Toward Rational Design of Orally Bioavailable PROTACs: Development of a Fully Computational Oral Absorption Score

Abstract

Achieving adequate oral absorption for proteolysis-targeting chimeras (PROTACs) remains challenging due to their size, polarity, and conformational complexity. While existing approaches often rely on physicochemical property space or chameleonicity-driven models, a generalizable framework for guiding compound design prior to synthesis is lacking. Here, we introduce comPASS, a fully computational scoring function derived from molecular dynamics-based descriptors that is designed to capture the balance between molecular extension and exposed polarity through normalized radius of gyration and three-dimensional polar surface area of PROTACs. Across mouse and rat data sets, comPASS shows consistent predictive performance and may facilitate prioritization of compounds at early design stages. We further show that within the data sets studied, extended molecular geometries are associated with improved absorption, when increased elongation is not driven solely by increases in molecular size. To assess the potential benefit of incorporating experimental information, we also developed an enhanced variant, ePASS, which incorporates experimental polar surface area (EPSA) and lipophilicity terms consistent with prior approaches and shows promising predictive performance relative to similar literature methods. Together, these results establish practical guidance for presynthetic prioritization of PROTACs and provide a potential framework for guiding the design of orally absorbed PROTACs.

Proteolysis-targeting chimeras (PROTACs) are bifunctional molecules that enable degradation of target proteins by recruiting an E3 ubiquitin ligase to the protein of interest, leading to ubiquitination and elimination through the proteasome. This mechanism of action distinguishes them from inhibitors and enables modulation of targets that are otherwise difficult to address with conventional approaches. − Recently, the FDA approval of vepdegestrant (VEPPANU), the first PROTAC therapeutic, underscores the growing clinical validation and promise of this approach.

PROTACs typically fall outside the physicochemical property space of historical orally bioavailable small molecule drugs, with higher molecular weight, increased polarity, and high conformational flexibility hindering absorption and permeability. As a result, traditional guidelines such as Lipinski’s Rule of 5 are not directly applicable, highlighting the need for alternative frameworks to evaluate and optimize properties governing oral exposure for this class of compounds. − These challenges have motivated multiple approaches to better understand and predict absorption behavior of PROTACs.

Early efforts to rationalize oral absorption of PROTACs have, among other strategies, focused on defining a favorable physicochemical property space, where parameters such as topological polar surface area (TPSA), molecular weight (MW), number of rotatable bonds (RB), and counts of hydrogen bond donors (HBD) and acceptors (HBA) are constrained to achieve compounds with improved fraction absorbed. For example, in our previous work, systematic analyses of large internal data sets proposed empirical boundaries in these descriptors to increase the likelihood of oral exposure in rats. These boundaries define a region beyond which oral absorption is unlikely to be achieved. However, within this defined descriptor space, many compounds still show poor absorption despite satisfying all criteria, highlighting the need for additional design principles beyond simple property filters to reduce the number of poorly absorbed compounds. A recent review has emphasized the importance of balancing these properties when navigating beyond-Rule-of-Five chemical space.

Within this broader framework, multiple perspectives have emerged regarding the determinants of oral absorption for PROTACs. One perspective regarding PROTAC permeability emphasizes chameleonicity, wherein molecules can dynamically adapt their conformation to aqueous and hydrophobic environments, reducing exposed polarity through intramolecular interactions to facilitate membrane permeation and unmasking to maintain aqueous solubility. − In this context, descriptors such as three-dimensional polar surface area (3D PSA) and radius of gyration (R _g) have been used to capture environment-dependent changes in molecular shape and polarity, providing means to quantify conformational ensembles and the extent of polarity shielding. − In contrast, another perspective highlights the importance of molecular shape, proposing that more elongated, rod-like PROTACs may permeate membranes more efficiently without requiring extensive conformational collapse. Consistent with this view, incorporation of molecular length descriptors such as SMID, the smallest maximum intramolecular distance, into multiparameter indices has been shown to correlate with improved oral absorption, introducing rod-like structure as a key driver. Furthermore, systematic evaluation of experimental and calculated chameleonicity descriptors across large PROTAC data sets has not revealed a consistent relationship with oral absorption, instead identifying factors such as efflux as more influential determinants. Together, these perspectives highlight that because multiple factors contribute to oral absorption, optimization within a defined physicochemical property space may involve contributions from molecular shape, polarity, lipophilicity, polarity shielding, and other mechanisms depending on the physicochemical class of the compound.

For orally administered compounds, systemic exposure is determined by the extent of absorption and metabolism. The fraction absorbed (f _a) describes the proportion of drug that crosses the intestinal epithelium, while f _g represents the fraction that escapes metabolism in the gastrointestinal tract. The hepatic component (f _h) accounts for the fraction of drug that survives first-pass extraction in the liver and can be approximated from systemic clearance (Cl) and hepatic blood flow (Q _h) as

$$f_{\mathrm{h}}=1-\mathrm{Cl}/Q_{\mathrm{h}}$$ 1

The combined term f _a × f _g therefore represents the fraction of an oral dose that reaches hepatic portal circulation, providing an estimate of a compound’s absorption performance. Oral bioavailability (F) incorporates all three contributions and can be expressed as

$$F=f_{\mathrm{a}}\times f_{\mathrm{g}}\times f_{\mathrm{h}}$$ 2

Rearrangement gives

$$f_{\mathrm{a}}\times f_{\mathrm{g}}=F/f_{\mathrm{h}}$$ 3

Since F is dependent on the extent of hepatic clearance, it does not directly reflect absorption, and similar F values may arise from different absorption and clearance contributions. Consequently, interpretation based solely on F can obscure underlying differences in absorption-related behavior. , Accordingly, here we focus on the calculated term f _a × f _g as a more direct measure of oral absorption, noting that it represents an approximation of intestinal absorption and assumes minimal contribution from gut metabolism.

In this work, we aimed to develop a fully computational scoring framework that can be applied prior to synthesis to guide the design and prioritization of PROTACs, filling a current gap in the field. Our objective was to develop a practical and general computational metric for medicinal chemists to use for prospective compound prioritization rather than a comprehensive mechanistic model of oral absorption. By enabling prioritization prior to synthesis, this framework has the potential to reduce synthesis efforts, experimental burden, time, and cost toward the identification of orally bioavailable molecules. Our approach builds on recent efforts that use conformational descriptors derived from molecular dynamics simulations, including metrics such as R _g and 3D PSA, which have previously been applied in studies of chameleonicity. Here, however, we investigate these descriptors as empirical predictors of oral absorption rather than as direct measures of chameleonicity. For instance, R _g is intrinsically dependent on molecular size, such that larger PROTACs tend to exhibit higher values regardless of their conformational behavior, complicating interpretation across diverse chemical series. Likewise, strategies aimed at reducing 3D PSA to improve permeability may inadvertently favor overly lipophilic compounds, potentially compromising solubility and overall developability. Consequently, minimizing 3D PSA in isolation does not adequately capture the solubility/permeability balance required for optimal passive transport.

To address these challenges, we introduce normalized variants of these descriptors designed to decouple size effects while preserving meaningful comparisons across compounds. Specifically, the normalized radius of gyration is defined as nR _g = R _g /HAC, where HAC is the heavy atom count, and the normalized polar surface area is defined as nPSA = 3D PSA/PC, where PC represents the polar content, calculated as the fraction of polar atoms (N, O, P) relative to HAC. The performance of these normalized descriptors against fraction absorbed obtained from pharmacokinetic (PK) studies in rodents is illustrated in Figure . The data are presented as bar plots, where the y-axis represents the fraction of compounds exceeding this threshold, thereby reflecting the probability of identifying orally absorbed PROTACs across different descriptor ranges. These descriptors were computed from conformational ensembles generated by molecular dynamics (MD) simulations of PROTACs in chloroform as a hydrophobic solvent (See Supporting Information, Section S1, for details on R _g and 3D PSA computation and MD simulations). The data set consists of PK studies on 1,806 unique PROTACs, all evaluated in rat, where 1,237 were evaluated in mouse as well. Among these, 555 compounds in rat and 752 compounds in mouse exhibited f _a × f _g ≥ 0.25. This data set has been described previously in the context of defining physicochemical property boundaries associated with oral absorption in PROTACs. The present work builds upon that foundation by investigating whether molecular conformation provides complementary predictive information beyond the physicochemical descriptors. The calculated parameter of f _a × f _g was curated similar to ref . To illustrate the physicochemical space represented in this data set, distributions of the principal physicochemical descriptors are provided in Supporting Information, Section S2.

1.

Relationship between normalized descriptors and oral absorption for PROTACs. (a) Fraction of compounds with f _a × f _g ≥ 0.25 as a function of normalized radius of gyration (nR _g = R _g /HAC). (b) Fraction as a function of normalized polar surface area (nPSA = 3D PSA/PC). Data are shown for mouse (orange, n = 1,237; 752 above threshold) and rat (blue, n = 1,806; 555 above threshold). Increasing nR _g is associated with improved absorption probability, while decreasing nPSA shows a generally favorable trend.

The analysis indicates that increasing nR _g is associated with an increased probability of favorable oral absorption within the data sets examined, while decreasing nPSA shows a generally favorable trend. In comparison to the corresponding plots of the raw data for R _g and 3D PSA (Section S3, Supporting Information), the normalized descriptors exhibit substantially improved behavior, with clearer trends. These observations suggest that balancing molecular shape with controlled reduction in exposed polarity may increase the likelihood of identifying orally absorbed PROTACs.

To provide independent support for this interpretation, we performed principal moments of inertia (PMI) analysis. As shown in Figure , compounds with higher nR _g preferentially occupy the rod-like region of PMI shape space, confirming that nR _g captures molecular elongation. A complementary PMI analysis (Section S4, Figure S4, Supporting Information) further shows that PROTACs with favorable oral absorption are more frequently found in these rod-like regions.

2.

Principal moments of inertia (PMI) analysis of PROTAC molecular shape. Each point represents the ensemble-averaged PMI coordinates for a single PROTAC and is colored according to the corresponding ensemble-averaged normalized radius of gyration (nR _g = R _g /HAC).

A natural extension of these observations is to combine the normalized descriptors into a single metric that captures the balance between molecular extension and exposed polarity. To this end, we define a computational PROTAC Absorption Structural Score (comPASS) as

$$\mathrm{comPASS}=\alpha \times \frac{\mathrm{n}{R}_g}{\mathrm{nPSA}}$$ 4

which can be expanded as

$$\mathrm{comPASS}=\frac{\alpha \times R_g}{\mathrm{HAC}\times \left(\frac{3\mathrm{DPSA}}{\mathrm{PC}}\right)}$$ 5

where α is a scaling constant set to 10⁵ to yield numerically convenient values and R _g and 3D PSA are in Å and Ų, respectively. By design, higher comPASS values correspond to compounds with greater normalized molecular extension and reduced normalized polarity and are therefore expected to be associated with an increased likelihood of oral absorption.

This formulation has conceptual similarity to the recently reported Bifunctional Bioavailability Index (BBI). However, several key distinctions exist. First, comPASS is fully computational and does not rely on experimentally derived measurements, enabling application at early design stages prior to synthesis. Second, while BBI employs SMID as a descriptor for molecular shape derived from conformer generation, comPASS utilizes R _g obtained from molecular dynamics simulations, thereby incorporating conformational sampling of the ensemble. Notably, SMID corresponds to an order statistic (the minimum over sampled conformations) and therefore depends strongly on the extent of sampling. As sampling increases, rare compact conformations become accessible for many PROTACs, causing SMID values to converge toward similar lower bounds across molecules. In contrast, R _g reflects an ensemble-averaged property and is therefore less sensitive to rare events. We note, however, that both descriptors depend on the quality of conformational sampling, and our approach relies on standard MD simulations without enhanced sampling, which may limit exploration of the full conformational space. Nevertheless, R _g is expected to approach a stable value with increased sampling, whereas SMID systematically decreases as more compact conformations are discovered. This fundamental difference suggests that R _g provides a more statistically robust descriptor for comparing PROTAC geometries (see Section S5, Supporting Information, for an illustration). Third, comPASS replaces experimentally derived EPSA with a normalized, computationally derived 3D PSA descriptor (nPSA) and not a raw 3D PSA value. Finally, BBI includes a lipophilicity term (ChromLogD), whereas no such term is included here; in our analysis, computed lipophilicity descriptors (e.g., cLogD) did not provide predictive value and were therefore excluded to preserve both simplicity and accuracy of the model (see Section S6; Supporting info). The performance of comPASS was evaluated on the same data sets described above for mouse and rat using receiver operating characteristic (ROC) analysis (Figure ). The resulting ROC curves demonstrate consistent predictive capability across both species, with area under the ROC curve (AUC) values of 0.72 ± 0.02 for mouse and 0.73 ± 0.02 for rat. In addition, precision–recall analysis yields area under the curve (AUC-PR) values of 0.78 ± 0.03 for mouse and 0.57 ± 0.04 for rat. Differences in AUC-PR between mouse and rat data sets reflect variations in class distribution, as AUC-PR is sensitive to the baseline fraction of positives. Overall, these results suggest that comPASS provides a practical and transferable predictor of oral absorption likelihood using only computational descriptors and performs consistently across the rodent data sets examined. comPASS also shows significantly better performance compared to our previous rule-based approach for identifying orally favorable PROTACs. To make this comparison, the comPASS threshold was chosen to match the sensitivity achieved by the rule-based definition using the cutoffs reported in ref . for desired physicochemical properties. At matched sensitivity, comPASS showed a substantially lower false positive rate and higher precision and accuracy for both species (Section S7, Table S1, Supporting Information). This improvement highlights the ability of comPASS to more effectively reduce poorly absorbed compounds compared to the physicochemical property space approach. We also emphasize that analysis of individual descriptor contributions indicates that R _g is the dominant predictor of absorption, while polarity and size descriptors alone show limited or near-random performance (Section S8; Supporting Information).

3.

Receiver operating characteristic (ROC) curves for comPASS predicting oral absorption (f _a × f _g ≥ 0.25) in mouse (orange) and rat (blue) data sets described in Figure . The dashed line represents random classification. Performance is quantified by area under the ROC curve (AUC) as well as AUC-PR, the area under the precision–recall curve, as indicated in the legend.

To benchmark comPASS against existing approaches, we compared its performance with BBI as well as the experimentally derived descriptor ETR, defined as the EPSA to TPSA ratio reported by researchers at AbbVie. ETR was originally proposed as a descriptor for identifying polarity-reducing compounds and has been shown to be informative for specific PROTAC subclasses when combined with clustering analyses. Here, we evaluated its performance as a stand-alone predictor of oral absorption to compare its predictive power for compound absorption with that of comPASS and other scoring functions.

This analysis was performed on the subset of compounds for which experimental EPSA and ChromLogD measurements were available, comprising 186 rat studies, of which 107 also included mouse data, where 62 and 60 compounds, exhibit f _a × f _g ≥ 0.25, respectively (see bottom panel of Figure S2 for a distribution plot of the physicochemical properties of this set). In addition, we introduce an experimentally enhanced metric, ePASS, which replaces 3D PSA with experimentally measured EPSA, consistent with BBI, while retaining normalization through the polar content (PC) term. This formulation also substitutes SMID with R _g and further incorporates a lipophilicity component (ChromLogD), again aligned with BBI:

$$\mathrm{ePASS}=\beta \frac{R_g\times \mathrm{Chrom}\mathrm{Log}\mathrm{D}}{\mathrm{HAC}\times \left(\frac{\mathrm{EPSA}}{\mathrm{PC}}\right)}$$ 6

where β is a convenience scaling factor equal to 100. As shown in Figure , ETR exhibited limited predictive performance as a stand-alone predictor of oral absorption in the present data set. We note that ETR was originally proposed to identify polarity-reducing compounds within specific PROTAC subclasses rather than as a universal predictor of oral absorption. In contrast, ePASS showed the highest predictive performance among the methods evaluated. Although this comparison was performed on a relatively small subset of compounds for which experimental EPSA and ChromLogD measurements were available, it suggests that incorporation of experimentally determined descriptors has the potential to further improve prediction. The predictive performance of the fully computational comPASS score is established using the large computational data set and remains comparable to the experimentally enhanced scoring functions on the smaller experimental subset, despite requiring no experimental measurements. Together, these results support the use of comPASS as a practical computational tool for prospective compound prioritization, while illustrating the potential value of incorporating experimentally derived descriptors, EPSA and ChromLogD. Broader experimental data sets will be valuable for assessing the generality of this observation.

4.

Comparison of comPASS with alternative scoring approaches for predicting oral absorption (f _a × f _g ≥ 0.25) in mouse (top) and rat (bottom) data sets. ROC curves are shown for ePASS, comPASS, BBI, and ETR. The dashed line represents random classification. Performance metrics (AUC and AUC-PR) are reported in the legend. The data set comprises 107 mouse compounds (62 above threshold) and 186 rat compounds (60 above threshold). ePASS shows the strongest performance, while comPASS achieves comparable results in rat using only computational descriptors. BBI performs similarly in rat but underperforms in mouse, and ETR shows consistently poor predictive capability. Error bars represent asymmetric 90% bootstrap percentile confidence intervals, computed from 1,000 resamples with replacement of the outcome–score pairs for each method.

To further interrogate the interpretability of the scoring functions, we examined the distribution of scores across compounds (Figure ) with low and high absorption (f _a × f _g < 0.25 vs ≥ 0.25). This plot directly assesses the extent to which each score separates the two populations. As shown in Figure , both comPASS and ePASS exhibit shifts toward higher values for absorbed compounds in both species, indicating potential discrimination at the distribution level. The statistical significance of the separation between absorbed and poorly absorbed compounds was established using the Mann–Whitney U test (see Supporting Information Section S9). In particular, ePASS shows the strongest separation, consistent with its superior ROC performance. In contrast, BBI demonstrates more limited separation, with substantial overlap between the two classes, while ETR shows minimal differentiation. These results reinforce that the proposed normalization framework captures relevant structural features associated with absorption and provides an interpretable basis for prioritizing compounds, complementing the global performance metrics shown in Figure .

5.

Distribution of scoring functions for compounds with f _a × f _g < 0.25 and ≥ 0.25 in rat (blue) and mouse (orange). Box plots are shown for ePASS (a), comPASS (b), BBI (c), and ETR (d). Mean values are indicated below each group. ePASS and comPASS show clear separation between classes, whereas BBI shows moderate separation and ETR exhibits minimal discrimination.

A key component of our framework is the inclusion of polar content (PC) as a normalization factor. To assess its contribution, we evaluated variants of both comPASS and ePASS with the PC term removed (Section S10, Table S2, Supporting Information). In all cases, exclusion of PC leads to a consistent decrease in predictive performance, as reflected by lower AUC values in both mouse and rat data sets. This trend underscores the importance of incorporating PC, as its inclusion improves the discriminative power of both computational and experimentally enhanced scoring functions. From a design perspective, this analysis highlights that reducing polarity alone is insufficient; rather, polarity must be optimized relative to molecular size and composition to achieve favorable absorption.

To evaluate whether alternative functional forms could improve performance, we explored generalized logarithmic combinations of the underlying descriptors and optimized their coefficients to maximize AUC. These analyses yielded only marginal improvements and converged to formulations closely resembling comPASS. Given the minimal gains and increased complexity, the simple form of comPASS (eqs and ) is preferred. Details are provided in the Supporting Information (See Section S11, Supporting Information).

In summary, our results indicate that extended molecular geometries are associated with higher likelihood of improved oral absorption and can be utilized as a practical design principle for prospective PROTAC prioritization. The mechanistic basis for this association remains to be established. Several factors, including molecular shape, polarity, passive permeability, active transport, and efflux, are all expected to contribute to oral absorption, and the present study was not designed to distinguish their individual contributions. Although the present study does not evaluate environment-dependent conformational changes associated with chameleonicity, our results are inconsistent with one of the central features of the chameleonicity hypothesis that orally absorbed PROTACs preferentially adopt compact conformations in a hydrophobic environment. We introduced comPASS, a fully computational and design-ready scoring function that captures the balance between molecular shape and polarity and provides a practical framework for presynthetic PROTAC prioritization. We demonstrated that this framework may be potentially enhanced through incorporation of chromatographic measures of polarity (EPSA) and lipophilicity (ChromLogD), although broader experimental data sets will be valuable to establish the generality of this observation. These findings provide a practical and scalable framework for medicinal chemists to prioritize and design PROTACs with a higher likelihood of oral absorption prior to synthesis.

Glossary

Abbreviations

PROTAC

proteolysis-targeting chimera

comPASS

computational PROTAC absorption structural score

ePASS

experimentally enhanced PROTAC absorption structural score

MD

molecular dynamics

PK

pharmacokinetics

R _g

radius of gyration

PSA

polar surface area

3D PSA

three-dimensional polar surface area

TPSA

topological polar surface area

EPSA

experimental polar surface area

nR _g

normalized radius of gyration

nPSA

normalized polar surface area

HAC

heavy atom count

PC

polar content

PMI

principal moments of inertia

SMID

smallest maximum intramolecular distance

BBI

Bifunctional Bioavailability Index

ETR

EPSA-to-TPSA ratio

HBD

hydrogen bond donor

HBA

hydrogen bond acceptor

MW

molecular weight

RB

rotatable bond

ROC

receiver operating characteristic

AUC

area under the receiver operating characteristic curve

AUC-PR

area under the precision–recall curve

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsmedchemlett.6c00305.

• Data set used for comparison of comPASS and ePASS with BBI and ETR, including experimental oral absorption and molecular descriptor data (XLSX)

• Additional analyses of molecular dynamics convergence, physicochemical property distributions, molecular shape, alternative descriptors and scoring functions, and statistical evaluation of comPASS and ePASS (PDF)

The manuscript was prepared with contributions from all authors.

Safety: No unexpected or unusually high safety hazards were encountered.

The authors declare the following competing financial interest(s): All authors are current employees of Arvinas Inc. and may own shares of stock and/or stock options in this company.