Optimized infusion rates for N,N-dimethyltryptamine to achieve a target psychedelic intensity based on a modeling and simulation framework

Interest in new treatments for psychiatric disorders has renewed investigation of psychedelic compounds, including N,N-dimethyltryptamine (DMT). Previous population pharmacokinetic/pharmacodynamic work linked plasma DMT concentrations to a minute-by-minute subjective intensity rating on a discrete 0–10 scale and estimated an EC50 of about 92 nM. Most clinical studies to date have used intravenous bolus dosing, but because DMT has a very short half-life, continuous infusion has been proposed to extend the psychedelic experience and thus could be clinically useful. A methodological issue is that the commonly used continuous-variable pharmacodynamic models treat the 0–10 intensity score as effectively continuous and unbounded, whereas the data are discrete and bounded; bounded integer models have been proposed to respect those properties and may fit such data better. Eckernäs and colleagues set out to design an infusion protocol for DMT that targets a specified psychedelic intensity, and to examine whether choosing a continuous-variable versus a bounded-integer pharmacodynamic model materially affects dose selection. The work uses a previously published continuous-variable model together with two newly developed bounded integer variants (one including a Markov element for serial correlation) to simulate bolus-plus-infusion regimens and identify combinations that keep simulated subjects within predefined target intensity ranges over an extended period.

The analyses used data from a previously published, placebo-controlled clinical study at Imperial College London in which 13 healthy subjects (placebo on first visit, DMT on second) received intravenous bolus DMT fumarate at four dose levels. Plasma samples were taken up to 60 minutes post-dose and subjective intensity ratings (0–10) were recorded every minute for the first 20 minutes. The investigators analysed the data using nonlinear mixed-effects modelling in NONMEM v7.4.3 with Pirana and Perl-speaks-NONMEM for automation and R for diagnostics and plots. Individual pharmacokinetic parameters were estimated simultaneously with pharmacodynamic parameters, and effect-compartment models were used to describe the slight delay between plasma and effect-site concentrations. Model comparison used changes in objective function value (ΔOFV) with a −3.84 threshold for p = 0.05, and parameter precision was assessed by sampling importance resampling (5000/1000). The continuous-variable model was the previously published effect-compartment Emax (sigmoid Emax) model with E0 fixed to 0.001 and a logit transformation applied to keep predictions within the 0–10 boundaries on the transformed scale. For the bounded integer approach, the 11-category scale was modelled by dividing the standard normal distribution into 11 equal-area intervals via probit cutoffs; a normal mean and variance were modelled as functions of time, covariates and random effects, and the probability of each integer score was calculated from the cumulative normal distribution. Between-subject variability (BSV) was included on slope (drug effect) and on the variance function; a Markov element for serial correlation was also evaluated and implemented in one bounded-integer variant. Simulations incorporated BSV and residual variability to evaluate optimal bolus-plus-infusion combinations that would maintain steady ratings over an extended period (target considered over ≈60 minutes). Two targets were evaluated: a higher target of ratings 7–9 (considered desirable here) and a lower target of 4–6 for model comparison. Simulations used 1000 virtual subjects; peak effect-compartment concentrations were expected near 2 minutes after bolus so that timepoint was used for early assessment, and steady-state was defined as a time when more than five terminal plasma half-lives had elapsed in all simulated individuals. For mapping continuous-model outputs to integer targets, the investigators treated continuous predictions ≥6.5 and <9.5 as equivalent to 7–9, and ≥3.5 and <6.5 as 4–6. Once optimal doses were proposed, typical-subject time-course simulations (no variability) illustrated expected ratings over 60 minutes.

Model development and diagnostics indicated that a linear concentration–effect function best described the bounded integer data, with BSV on slope and on the variance term; including BSV in the variance improved fit. Adding a Markov element produced a substantial objective function improvement (ΔOFV = −83), but the authors also present a bounded-integer model without the Markov term to allow comparison with the continuous model that lacks a serial-correlation element. Visual predictive checks (VPCs) and parameter precision suggested broadly similar fits across the continuous-variable and bounded-integer models; individual pharmacokinetic parameter estimates were judged similar across models. Simulations identified different optimal bolus-plus-infusion combinations depending on the model and the target. For the target intensity 7–9, the continuous-variable model predicted the highest proportion within target with a 16 mg DMT fumarate bolus followed by an infusion of 1.4 mg/min. The bounded-integer models identified 14 mg bolus plus 1.2 mg/min infusion as the corresponding optimal combination. At those optimal combinations the proportion of the simulated population within the 7–9 target was low overall: about 45% for the continuous-variable model and 24%–26% for the two bounded-integer models, indicating substantial residual variability and the need for individual dose adjustments. When considering median responses across dose ranges, the continuous model predicted median 7–9 ratings for bolus doses 12–26 mg and infusion rates 1.2–2.6 mg/min, whereas the bounded-integer models produced narrower intervals (bolus 12–16 mg, or 14–18 mg with the Markov term; infusion 1.0–1.4 mg/min). At the high end of dosing the models diverged markedly in predicted extreme scores. The bounded-integer models tended to predict a much larger proportion of subjects attaining the maximum score of 10 at high doses (up to 94% reporting 10 at the highest simulated dose) compared with the continuous model (about 52% at the highest dose). For the lower target 4–6, all models converged on an optimal bolus of 10 mg with an infusion of 0.8 mg/min and similarly low proportions within target (<40%), again indicating that individually tailored dosing would be required. Typical-subject simulations (no variability) showed that the proposed bolus-plus-infusion combinations could maintain steady ratings over 60 minutes in a hypothetical typical individual.

The investigators used both a previously published continuous-variable pharmacodynamic model and two newly developed bounded-integer models to propose infusion protocols intended to achieve target subjective intensity ratings for DMT over an extended period. Eckernäs and colleagues report that, although the optimal bolus-plus-infusion combinations were similar between modelling approaches (16 mg + 1.4 mg/min for the continuous model versus 14 mg + 1.2 mg/min for the bounded-integer models to target 7–9), the predicted distribution of responses differed importantly. The bounded-integer models forecasted lower proportions within the 7–9 target at the proposed optimal doses and a higher frequency of maximum (10) ratings at high doses, whereas the continuous model predicted a higher proportion within target and fewer subjects exceeding the target range. The authors attribute these discrepancies principally to the bounded integer model’s respect for the discrete scale and boundaries, which changes behaviour near the scale endpoints. The study team notes that the Markov element improved fit to the observed data but had only minor influence on the dose-selection simulations. They emphasise that inter-individual variability is substantial: simulated proportions within target were generally low (<53% for the higher target and <40% for the lower target), implying that a single fixed regimen is unlikely to place a large fraction of subjects in the desired intensity window and that individualised titration will be needed. Potential covariates such as body weight, metabolic enzyme polymorphisms, or baseline neuropsychological factors could explain some variability, but the small sample size (13 subjects) precluded meaningful covariate analysis. Limitations acknowledged by the authors include reliance on a single small dataset drawn from subjects with prior psychedelic experience, which may underestimate intensity in the general population, and the fact that the recommended doses have not been clinically tested. They also point out modelling-structure differences that complicate a direct comparison (for example, BSV placed in the bounded-integer variance function versus in residual error for the continuous model), though sensitivity simulations without the variance BSV did not materially change the dosing conclusions. Visual predictive checks indicated that all models fitted the observed data comparably well, but the authors underline that model choice matters most near the boundaries of the rating scale. In conclusion, the paper presents dose recommendations intended as starting points for infusion studies, emphasises the need for individual dose adjustment, and highlights that bounded-integer versus continuous-variable model choice can materially influence predicted proportions above or at the maximum rating scale.