trachea and implanted with a nanoclip interface, and the skin incision closed (around the protruding electrode interconnect) with 2–3 simple interrupted sutures. With the bird in lateral recumbency, a 2 mm silicone cannula was placed in the abdominal air sac, secured to the skin with finger-trap suturing, and the air sac-cannula interface sealed with Kwik-Kast (WPI). The bird was wrapped in elastic nylon mesh and isoflurane delivery switched from the mask to the air sac cannula for the remainder of the experiment. At the conclusion of the experiment, the animals were sacrificed and the devices recovered. Individual fictive singing experiments lasted 3–5 h in total.

Chronic implant for nerve function assessment (n = 12 birds): the bird was prepared for surgery and trachea exposed via blunt dissection as in the “acute preparation.” Sections of the nXIIts (each 3–4 mm) were isolated bilaterally from the trachea. Each bird received one of the following manipulations: (1) bilateral implant of nanoclip nerve interface (n = 3 birds); (2) bilateral sham implant (n = 3 birds); (3) bilateral nerve crush (transient ~1 N force applied to nXIIts with 2.5 mm wide flat forceps; n = 3 birds). Anti-inflammatory splash block (~250 μL 1% Meloxicam) was applied directly to intervention site, and incisions were closed with sutures. All birds exhibited normal rates of singing within 1 day of surgery. For intact controls (n = 3 birds), no surgical, anesthetic, or analgesic measures were taken.

Chronic implant for nerve recording (n = 7 birds): the bird was prepared for surgery as in the “acute preparation” and placed in a stereotax. A sagittal incision was made along the top of the head and the tissue retracted. Four to six stainless steel anchor pins (26002–10, Fine Science Tools) were threaded between layers of the skull, and a head cap was made from dental acrylic. The device connector was secured to the head cap with additional dental acrylic, and the nanoclip and a platinum reference wire (0.003 in dia., Teflon coated; AM-Systems) were trocared beneath the skin to the neck. The animal was then removed from the stereotax and placed in supine position with neck support. A 10–15 mm incision was made at the base of the neck, and the trachea and nXIIts were isolated as described above. The nanoclip interface was implanted on the nerve, and the reference wire secured to the underside of the skin. In n = 2 birds, a second interface with an off-nerve nanoclip (see Fig. 5a) was implanted on the same nXIIts nerve ~5 mm caudal to the first device. Anti-inflammatory splash block (~250 μL 1% Meloxicam) was applied directly to the implant site, and incisions were closed with sutures. All birds exhibited normal rates of singing within 1 day of surgery; birds acclimated to tethering and resumed singing within 1–2 days.

In vivo electrophysiology

All experiments were implemented and controlled using custom LabVIEW (National Instruments) and MATLAB (MathWorks) software applications.

Acute electrophysiology: acute electrophysiological data were recorded on the right-side nXIIts using nanoclip interfaces with a RZ5 BioAmp Processor and an RA16PA Medusa Preamplifier (Tucker-Davis Technologies). Neural signals were digitized at 24.4 kHz and 16-bit depth and were Bessel bandpass filtered (1 Hz–10 kHz, zero-phase). Stimulation currents were delivered—through either bipolar silver hook electrodes or a nanoclip interface—using a PlexStim programmable stimulator (Plexon). For all acute electrophysiology experiments, current pulses were biphasic, 200 μs phase−1 in duration, delivered at 1 Hz, and varied in amplitudes from −200 to 200 μA. By convention, positive current amplitudes are cathodic; negative amplitudes anodic.

Fictive singing: 1–1.5% isoflurane in dissolved oxygen, warmed and humidified to near-physiological levels (35 C, 80% humidity) by bubbler cascade to minimize respiratory tissue damage and prolong work time, was delivered by an air sac cannula (Fig. 7a). Flow rates were adjusted to maintain a stable anesthesia plane while eliminating broadband noise from forced gas flow over the passive syringeal labia (typically 100–200 mL min−1 at 0.5–1.5 kPa). Fictive vocalizations were recorded with an omni-directional condenser microphone (AT-803, Audiotechnica) placed 5–10 cm from the bird’s open beak, amplified and bandpass filtered (10x, 0.1–8 kHz; Ultragain Pro MIC2200, Behringer), and digitized by data acquisition boards (PCIe-6212, National Instruments) and LabVIEW software at 44.15 kHz and 16-bit depth. Current-controlled stimulation was delivered to the right-side nXIIts through a nanoclip interface using a programmable stimulator (PlexStim, Plexon) with a custom MATLAB interface. 1 kHz bursts of 100 biphasic pulses, 200 μs phase−1 in duration, delivered at 0.25 Hz, and varying in amplitudes from −200 to 200 μA were delivered at each active stimulation site (of up to six in total). The current-steering parameters used here consisted of spatially distinct multichannel stimulation patterns (all current-steering patterns tested are listed in Supplementary Fig. 7).

Chronic electrophysiology: all birds were recorded continuously using song-triggered LABVIEW software as above, generating a complete record of vocalizations and nerve activity for the experiment. Song-detection thresholds were set to detect periods in which power in the 2500–8000 Hz band (corresponding to song) exceeded 10–50 times the power in the 50–250 Hz band (corresponding to low frequency background noise); recordings continued 1.5 s after the cessation of song. Neural recordings were simultaneously acquired with an RHD 2000 system with a 16-channel unipolar input headstage (Intan Technologies), amplified, and bandpass filtered (0.3–15 kHz). Singing-related nerve activity was recorded from up to six sites on the right-side nXIIts in n = 7 birds. In n = 3 birds, recordings were made for >30 days before implant failure; these animals are reported in Fig. 4. In n = 2 birds, chronic recordings were made for 17 and 19 days before implant failure; postmortem inspection found evidence of fatigue-related fracture of the polyimide electrode along the interconnect. In an additional n = 2 birds, nonneuronal signals from a second off-nerve interface were recorded simultaneous with neuronal signals from an on-nerve interface (Fig. 5 and Supplementary Fig. 6). Across all chronically recorded birds, we found high correlation between simultaneous nerve recordings made at adjacent recording sites; we report on data collected at the most stable recording site in each bird, though we note that the trends were similar across all channels.

Data analysis

All song and electrophysiology data analysis was performed off-line using MATLAB.

Stimulation evoked responses: we sampled activity ~5 ms before and up to 25 ms after stimulation onset and used the onset of the stimulation artifact (Figs. 3b, d and 6b, c at 0 ms) to temporally align individual trial responses. Absolute response amplitudes were observed and quantified in a stimulation response window of 0.75–4 ms after stimulation onset—a latency consistent with estimated nerve conduction velocities for 4–6 μm diameter myelinated axons (i.e., 4–24 m s−1)11. SNR was calculated from mean recordings (n = 20 trials) made in the stimulation response window (i.e., signal) and 10.75–14 ms after stimulation (i.e., noise). We considered an evoked response to be detected if the SNR within the signal response window exceeded a 90% confidence interval calculated by bootstrap (i.e., resampling with replacement the signal and noise intervals over n = 10,000 trials). Figures 3b and 6b show individual stimulation trials from single experimental sessions. Data points in Fig. 3c show mean response over n = 20 trials for each bird; symbols identify individual birds. Figure 3d shows the mean (solid line) and standard error (shaded region) across trials. Figure 3e shows the mean (bar) and standard deviation (error bars) across animals; symbols identify individual animals.

Syllable segmentation and annotation: raw audio recordings were segmented into syllables using amplitude thresholding and annotated using semi-automated classification methods. Briefly, spectrograms were calculated for all prospective syllables, and a neural network (5000 input layer, 100 hidden layer, 3–10 output layer neurons) was trained to identify syllable types using a test data set created manually by visual inspection of song spectrograms. Accuracy of the automated annotation was verified by visual inspection of a subset of syllable spectrograms.

Syllable acoustic feature quantification and acoustic similarity: both natural song syllables and fictive vocalizations were characterized by their pitch, frequency modulation, amplitude modulation, Wiener entropy, and sound envelope—robust acoustic features that are tightly controlled in adult zebra finch song. Each feature was calculated for 10 ms time windows, advancing in steps of 1 ms, such that an estimate was computed for every millisecond. Acoustic similarity between vocal elements was calculated by pairwise millisecond-by-millisecond comparisons of acoustic features of identified vocalizations in five-dimensional space. The median Euclidean distance between points was converted into a P value based on the cumulative distribution of distances calculated between 20 unrelated songs. Thus, on this scale a similarity score of 1 means that syllables are acoustically identical, while a score of 0 indicates that the sounds are as different as two unrelated syllables.

Assessing nerve function via acoustic similarity: to assess nerve function following bilateral nanoclip implant (or sham, crush, and intact controls), we calculated the acoustic similarity of songs produced before implant with those produced following manipulation. This was quantified for each experimental group as the mean across syllables within a bird (100 renditions each of 3–6 unique syllable types) and then the mean over birds within a treatment group (n = 3 birds per group). Figure 2b shows this as the mean (solid line) and standard deviation (error bars) for each treatment group, normalized to the group mean of pre-manipulation acoustic similarity. Figure 2c summarizes post-implant changes in acoustic similarity as the mean (bars) and standard deviation (error bars) changes from pre-manipulation baseline on days 1 and 8; circles identify individual animals.

Quantifying specificity of stimulation via acoustic similarity: to quantify the specificity of current-steering stimulation, we compared the acoustic similarity of fictive vocalizations elicited within the same stimulation pattern with those produced across different stimulation patterns. (See Supplementary Fig. 7 for stimulation pattern definitions.) This is calculated as the mean over vocalizations elicited by stimulation type (i.e., within or across stimulation patterns; ~20 trials of up to 24 patterns) within a bird and then the mean over birds for each stimulation type (n = 6 birds). For comparison, the same analysis was performed on naturally produced vocalizations from unmanipulated birds (n = 6). Figure 7d shows this as the mean (bar) and standard deviation (error bars) for each stimulation type and experimental group (i.e., fictive or natural vocalizations); circles identify individual animals.

t-SNE embedding of fictive vocalizations: fictive vocalizations produced from ~20 trials of up to 24 stimulation patterns were included in this visualization. Audio recordings of each fictive vocalization, starting at the onset of stimulation and lasting for 200 ms, were converted into spectrograms (5 ms window, 1 ms advance, 512-point nfft). Spectrogram rows corresponding to frequencies above 8 kHz were discarded, and the remaining matrices (200 timesteps by 91 frequency bins for each vocalization) were transformed into 18,200-dimension vectors. The data set dimensionality was reduced to 50 with principal components analysis, and these data were subsequently embedded in two-dimensional space using t-SNE with distances calculated in Euclidean space and a perplexity of 35. Figure 7c shows this embedding in a representative bird.

Alignment of the electrophysiology recordings to song: a dynamic time warping algorithm was used to align individual song motifs to a common template60. The warping path derived from this alignment was then applied to the corresponding common mode subtracted and bandpass filtered voltage recordings (0.3–6 kHz, zero-phase, two-pole Butterworth) with no premotor time shifting. Inspired by refs. 17,60, the aligned signals were squared (to calculate signal envelope) and smoothed (Fig. 4b: 20 ms boxcar window; all other analyses: 5 ms boxcar window, 1 ms advance).

Neural activity envelope correlation: the trial-by-trial stability of singing-related nerve activity was calculated as the correlation between the song-aligned neuronal signal envelope on the 1st day of recording with those at later time points. The running correlation (Fig. 4d) shows Pearson’s correlation between the mean activity envelope of motifs on the 1st day of recording and the mean envelope in a sliding window (width: 25; advance: 1). The data points in Fig. 4e denote the mean correlation between the mean of all signal envelopes produced on day 1 and the mean of signal envelopes in a sliding window (width: 25; advance: 1) produced in a day; symbols identify individual birds.

Neural activity Vpp: the trial-by-trial Vpp of singing-related nerve activity was calculated as the difference of the maximum and minimum voltage recorded for each song motif. The data points in Fig. 4e denote the mean Vpp over all trials produced in a day; symbols identify individual birds.

Neural activity SNR: the trial-by-trial SNR of singing-related nerve activity was calculated using the formula 10 × log10(RMSS/RMSN), where RMSS and RMSN are the root mean square of the signals corresponding to singing and non-singing, respectively. Because our song-triggered recording system captures audio and electrophysiology data after singing ends, we used 500 ms segments of this vocalization-free recording as the in vivo “noise” floor. A subset of these “noise” recordings was visually inspected to confirm the absence of vocalizations or other artifacts. The data points in Fig. 4e denote the mean SNR over all trials produced in a day; symbols identify individual birds.

Neural activity event rate: the trial-by-trial event rate of singing-related nerve activity was calculated as the number of envelope threshold crossings per unit time. A unique threshold was calculated for each motif at 5 standard deviations over the mean during singing; duration of unwarped song was used to calculate rates. The data points in Fig. 4e denote the mean event rate over all trials produced in a day; symbols identify individual birds.

Comparison of on-nerve and off-nerve recordings: voltage signals recorded simultaneously during singing from on-nerve and off-nerve nanoclips were common mode subtracted (calculated independently for each device) and processed as described above. Trial-by-trial SNR for on- and off-nerve signals was calculated as described above; the data points in Fig. 5c and Supplementary Fig. 6d denote the SNR for on-nerve and off-nerve recordings for individual trials. Trial-by-trial correlation between on- and off-nerve recordings was calculated as the Pearson’s correlation of the signal envelopes; the data points in Figs. 5d and S6c denote the correlation between on-nerve and off-nerve recordings for individual trials. Data points in Fig. 5e show mean (marker) and standard deviation (error bars) for all trials in each bird (n = 154 and 57); symbols identify individual birds.

Statistical analysis

All statistics on data pooled across animals is reported in the main text as mean ± SD and depicted in figure error bars as mean ± SD, unless otherwise noted. Figure starring schema: *P < 0.05, **P < 0.01, ***P < 0.001. Where appropriate, distributions passed tests for normality (Kolmogorov-Smirnov), equal variance (Levene), and/or sphericity (Mauchly), unless otherwise noted. Multiple comparison corrected tests were used where justified. Statistical tests for specific experiments were performed as described below:

Figure 2c: changes in the acoustic structure of baseline and post-manipulation song following bilateral nanoclip implant (n = 3 birds), nerve crush (n = 3 birds), sham implant (n = 3 birds), and intact controls (n = 3 birds). A two-tailed, paired t-test revealed that the changes in acoustic similarity to baseline following nerve crush were significantly different from 0 on day 1 (P = 0.02) and day 8 (P = 0.016). Days 1 and 8 changes in acoustic similarity were not significantly different from 0 for nanoclip implants (P = 0.073 and 0.417), sham implant (P = 0.53 and 0.2786), and intact controls (P = 0.70 and 0.73). In addition, two-tailed, unpaired t-tests showed significant differences between nerve crush and intact controls at all post-surgery time points (i.e., through the 8th day: P < 0.01); no significant differences were found between nanoclip implants and intact controls at any time point (P > 0.22).

Figure 3e: comparison of stimulation evoked response amplitudes before and after lidocaine/saline application in n = 3 birds. Mauchley’s test indicated a violation of sphericity (W = 0, P = 0), and a Huynh–Feldt degree of freedom correction was applied. Subsequent repeated-measures ANOVA revealed significant differences between the treatments (F(1.11,2.21) = 189.02, P = 0.003). Post hoc comparisons using Dunnett’s test showed significant differences between saline (control) and lidocaine application (P = 6 × 10−6); no other condition significantly differed from control (P > 0.15).

Figure 4e: quantification of the stability of nXIIts dynamics, Vpp, SNR, and event rate over 30 days of continuous recording in n = 3 birds; no statistics were found to change significantly from day 1 over the duration of the experiment. A two-tailed, paired t-test revealed no significant differences in correlation between recording day 1 and days 10, 20, and 30 (P = 0.36, 0.089, and 0.179, respectively). A two-tailed, paired t-test revealed no significant differences in the Vpp of singing-related activity on day 1 and days 10, 20, and 30 (P = 0.831, 0.739, and 0.156, respectively). A two-tailed, paired t-test revealed no significant differences in the SNR of singing-related activity on day 1 and days 10, 20, and 30 (P = 0.504, 0.208, and 0.311, respectively). A two-tailed, paired t-test revealed no significant differences in the event rate of singing-related activity on day 1 and days 10, 20, and 30 (P = 0.714, 0.478, and 0.147, respectively).

Figure 5e: comparison of simultaneous on-nerve and off-nerve recordings in n = 2 birds. A two-tailed, paired t-test revealed no significant differences between mean on-nerve and off-nerve SNRs (P = 0.2567). Two-tailed, paired t-tests also showed that the mean correlation between on- and off-nerve recordings was not significantly different from 0 (P = 0.127).

Figure 7d: mean acoustic similarity of vocalizations within and across multichannel stimulation patterns (n = 6 birds) and natural vocalizations (n = 6 birds). A two-tailed, paired t-test showed that fictive vocalizations were less similar across different stimulation patterns than within repeated application of the same pattern (P = 3.5 × 10−5). A two-tailed, paired t-test showed that naturally produced song syllables were less similar across identified syllables types than within the same syllable type (P = 0.0016).

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.


Source link

Leave a Reply

Your email address will not be published. Required fields are marked *