TAK-653

Chemical identity

TAK-653 is not a peptide, although the secondary literature catalogues and discusses it alongside them. Beyond that, its identity is unambiguous. PubChem holds a single compound record under the name: CID 56655833, molecular formula C19H23N3O3S, molecular weight 373.5 g/mol, InChIKey PXJBHEHFVQVDDS-UHFFFAOYSA-N. The FDA substance registry carries the same structure as UNII 9E3TOE5RIZ, with CAS 1358751-06-0 as the primary registry number and INN number 12689, the USAN code MN-277 and PubChem CID 56655833 recorded alongside it. ChEMBL4594403 repeats the formula and molecular weight and records no first approval date. The registry describes the structure as achiral.

Three names attach to this one structure. TAK-653 is the Takeda development code, NBI-1065845 is the code used by Neurocrine Biosciences after the compound was licensed, and osavampator is the International Nonproprietary Name. All three appear as synonyms on the same PubChem CID and the same FDA substance record, and the trial registry entry for the phase 2 study lists TAK-653 explicitly as another name for NBI-1065845. Papers, registry entries and conference material published between 2015 and 2025 use whichever code was current at the time, which makes the literature look larger and more fragmented than it is.

The systematic name is 9-[4-(cyclohexyloxy)phenyl]-7-methyl-3,4-dihydropyrazino[2,1-c][1,2,4]thiadiazine 2,2-dioxide. Suzuki and colleagues describe it as a deliberate modification of an earlier compound from the same laboratory, TAK-137, with a cyclohexyl group substituted for a terminal phenyl specifically to increase steric bulk. The stated reason was to induce steric repulsion at Ser750 in the ligand-binding domain of GluA2o, the residue that group had previously identified as governing how much agonist-independent activation this chemical series produces.

The agonism problem the molecule was built around

Earlier AMPA receptor potentiators carried two linked liabilities: they activated the receptor to some degree in the absence of glutamate, and they produced bell-shaped dose-response curves and convulsions at higher exposures. Suzuki and colleagues set out the design premise directly, arguing that the bell-shaped curve is inconsistent with a desensitisation explanation because the same compounds induce seizures at the doses where the curve turns down. The whole preclinical package for this molecule is organised around one question: whether an AMPA receptor potentiator can be made to work only when glutamate is already present.

Their 2021 report gives the binding measurements. TAK-653 displaced the radiolabelled potentiator [3H]-HBT1 from a His-tagged GluA2o ligand-binding domain with an IC50 of 0.26 μM (n = 4). Binding of radiolabelled TAK-653 to the same protein rose in a glutamate-dependent manner and was not detected against an unrelated control protein. It did not displace [3H]-AMPA from the agonist site; it mildly increased that binding instead, with an EC50 of 1.5 ± 0.2 μM and an Emax of 19.9 ± 3.4%. In CHO cells expressing GluA1i, glutamate-dependent calcium influx was potentiated with an EC50 of 3.3 μM. When Ser743 — the GluA1i equivalent of GluA2o Ser750 — was mutated to alanine, the maximum response to TAK-653 was higher than in wild-type at every glutamate concentration tested, which the authors read as removal of the steric interference the molecule was designed around (PMID 34267258).

The convulsion data come from Good Laboratory Practice studies in male Sprague-Dawley rats dosed orally at 100, 15 and 5 mg/kg in one study and 50, 15 and 5 mg/kg in a second. Convulsions were recorded in one animal at 100 mg/kg, four hours after dosing. From that observation the authors calculated exposure margins of 419-fold on plasma Cmax and 1017-fold on plasma AUC, against comparator margins of 42 and 122 for TAK-137 and 3.4 and 4.0 for LY451646. The denominator matters and is often dropped: the margin is measured against the rat exposure that improved novel object recognition, not against any dose given to a person.

What the animal work measured

Cognitive readouts in rats came from two paradigms. In the novel object recognition test in male Long-Evans rats, oral doses of 0.03, 0.1 and 0.3 mg/kg raised the novelty discrimination index against vehicle, with ten animals per group. In a radial arm maze test in which working memory was disrupted by MK-801 at 0.08 mg/kg subcutaneously, oral doses of 0.1, 0.3, 3 and 10 mg/kg reduced mean errors, with group sizes reported as 6 to 18. Two comparator potentiators, LY451646 and PF-04958242, improved the recognition task but not the maze task in the same report, which is the dissociation the paper was written to demonstrate.

Primate data in that report are thin by design. Three male cynomolgus monkeys received 0.06 mg/kg orally six hours before a delayed match-to-sample session of 96 trials, a dose chosen because it produced a Cmax matching 0.1 mg/kg in rats. Accuracy improved at the 16-second delay interval on a paired t test and did not reach significance at the 0, 4 or 8-second intervals. Three animals and one significant cell across four delay conditions is a small result, and the paper presents it as one panel among many rather than as a standalone finding.

Two studies address depression-like behaviour. Hara and colleagues reported in 2021 that ketamine at 30 mg/kg intraperitoneally produced an antidepressant-like effect in the rat reduction of submissive behaviour model, that pretreatment with the AMPA antagonist NBQX at 10 mg/kg blocked it, and that six days of TAK-653 produced a comparable effect without the hyperlocomotion ketamine caused; the TAK-653 dose and the group sizes do not appear in the retrievable abstract and the full text is subscription-only. Li and colleagues reported in 2025 on three male cynomolgus monkeys taken through twelve weeks of chronic unpredictable mild stress and then two weeks of oral TAK-653 at 0.346 mg/kg, a dose derived by body-surface-area scaling from a rat figure. There was no control group and no separate vehicle arm; each animal served as its own baseline, the compound was purchased from a commercial chemical supplier, and the authors state plainly that the sample size prevented any dose-response assessment.

Three studies, and one cohort of twenty-four

The first-in-human work is registered as NCT02561156 and ran at a single site in the United Kingdom from August 2015 to September 2017 in 88 participants. Single doses of 0.3, 1, 3, 5, 9 and 18 mg were given in part one; part two gave 0.3, 1, 3, 6 and 9 mg once daily from day 6 to day 18, six participants per cohort against pooled placebo groups of 12 and 10. The posted results record no serious adverse events in any arm, no clinically significant electroencephalogram abnormality in either part, no treatment-emergent suicidal ideation or behaviour on the Columbia scale, and one discontinuation for an adverse event across the whole study. Treatment-emergent adverse event rates did not track dose: 83.3% at 5 mg against 33.3% at both 9 and 18 mg, on six participants per cell.

O'Donnell and colleagues published the translational transcranial magnetic stimulation work in 2021. Thirty-one adult male Sprague-Dawley rats received TAK-653 by oral gavage at 0.1, 0.3, 1, 8 or 50 mg/kg or vehicle; mechanomyographic response amplitude showed a significant effect of dose, F(5,25) = 4.399, p = 0.005, with all doses except 0.1 mg/kg producing responses 30 to 70% above vehicle. The human arm enrolled 24 healthy volunteers, 23 men and one woman, in a four-period design: three double-blind periods (placebo, 0.5 mg and 6 mg) followed by an open-label ketamine 0.5 mg/kg intravenous period, which four of the 24 did not take part in. Motor-evoked potential amplitude at 2.5 hours was higher on 6 mg than placebo (p = 0.0269, Dunnett-adjusted p = 0.0497). Resting motor threshold did not change at either dose, and of the paired-pulse measures only long intracortical inhibition at a 300 ms interval on the 0.5 mg dose reached significance.

Dijkstra and colleagues reported the pharmacodynamic battery run in that same cohort, with 90% confidence intervals throughout. At 6 mg, adaptive tracking improved by 1.68% (90% CI 0.51 to 2.84, P = 0.02) and smooth pursuit by 2.32% (0.37 to 4.27, P = 0.05); the saccadic peak velocity change was 15.40 degrees per second (1.91 to 28.90) and did not reach significance (P = 0.06). The 0.5 mg dose produced the study's only significant saccadic effect, 19.49 degrees per second (5.98 to 32.99, P = 0.02), while also increasing the Stroop congruent-incongruent reaction-time difference by 22.0 ms (P = 0.05) and reducing the number of correct responses in incongruent trials (P = 0.02); 6 mg affected no Stroop parameter. Body sway and every subjective visual-analogue subscale were unchanged, and no dissociative or euphoric effects were recorded. Somnolence was reported by 3 of 24 at each dose against 2 of 24 on placebo, and headache by 4 of 24 at 6 mg against 2 of 24 on placebo.

SAVITRI, and two versions of the same result

The phase 2 trial is NCT05203341, run under the acronym SAVITRI between February 2022 and February 2024. Adults aged 18 to 65 with recurrent major depressive disorder or persistent depressive disorder, an inadequate response to antidepressant treatment and a 17-item Hamilton score of at least 22 were randomised 2:1:1 to placebo, 1 mg or 3 mg once daily for eight weeks, on top of the antidepressant they were already taking. One hundred and eighty-three were randomised and 161, or 88%, completed treatment. The primary endpoint was change in total MADRS score at day 28, with day 56 among the secondary endpoints; the poster labels every p-value it reports, day 28 included, as nominal. The registry entry itself never names the two dose levels; it calls them low and high.

Two sponsor accounts of the outcome exist and they do not match. The April 2024 press release reported, for the dose that separated from placebo, a least-squares mean difference of -4.3 at day 28 (p = 0.0159) and -7.5 at day 56 (p = 0.0016), with the other dose at -3.0 (p = 0.0873) and -3.6 (p = 0.1082). The poster presented at Psych Congress in 2025 identified the doses as 1 mg and 3 mg and reported the 1 mg arm at -4.3 at day 28 (95% CI -7.8 to -0.8, nominal p = 0.0159) and -6.9 at day 56 (95% CI -11.2 to -2.6, nominal p = 0.0018, effect size 0.73), with the 3 mg arm at nominal p = 0.0573 at day 56. Day 28 is identical across the two accounts. Day 56 differs on both arms, and nothing in the retrievable material explains why.

Whichever figures are taken, the lower dose separated from placebo and the higher dose did not. No published analysis addresses why. No results have been posted to the registry, and no peer-reviewed publication of the trial exists: a PubMed search for the compound name returns ten records and a Europe PMC title-and-abstract search returns eleven, none of which is this study. Five phase 3 trials are now registered — three efficacy studies of roughly 200 participants each, a 550-participant maintenance study and an 850-participant long-term safety study — with completion dates between 2027 and 2030. None of the registry entries states the phase 3 dose.

The interaction study, and the legal position

One further human study has been published in full, and it carries no trial registry record at all. Lin and colleagues ran a parallel-arm phase 1 in 2024 testing whether repeated dosing induces CYP3A. The midazolam arm comprised 14 men and 4 women, of whom 16 completed; the oral contraceptive arm comprised 17 women, of whom 16 completed. After multiple daily doses, geometric mean ratios with 90% confidence intervals were 0.94 (0.79 to 1.13) for midazolam Cmax and 0.88 (0.78 to 0.98) for midazolam AUC to infinity; 1.00 and 1.01 for ethinyl estradiol; and 0.99 for levonorgestrel Cmax with 0.87 (0.78 to 0.96) for AUC to the last quantifiable concentration. The authors concluded that the compound is not a CYP3A inducer.

Regulatory status is straightforward. ChEMBL records no first approval date, phase 3 trials began recruiting in 2025, and there is no marketing authorisation in any jurisdiction. Every registered study naming the compound has an industry sponsor: eight from Neurocrine Biosciences and one from Millennium Pharmaceuticals, a treatment-resistant depression study with a February 2018 start date that was withdrawn without enrolling anyone. The material sold on vendor product pages and nootropics aggregator sites has no relationship to the tablets manufactured for these trials, and nothing established in the trials transfers to it.

The published human exposure is also short. The longest characterised repeat-dosing interval in any published or registry-posted record is thirteen consecutive days in part two of the first-in-human study. The eight-week exposure in the phase 2 has produced no published safety table, no adverse-event listing beyond the sponsor's statement that there were no deaths, serious adverse events or adverse events of special interest, and no posted registry results.

What is not known

The largest dataset on this molecule is not public. SAVITRI randomised 183 people, completed in February 2024, and has posted no results to ClinicalTrials.gov and produced no peer-reviewed publication; everything known about it comes from a sponsor press release and a conference poster, which disagree with each other on the day 56 figures for both dose arms, and the day 56 comparisons are unadjusted secondary-endpoint p-values that the poster itself labels nominal. Human pharmacokinetics are in the same position: the half-life range every secondary source repeats traces to a 2018 conference abstract that is not indexed in PubMed, and the registry entry for the study that generated it posted no half-life outcome. Nothing has been published on exposure beyond thirteen days of daily dosing, on what happens after discontinuation, on pharmacokinetics in older adults, in hepatic or renal impairment, or across ancestry groups. The phase 1 cohorts were six participants per dose level and the two pharmacodynamic studies used one group of 24 volunteers, 23 of them men, so the human physiological data rest on a single small and almost entirely male sample. One published human study, the CYP3A interaction trial, carries no trial registry record at all. Seizure risk in people is uncharacterised: no convulsion has been reported in any registered trial, but the total human exposure across all published studies is a few hundred participants and no study was powered for a rare event. On the preclinical side, the primate cognition finding rests on three animals and one significant delay interval, the primate depression study on three animals with no control group, and no independent laboratory has replicated the founding pharmacology, which comes entirely from the originating company.

Questions