Modeling with payload inactivation captures mouse and NHP PK

The mouse PK model was collectively fit (Fig. 2A) to the active ADC, inactive ADC, and total antibody (mAb) for a total of six treatment groups: two single dose DYP688 treatment groups (3 mg/kg and 6 mg/kg), and two sets of dose fractionation groups (1.5 mg/kg Q2W & 3 mg/kg Q4W; 3 mg/kg Q2W & 6 mg/kg Q4W). The PK model was estimated with satisfactory accuracy (Table 1, with relative standard errors of all estimates < 50%). The model captures the total mAb and active ADC in an overall unbiased manner, with only a slight underestimation of the total mAb at 6 mg/kg (Figure S1 – S2) and slight overestimation of the inactive ADC in nearly all treatment groups. The NHP PK model was fit to the conjugated active payload, conjugated inactive payload, and total mAb for a total of two single dose treatment groups: 30 mg/kg and 75 mg/kg (Fig. 2B; Table 2). As with the mouse PK, the total mAb and conjugated active payload (“active ADC”) are well captured in an unbiased manner by the PK model. However, the conjugated inactive payload (“inactive ADC”) is overestimated by the model, though the discrepancy is much larger in NHPs than in mice.

Fig. 2

Population PK fitting of total antibody, active ADC, and inactive ADC in (A) mouse plasma, (B) NHP plasma. (C) In vitro stability and mass balance assessment in mouse, NHP, and human plasma

Table 1 pop-PK parameters inferred from mice for DYP688Table 2 pop-PK parameters inferred from NHP for DYP688

To further examine the source of discrepancy in the inactive ADC/conjugated inactive payload PK fitting, we evaluated the in vitro stability of DYP688 in mouse (LBA), NHP (PRA), and human plasma (PRA) (Fig. 2C). Since there is no elimination, only biotransformation in vitro, the mass balance of the total drug should be maintained i.e., total mAb/ADC should be a sum of the active ADC and inactive ADC. However, as is evident from the data, the mass balance is not conserved throughout the duration of the study. Specifically, the sum of active and inactive ADC corresponds to the amount of initial DYP688 spiked into the plasma up to ~ 48 h, but not beyond that. This suggests that additional species of the inactive ADC are formed, which cannot be detected by LBA and PRA assays developed against a specific molecular species of inactive payload (“IP-1”). Additionally, a larger fraction of the ADC is unaccounted for by 168 h in NHP plasma (~ 75%) compared to mouse plasma (~ 25%). This is consistent with the in vivo observation of higher inactive ADC overestimation in mice compared to NHP, where inactive ADC/conjugated IP-1 is calculated as [total mAb - active ADC] to maintain mass balance. Overall, since the model calibration captures the active ADC/conjugated active payload (molecular species driving efficacy) profile well in both animal species, and the overestimation of the inactive ADC is likely due to untracked ADC species (all likely inactive), the model calibration was deemed fit for purpose.

Additionally, the in vitro inactivation half-life (calculated as loge2/ki.a.) in mouse plasma (39–42 h, Figure S3) was consistent with that estimated from the in vivo PK model (~ 52 h), but the NHP values diverged. In vitro NHP plasma stability indicated an inactivation half-life of ~ 17 h, while fitting from in vivo PK model estimated a half-life of ~ 61 h. This difference was not due to parameter identifiability issues – indeed forcing the inactivation half-life to a constant value of 17 h did not yield good fits.

PK-TGI modeling of mouse tumors indicates both Cmax and exposure driven efficacy

The Gompertz growth model described the vehicle-treated tumor growth kinetics well. Keeping the population mean and variance estimates of the tumor growth rate (β) and PK parameters constant (but linked to individual animal IDs for the latter), the impact of four DYP688 treatment groups from a single study was evaluated using a linear killing, one-transit compartment TGI model. Parameters from the model are listed in Table 3. Tumor responses within each treatment group of the 92.1 CDX tumor model was heterogeneous, though the individual data (Figure S4) and overall trends in the data were captured well by the median and 90% prediction quantiles (Fig. 3A). Interestingly, the estimates of tumor killing rate constant (kkill) in individual mice showed a skewed distribution and co-varied with dose (Figure S4). More complex models, including those with different reaction schemes for non-linear killing effects, did not further improve the predictability of the model, certainty of estimates, or eliminate the dose covariate on killing rate constant. We additionally tested a tumor growth modulation model where tumor growth rate influences tumor killing (Figure S5). While this model did reduce the dose covariate on the killing rate constant, the predictions considerably worsened. Since the goal was to capture the observed tumor kinetics in all dose groups, the initial TGI model with the dose-covariate was retained and used for further simulations. Though not directly confirmed, it is likely the kkill-dose covariance arises from the physiological response of 92.1 cells to FR, wherein treatment with increasing amounts of FR resulted in re-differentiation of 92.1 cells, adopting a “flat” morphology, growing slower than the parent cells, and demonstrated increased melanosome content [18]. This is consistent with the observations of “flat” tumors (tumors with flattened, disc like morphology with immeasurable height) with much slower regrowth was prominent in 3 mg/kg and 6 mg/kg groups with increasing prevalence. Note, the ‘flat’ morphology is not universally observed in all melanomas. Indeed, published accounts of melanoma cell lines treated with FR indicate that the dark/flat morphology is more commonly observed with Class 1 melanoma cell lines (which includes 92.1), while Class 2 melanomas which have a higher metastatic potential, do not necessarily exhibit this morphology upon treatment with FR [18,19,20,21]. In-house evaluation of DYP688 with a PDX tumor model did not exhibit the appearance of the flat morphology even at 5x higher doses than in the 92.1 model (data not shown), indicating this morphological change may not be universally applicable to patients, particularly those with metastases. Since the volume calculation for such tumors was difficult, they were censored as below limit of quantification (‘BLOQ’) in the dataset during model fitting. Additionally, the fitted PK (Figure S6, Supplementary Table 2) of a mutant version of DYP688 that does not bind to FcRn receptors and thus exhibits rapid clearance (‘FcRn-mut’) was performed. Validation of the PK-TGI model developed using 92.1 CDX tumor data was performed by overlaying the predicted PK and/or efficacy profiles of DYP688 (Figure S7A) and its ‘FcRn-mut’ variant (Figure S7B) to observed data from an independent study not employed for the calibration. Predictions agreed qualitatively well with the independent datasets, highlighting reliability of PK-TGI model for further exploratory analysis.

Table 3 92.1 CDX tumor growth inhibition model parameters inferred from miceFig. 3

(A) Population PK-TGI fitting of tumor growth inhibition data in 92.1 CDX tumors, (B) PK-TGI correlation analysis identifying exposure above tumorostatic concentration (TSC) as the PK-driver of efficacy

Using the validated PK-TGI model developed using 92.1 CDX data, dose fractionation simulations were performed and different PK metrics, namely peak concentration (Cmax), exposure (AUC), and time over tumorostatic concentration, were correlated to the predicted tumor growth inhibition (Fig. 3B). Tumorostatic concentration (TSC) is defined here as the conjugated active payload concentration at which achieves tumor stasis and is calculated according to the equation below. Note, since PK-TGI model based on the 92.1 tumor model (CDX) had dose as a covariate on kkill, we selected the smallest kkill i.e., largest TSC (Table 4) as a conservative estimate for subsequent analysis.

Table 4 Killing rate constant and tumorostatic concentration (TSC) inferred from 92.1 CDX TGI model

All three metrics only showed limited/incomplete correlation to TGI response. For example, Cmax showed reasonable correlation with response with DYP688; however, a confirmatory assessment with equivalent doses of ‘FcRn-mut’, the Fc-mutant variant (i.e., same Cmax, different AUC) breaks the correlation, indicating that Cmax alone does not predict response. However, further assessment of AUC showed that AUC-matched groups with lower Cmax showed less response (e.g., 3 mg/kg Q2W vs. 6 mg/kg Q4W). Time over TSC showed a similar incomplete correlation pattern. To account for both Cmax, AUC, and time over TSC dependent response, we calculated a composite metric denoted exposure above TSC (AUCTSC). AUCTSC captures the improved response from 1.5 to 3 to 6 mg/kg, as the peak concentration progressively surpasses the TSC, while the concentration stays above the TSC longer with higher Cmax or more frequent dosing, thereby improving response. Therefore, our analysis indicated that AUCTSC is a good PK surrogate for TGI response, establishing it as a reliable metric for human dose translation and optimization. Additional mechanistic exploration of the AUCTSC-driven response and related impact of DYP688 tumor distribution, antigen binding, intracellular payload accumulation/binding on anti-cancer activity is discussed in the following section.

Mechanistic modeling highlights AUCTSC-driven response attributable to payload mechanism

Tumor response driven partially by the peak ADC concentration may appear atypical for ADCs, as most are assumed to exhibit cumulative exposure-driven response. Cmax-driven response has previously been hypothesized as an outcome when the tumor penetration of the ADC is heterogeneous i.e., tumor cells in the furthest edge of the tumor do not encounter any ADC due to the binding site barrier effect [22]. To assess if tumor penetration depth limitations explain the partial Cmax driven response of DYP688, we performed mechanistic Krogh cylinder simulations (Fig. 4A) based on a previously validated model [15]. The Krogh cylinder model simulations are fully predictive and not fit to any data. Parameters for the model are listed in Supplementary Table 1. In particular, PMEL17 expression was determined to be moderate (~ 100,000 receptors/cell), based on in-house immunohistochemistry (IHC) scoring of ~ 1.5–2 (data not shown, [23]). Immunohistochemistry detects both surface and intracellular PMEL17, and literature evidence that suggests that up to 45% of the total cellular PMEL17 can be present on the cell surface [24]. Assessment of conjugated active payload (ADC species driving efficacy) tumor penetration for 100,000 receptors per cell shows that at the peak concentration, some active ADC reaches the edge of the tumor, even at 1.5 mg/kg (Figure S8A). This indicates that the Cmax-driven response with DYP688 is likely not due to the canonical hypothesis that increasing Cmax improves tumor penetration of the ADC, and therefore efficacy. Even, at the extreme limit of IHC 2 expression of 500,000 receptors per cell, active ADC at 1.5 mg/kg is predicted to target cells at the edge of the tumor (R = 100 μm). This suggests that Cmax-driven improvement in efficacy at higher dose levels of DYP688 is likely not related to better tumor penetration.

Fig. 4

(A) Graphical schematic of Krogh cylinder model incorporating tumor uptake of three ADC species deriving from plasma inactivation of DYP688 (2 active payloads, one active payload, and no active payloads), competitive binding of the ADC species to surface receptor (PMEL17), and internalization of the receptor-ADC complex. Payload inactivation was assumed to not occur in the tumor microenvironment, based on experimental observations. In the lysosome, the active payload is released either from linker cleavage or antibody backbone degradation, and the amount of active payload release is dependent on the corresponding DAR of each ADC species. For simplicity, the presence of any inactive payload was considered functionally equivalent to no active payload (e.g. unconjugated antibody formed from release of both active payloads was considered equivalent to an ADC with two inactive payloads and pooled in the same state variable). The released active payload permeates from the lysosome to the cytoplasmic space, where it could either reversibly bind Gαq/11 or escape the cell, diffuse into adjacent cells, and mediate bystander Gαq/11 inhibition. (B) Overlay of predicted fraction of free Gαq/11 (lines) with the observed pERK score (symbols) showed the model captures the kinetics of DYP688 anti-cancer activity well. (C) Predicted fraction Gαq/11 free at peak inhibition showed good correlation between time-independent tumor growth inhibition at different doses over a 14-day period (single administration). Time-independent tumor growth inhibition is presented as area-under-curve of tumor volume ratio (Figure S9), normalized to the area-under-curve of untreated group

We further assessed the mechanism behind the AUCTSC-driven response by simulating the payload activity in the cell. Payload release into the lysosome was modeled as linker cleavage based on the known stability of the valine-citrulline linker [25] and payload escape, bystander effects, and intracellular Gαq/11 binding was adapted based on the previously established Krogh cylinder model for cytotoxic agents [15]. In addition to using FR and Gαq/11 specific parameters in the model (Supplementary Table 1), Gαq/11 target turnover was also included in the model. The predicted radial average and standard deviation of available fraction of intracellular Gαq/11 (“fraction Gαq/11 free”) was compared to the pERK score for three dose levels – 1.5 mg/kg, 6 mg/kg, and 12.5 mg/kg DYP688 (Fig. 4B). A dose of 1.5 mg/kg delivers an intracellular FR concentration sufficient inhibition only ~ 50% Gαq/11, consistent with only partial pERK downregulation. At 12.5 mg/kg, pERK downregulation was maximized and more durable, consistent with the predicted Gαq/11 inhibition. The peak Gαq/11 inhibition at 6 mg/kg was also consistent with maximum pERK downregulation, though the rebound of Gαq/11 levels appears to occur faster than pERK rebound. Overall, the trends in predicted Gαq/11 inhibition corresponds with the observed trend in pERK scores, and 90% Gαq/11 inhibition (i.e., 10% free Gαq/11) was approximated as the threshold at which growth signaling rebounds. Furthermore, we compared the predicted peak fraction Gαq/11 free to the observed tumor growth inhibition for different doses of DYP688 and a high dose of the FcRn-mut variant (Fig. 4C). A strong correlation was observed, with > 90% Gαq/11 inhibition correlating with tumor stasis and > 99% Gαq/11 inhibition correlating to durable tumor regression. These observations are consistent with the hypothesis that increasing Cmax is driving better tumor saturation rather than tumor penetration [22]. The detailed spatial profile of Gαq/11 inhibition (Figure S8A) highlights that the intra-dose radial range of peak Gαq/11 inhibition is relatively uniform at 100,000 PMEL17/cell (i.e. no tumor penetration limitations), but undersaturated at 1.5 mg/kg (40–60%) and increases proportionately with dose. Correlations between AUCTSC for a single administration and peak receptor occupancy (i.e., proxy for tumor penetration, Figure S8B) or average Gαq/11 inhibition over a 7-day period (i.e., proxy for tumor saturation, Figure S8C) further highlights a poor correlation to tumor penetration and better correlation to tumor saturation. Detailed PK, Gαq/11, and tumor growth inhibition profiles for different dosing schemes are shown in Figure S9.

Overall, these results provide a mechanistic explanation as to why both Cmax and AUC correlate to tumor response. Achieving a Cmax above the TSC is critical for ensuring the minimum threshold of Gαq/11 inhibition which drives downstream tumor cell growth inhibition (indicated by pERK). Progressively higher doses serve to (1) achieve maximum inhibition of the target, leading to no downstream growth signaling, and (2) prolongs > 90% inhibition of Gαq/11, driving drastically reduced downstream signaling for longer durations. Furthermore, at doses that exceed the TSC, more frequent dosing allows more sustained inhibition of Gαq/11, and therefore downstream growth signaling. Thus, the AUCTSC-driven response for DYP688 is attributed to the ‘targeted’ mechanism of action of the FR payload, which is different from the tumor payload-exposure mechanism of action of most non-specific cytotoxins [26] utilized on all currently approved ADCs.

DYP688 elimination and inactivation rates are well-matched to optimize payload activity

Simulations presented in the previous sections specifically focused on analyzing the behavior of DYP688, accompanied by a mechanistic understanding of its pharmacology. In this section, we use the above validated DYP688 Krogh cylinder model to extrapolate generalized learnings to aid in the design of targeted-targeted ADCs that exhibit linker/payload metabolism in circulation. Note, the following analysis is not limited just to payload inactivation – it is also applicable to common ADC pharmacology such as payload deconjugation.

One of the key findings of the semi-mechanistic systemic PK-TGI modeling of DYP688 showed that the payload inactivation rate (ki.a.) and the antibody elimination rate (kel) are approximately the same in both mice and NHPs. However, it is not necessary that these two rates are always matched to enable sufficient exposure of active payload. We therefore used the PK-TGI model to explore the impact of designing ADCs with varying biotransformation (payload inactivation, deconjugation, etc.) half-life and elimination half-life. A heat map of the impact of biotransformation versus elimination on the predicted preclinically efficacious exposure (AUCTSC) is shown in Fig. 5A. Compounds that lie further to the right of the plot represent progressively more stable ADCs, thus a longer half-life of the ADC (moving up the plot) progressively approaches the maximum effective ADC exposure at any given dose. Compounds that lie further left of the plot represent progressively decreasing stability, so slowing the clearance half-life (moving up the plot) does not significantly change preclinically efficacious exposure of the conjugated active payload. The elimination half-life of DYP688 is near-ideally matched to the inactivation half-life, such that the ADC species begin to clear significantly just as they start to inactivate (i.e., do not contribute to efficacy). For a hypothetical DYP688 with much faster elimination (hDYP688-fast, blue star), the conjugated active payload is prematurely eliminated, and would require either a higher dose or more frequent dosing to achieve preclinically efficacious exposure. On the other hand, a hypothetical DYP688 with much slower elimination (hDYP688-slow, red star) does not significantly increase predicted preclinically efficacious exposure, as most of the payload has already inactivated. In this scenario inactive ADC would be circulating unnecessarily, creating greater risk for undesired side effects such as anti-drug-antibody (ADA) development. Therefore, DYP688 lies on the sweet-spot of the heat map where there is just enough exposure to achieve sufficient anti-cancer activity at low-to-moderate doses.

Fig. 5

(A) Heatmap of predicted preclinical ADC efficacious exposure as a function of antibody backbone elimination half-life and payload inactivation half-life. Efficacious exposure is reported as fold-factor of exposure above threshold (AUCTSC) estimated for DYP688. (B) DYP688 (black circle) and two hypothetical ADCs exhibiting same payload inactivation half-life as DYP688, but faster antibody elimination (hDYP688-fast, blue star) or slower antibody elimination (hDYP688-slow, red star) were simulated using the Krogh cylinder tumor model. The stoichiometric ratio of active ADC to inactive ADC at steady state active ADC Cmax, surface PMEL17 occupancy at active ADC Cmax, and fraction Gαq/11 available at peak inhibition were predicted for each ADC design

The above exploratory analysis was based purely on plasma exposure. In this section, we compare the plasma pharmacokinetics, tumor PMEL17 occupancy, and predicted free fraction Gαq/11 (corresponding to efficacy, Fig. 4C) of DYP688 (tel ~ ti.a. ) and the two hypothetical compounds shown on the heatmap – DYP688 but with faster antibody elimination (‘hDYP688-fast’, tel < ti.a.) and DYP688 but with slower elimination half-life (‘hDYP688-slow’, tel > ti.a.). First, if we assume that no inactivation occurs, the total antibody (tmAb, dashed line) curve represents the hypothetical maximum ADC exposure, and corresponding tumor receptor occupancy and Gαq/11 activity (Figure S10). As expected, slower elimination yields more sustained tumor receptor occupancy and Gαq/11 inhibition. When the effect of inactivation is added in (same for all three compounds), an anticipated decrease in plasma ADC concentration, less tumor receptor occupancy, and less Gαq/11 inhibition (solid line) is predicted at the first dose compared to the no inactivation scenario.

For hDYP688-fast, elimination occurs faster than inactivation, leading to no accumulation of the inactive ADC species. Thus, the steady state profiles look identical to the first dose profile (Figure S10). In this scenario, the steady state plasma stoichiometry of active ADC to inactive ADC is extremely high at every new dose administration (containing 99% active ADC), consistently driving maximum binding of the active ADC to PMEL17 and maximum inhibition of Gαq/11 in tumors (Fig. 5B). For hDYP688-slow elimination occurs slower than inactivation, leading to significant accumulation of the inactive ADC species at steady state (Figure S10), to the extent that steady state plasma stoichiometry of active ADC to inactive ADC reduces to 1 at every new dose administration. This creates significant PMEL17 binding competition for the active ADC in the tumor, which in turn reduces payload accumulation in the cell, driving less Gαq/11 inhibition at steady state compared to the first dose (Fig. 5B). The design of DYP688 that near-perfectly matches the antibody elimination half-life to the inactivation half-life leads to minimal accumulation of the inactive ADC at steady state (Figure S10). While this results in some reduction in the peak plasma stoichiometry, it is still several-fold larger than hDYP688-slow. Thus, there is minimal competition for PMEL17 binding to the active ADC, and peak Gαq/11 inhibition remains strong and durable compared to the first dose (Fig. 5B). Thus, the design of DYP688 is ideally optimized for payload activity.

Predicted human PK parameters consistent with observed clinical PK

Human PK parameters were extrapolated by single species weight-based allometric scaling from cynomolgus monkey PK parameters and are listed in Table 5. Volume (V) and clearance (CL) scaling was based on the findings of Li et al. that showed from a retrospective analysis of 11 ADCs that an exponent of 1 when scaling from cynomolgus PK data alone can sufficiently predict human PK for ADCs exhibiting linear PK [12]. For the inactivation rate (ki.a.), the in vivo estimate of ~ 61 h was assumed the more likely value for parameter scaling. Based on the similar inactivation rates between NHP and human in vitro (Fig. 2C), human inactivation rate was considered the same as NHP inactivation rate. The predicted human PK closely matches the observed interim clinical PK (cut-off date 8 Sept 2023) [27], as demonstrated by the cycle 1 comparison for the 8 mg/kg dose group (Fig. 6), highlighting that the reliability of the preclinical to clinical translation employed, including the non-conventional payload inactivation kinetics.

Table 5 Allometric scaling of PK parameters from NHP to humanFig. 6

Interim clinical PK data (cut-off date 8 Sept 2023, 12 patients) for total antibody (black) and conjugated active payload (blue) overlayed with anticipated human PK scaled allometrically from NHP demonstrates reliability of the scaling method applied, particularly for the payload inactivation. Anticipated PK simulations performed assuming body weight of 70 kg and a 1.5-hour infusion duration, based on the median interim clinical infusion time for the displayed cohort