Physical feasibility gating in early drug discovery: a decision framework for identifying predictable physical failures, with retrospective illustration

1 Introduction

1.1 Scope and manuscript type

This work presents a conceptual perspective proposing a feasibility-gating framework for integrating physical constraints into early discovery decisions. It is not a comprehensive review of physical chemistry in pharmacology nor a methods paper describing new computational techniques. The scope is intentionally limited to outline how thermodynamic, kinetic, transport, and microenvironmental constraints establish the boundaries within which any successful drug candidate must operate. Furthermore, it aims to organize these constraints into a transparent decision logic that supports resource allocation prior to extensive synthesis. The framework is designed to function during the discovery stage as a lightweight, low-data triage layer positioned upstream of the more data-intensive QSP and PBPK modeling described in Section 1.5, and to hand candidates to these models once they justify the necessary investment.

What this manuscript does not attempt: prediction of structure-activity relationships from first principles; substitution for experimental validation of biological hypotheses; resolution of target-selection uncertainty; or comprehensive treatment of specialized topics such as nonlinear pharmacokinetics, active transport, or target-mediated drug disposition. These limitations are stated explicitly because they define the framework’s appropriate domain.

1.2 Definition of physical constraints

Physical constraints, as used here, are the thermodynamic, kinetic, and transport limitations that govern whether a drug can achieve adequate concentration, engagement, and persistence at its site of action. They include binding thermodynamics, determining equilibrium occupancy; association and dissociation kinetics, determining time-dependent engagement; permeability, diffusion, and partitioning, determining spatial distribution; and microenvironmental factors such as pH, ionic strength, and macromolecular crowding that modulate drug properties under disease-relevant conditions.

Physical constraints are distinguished from biological constraints such as pathway relevance, target essentiality, and disease-mechanism validity, which remain experimentally irreducible. A drug may fulfill all physical constraints by achieving sufficient free concentration at the target with appropriate kinetics; however, it may still fail in clinical outcomes if the biological hypothesis connecting target modulation to disease modification is proven incorrect. This distinction is fundamental to the framework.

1.3 The attrition problem and its causes

As summarized by Sun et al. (2022), analyses of clinical development data spanning roughly 2010 to 2017 indicate that approximately 90% of candidates entering Phase I fail to obtain approval. Phase-transition analyses show that attrition is most severe at Phase II, the stage with the lowest success rate of any development phase (Arrowsmith and Miller, 2013; Harrison, 2016). The causes have been categorized as inadequate efficacy (40–50 percent), unacceptable toxicity (about 30 percent), poor drug-like properties (10–15 percent), and commercial or strategic factors (about 10 percent) (Sun et al., 2022). The dominance of efficacy failures carries a direct implication: the primary cause of late-stage attrition is biological, not physical. When drugs engage their targets with adequate exposure and appropriate kinetics yet do not modify the disease, the error lies in the biological hypothesis (Table 1).

TABLE 1. Minimum evidence package for physical feasibility assessment at each gate.

This biological uncertainty cannot be resolved through physical analysis; it requires experimental validation in patients. Any framework proposing to address attrition must acknowledge this limitation. The reframing of the amyloid hypothesis in Alzheimer’s disease, repeated failures of anti-inflammatory approaches in sepsis, and the limited efficacy of many kinase inhibitors in solid tumors all exemplify situations in which drugs engaged their targets but failed to produce clinical benefit because the biology proved incomplete or incorrect. Importantly, as Section 2 develops, most such cases are multi-causal rather than clean single-cause illustrations.

1.4 Contribution statement

The novelty of this work does not lie in identifying physical constraints, which are foundational to pharmacology (Rang, 2006; Kenakin, 2018), but in their organization into a feasibility-gating decision aid. Experienced discovery teams already reason about free concentration, tissue penetration, residence time, and assay relevance, but typically do so informally, target by target, and often after a series has been committed to. The contribution here is therefore stated as demonstrable rather than merely asserted. The framework makes four specific contributions:

• A structured, checklist-style gating logic that sequences physical assessments prior to comprehensive synthesis, incorporating explicit pass and flag criteria for each gate grounded in measurable parameters. Its objective is to render the implicit reasoning of proficient teams explicit, auditable, and operational at stages where limited data are available and decisions are most economical to modify.
• A minimal evidence package specifying the required measurements, the common artifacts to exclude, and the interpretation limits for each constraint category, so that gating decisions are made with appropriate confidence and known uncertainty.
• A failure-mode taxonomy, presented as a hierarchical tree, which categorizes development failures into classes that are addressable by physics, partially addressable by physics, and addressable by biology, thereby clearly delineating the specific failures that the framework is capable of addressing and those it is not.
• A modality-specific constraint mapping that identifies predominant physical limitations across small molecules, covalent inhibitors, protein degraders, photopharmacological agents, biologics, and nucleic acid therapeutics, extending the logic to contemporary drug discovery.

Section 1.5 situates the framework in relation to QSP and PBPK; Section 6 demonstrates the gating logic retrospectively against documented development histories.

1.5 Relationship to quantitative systems pharmacology and PBPK modeling

Quantitative Systems Pharmacology and Physiologically Based Pharmacokinetic modeling have, for approximately 2 decades, conducted integrated physical and biological feasibility analyses at a level of mechanistic detail surpassing what is proposed herein (Bradshaw et al., 2019; Kuepfer et al., 2016). PBPK models forecast tissue and plasma exposure based on physiological and physicochemical inputs; QSP models incorporate exposure, pharmacodynamics, and downstream biology to predict responses within virtual populations (Bradshaw et al., 2019; Helmlinger et al., 2019). These platforms are capable of, and routinely perform, mechanistic testing of whether free drug levels can support target engagement and whether tissue penetration is adequate. The framework introduced herein is neither intended to compete with nor replace these established platforms, and its relationship to them is made explicit below.

The relationship involves stages, data intensity, and purpose. QSP and PBPK models are generally developed once sufficient data are collected and adequate commitment is secured to justify their construction, parameterization, and qualification; they are most effective from late discovery through phases of clinical development. The gating framework is intentionally positioned upstream and is lightweight: it utilizes a limited set of early measurements (such as Kd or Ki, free fraction, an estimated or predicted Kp,uu, an occupancy target, and a tolerated-exposure estimate) to pose a screening question, specifically whether a candidate series warrants further development to more costly mechanistic modeling and biological characterization. In this context, the framework acts as a complementary, subordinate layer: a triage step generating hypotheses for testing and, in the event of a pass, parameters to be integrated into PBPK and QSP models. When the two overlap, for example, in occupancy and exposure analysis, the gating calculation is best understood as a simplified, transparent special case of what a comprehensive PBPK or QSP model computes with greater precision. The value of this framework is thus organizational and procedural, rather than mechanistic: it advances questions to the point of minimal committed cost and documents the responses in an auditable manner.

2 The primacy of biological uncertainty: what physics cannot address

Prior to evaluating the contributions of physical analysis, it is imperative to consider its limitations. This section delineates the appropriate scope of physical feasibility analysis by recognizing the predominant influence of failures in biological hypotheses on clinical attrition, and by acknowledging that actual failures are typically multifactorial.

2.1 Biological hypothesis failures dominate late-stage attrition

Detailed analysis of Phase II and Phase III terminations shows that efficacy failures typically occur in compounds that demonstrate target engagement, achieve adequate plasma exposure, and exhibit acceptable early safety (Harrison, 2016). These drugs fail because inhibiting or activating their intended target does not produce the anticipated therapeutic effect, a failure of the biological hypothesis rather than of physical feasibility.

The development of drugs for Alzheimer’s disease exemplifies this point, with an important caveat regarding causation elaborated below. Historical efforts targeting amyloid-beta production or aggregation have demonstrated target engagement and substantial reduction of amyloid, yet they have repeatedly failed to significantly decelerate cognitive decline, thereby fueling skepticism concerning the amyloid hypothesis as originally proposed (Espay et al., 2024). These pharmacological agents reached the brain, bound their intended targets, and diminished amyloid burden; the core challenge persisted with the hypothesis itself. More recently, lecanemab and donanemab received approval from the FDA (January 2023 and July 2024, respectively) based on clinical trials that indicated statistically significant yet modest clinical slowing of decline (Van Dyck et al., 2023; Sims et al., 2023), with differences on the 18-point CDR-SB scale approaching 0.4 to 0.5 points at 18 months and ongoing debate concerning their clinical significance (Daly et al., 2024). These treatments are also associated with amyloid-related imaging abnormalities; for lecanemab, symptomatic and asymptomatic ARIA with edema occurred in roughly 13 percent of treated patients overall, rising to about 35 percent in APOE4 homozygotes, with comparable or higher rates reported for donanemab (Honig et al., 2024).

Crucially, the Alzheimer’s history is not a straightforward single-cause illustration. The observed outcomes encapsulate a complex interplay of hypothesis uncertainty, biomarker interpretation, patient selection, disease stage, trial design, and compound-specific safety liabilities. We present this not as definitive proof that biology alone accounts for the failures but as an example where target engagement was unequivocal, yet the clinical benefit remained uncertain, a disconnect that physical analysis cannot resolve.

2.2 Similar patterns across therapeutic areas, and the reality of mixed causation

The pattern of target engagement without reliable therapeutic translation recurs across therapeutic areas, and in each, the causation is typically mixed. In oncology, highly selective kinase inhibitors can fail when tumors evolve resistance or when the targeted pathway proves dispensable; the BRAF-inhibitor experience in melanoma demonstrates that even genetically validated target selection can be undermined by adaptive resistance, alongside pharmacokinetic, scheduling, and patient-selection factors. In cardiovascular disease, the CETP inhibitors torcetrapib and evacetrapib significantly increased HDL cholesterol levels yet failed to reduce, or even worsened, hard clinical outcomes (Barter et al., 2007); torcetrapib, in particular, was associated with an off-target, molecule-specific increase in aldosterone and blood pressure, indicating that its failure is attributable both to an incorrect hypothesis regarding HDL causality and to compound-specific liabilities rather than a singular, clear cause. In metabolic disease, agents that normalized glycemic control did not consistently prevent vascular complications, and some raised concerns regarding cardiovascular health, again reflecting a complex interplay of mechanism, endpoint selection, and compound effects.

We therefore explicitly clarify that these are illustrative, multi-causal histories, and the framework pertains solely to the subset of questions suitable for early physical analysis. It does not claim to explain, nor to have predicted, the biological, clinical, and program-level factors that predominantly influence most actual failures.

2.3 What physical analysis cannot resolve

Physical constraints delineate the essential conditions for drug action; however, they cannot establish the sufficiency of these conditions for therapeutic success. Numerous categories of questions fall beyond their scope and necessitate experimental investigation.

• whether the selected target is causally relevant to disease progression in patients rather than only in preclinical models;
• whether target modulation produces benefit or whether compensatory mechanisms and pathway redundancy attenuate the effect;
• whether preclinical models recapitulate the relevant human pathophysiology or whether species differences limit translation;
• whether efficacy seen in enriched early trials replicates in broader pivotal studies;
• and what structure-activity modifications will improve biological response in ways not predictable from binding measurements alone.

Approaches that partially de-risk biology, including human genetic validation, target-engagement biomarkers, pharmacodynamic readouts, and causal inference from observational data (Plenge et al., 2013), provide valuable information but do not eliminate the need for clinical validation. The framework proposed here addresses the complementary problem of identifying candidates that will predictably fail on physical grounds regardless of biological validity.

3 Physical constraints on drug action: quantitative foundations

All pharmaceuticals are required to comply with specific physical constraints to ensure the exertion of their intended biological effects. These constraints are governed by principles of thermodynamics, kinetics, and transport physics. When a candidate compound breaches these constraints, failure is likely, independent of the biological hypothesis. This section outlines the quantitative foundations and, where relevant, discusses the propagation of uncertainty in each input to the final decision.

3.1 Thermodynamic constraints: occupancy and energetics

Drug-target interactions adhere to the principles of the law of mass action, as formalized by Clark (1937) and subsequently refined by Stephenson (1956) and Black and Leff (1983). Fractional occupancy is:where [D] denotes the free drug concentration at the site of action, and K_d represents the equilibrium dissociation constant (Jarmoskaite et al., 2020). Attaining a specified occupancy necessitates either sufficient affinity, characterized by a low K_d, or adequate free concentration; in situations where neither condition is satisfied, target modulation will be insufficient irrespective of biological significance. A compound with a K_d of 100 nM requires approximately 100 nM free concentration to achieve 50 percent occupancy. If distribution constraints limit the attainable free concentration to 10 nM, the maximum occupancy will be roughly 9 percent.

Fractional occupancy = D / D + K d (1)

The free drug theory (Summerfield et al., 2022) posits that only unbound drug molecules, which are in dynamic equilibrium at the site of action, engage with the biological target. Affinity is associated with the standard Gibbs free energy of binding, ΔG° = RT ln (K_d) = ΔH° − TΔS° (Equation 2), where Kd is expressed relative to the standard-state reference concentration of 1 M so that its logarithm is dimensionless, and where enthalpic gains are often counterbalanced by entropic losses, a phenomenon known as enthalpy-entropy compensation (Chodera and Mobley, 2013; Geschwindner et al., 2015). This concept elucidates the diminishing returns frequently observed in the process of affinity optimization.

3.1.1 Uncertainty propagation

Since occupancy is a saturable function of [D]/K_d, the decision to pass or fail at this threshold is most sensitive near the target occupancy. A two-fold uncertainty in K_d or in the free fraction roughly translates to a two-fold uncertainty in the required free concentration, which may cause a borderline candidate to cross the decision boundary. Therefore, the evaluation of the gate should involve explicit uncertainty ranges for K_d, free fraction, occupancy requirement, and tolerated exposure. Additionally, the results should be conveyed as a probability of feasibility rather than a definitive yes or no, particularly when the inputs are uncertain (see Section 4).

3.2 Kinetic constraints: residence time and temporal engagement

Biological systems seldom exist at equilibrium. Under conditions of nonequilibrium, association (k_on) and dissociation (k_off) rates can influence pharmacodynamic outcomes independently of equilibrium affinity (Copeland, 2016; Tonge, 2018). The residence time of a drug-target interaction is expressed as τ_R = 1/k_off (Equation 3).

The residence-time model (Copeland, Pompliano, and Meek, 2006; Copeland, 2021; Knockenhauer and Copeland, 2024) holds that longer residence times can produce more durable engagement than free concentration alone would support, with supporting evidence for muscarinic antagonists such as tiotropium (Sykes et al., 2012), kinase inhibitors (Tonge, 2018), and vasopressin receptor antagonists (Zhang et al., 2022). Important caveats apply: measured residence times depend strongly on assay conditions, and in vivo duration also reflects tissue retention, local rebinding, pharmacokinetic half-life, and target turnover. Residence time is one variable among several and must be established empirically for each system. The kinetic gate is therefore framed as conditional, applying only when the biology requires sustained engagement.

3.3 Transport constraints: reaching the site of action

The pharmacological effect necessitates the presence of the drug at the site of action. The correlation between physicochemical properties and oral absorption was assessed by Lipinski et al. (2001), whose property parameters delineate the physics underlying passive diffusion and first-pass metabolism. Most approved oral medications fall within these property ranges, although well-recognized exceptions exist and the rule-of-five framework has been reassessed as chemical space has expanded (Mullard, 2018). Targets within the central nervous system impose more stringent restrictions, typically favoring lower molecular weight, diminished polar surface area, limited hydrogen bonding, moderate lipophilicity, and minimal P-glycoprotein (P-gp) efflux.

The quantity governing site-of-action exposure is the unbound tissue-to-plasma partition coefficient, K_p,uu, defined at steady state as the ratio of unbound tissue to unbound plasma concentration. For the brain, K_p,uu can be estimated experimentally using the brain-slice and brain-homogenate methods, whose divergence is explained by a pH-partitioning model (Fridén et al., 2011), or predicted in silico from structure-brain exposure relationships in which hydrogen bonding is the dominant determinant (Fridén et al., 2009; Loryan et al., 2022). A K_p,uu well below unity indicates that adequate target-site exposure necessitates proportionally higher systemic exposure, which may conflict with safety limits and delineate the achievable therapeutic window.

3.4 Microenvironmental factors: condition-dependent properties

The site of action is not the homogeneous buffer of a standard assay. Disease compartments exhibit pH gradients, ionic-strength differences, altered lipid composition, macromolecular crowding, and redox shifts that can change protonation state, solubility, partitioning, and binding (Curry, 2015). Weak bases concentrate in acidic compartments by ion trapping; tumor microenvironments near pH 6.5 to 6.9 can accumulate weakly basic drugs severalfold above plasma (Mahoney et al., 2003), enhancing tumor exposure but also off-target accumulation in lysosomes, while weak acids are excluded. These effects follow the Henderson-Hasselbalch relationship and are predictable from pKa, but only if the relevant microenvironmental pH is characterized; assays run only at pH 7.4 may misrepresent behavior in disease compartments.

3.5 Worked example: a simplified, illustrative CNS feasibility analysis

3.5.1 Important framing

The following is a deliberately simplified, pedagogical illustration, not a predictive calculation for any real program. It assumes that the occupancy threshold is known, that K_p,uu is reasonably established, and that steady-state free concentration is the relevant driver of response. In actual programs these three assumptions are usually the central uncertainties, not givens, and each would carry a distribution rather than a point value. The example shows the structure of the reasoning, not a claim about predictive accuracy.

Consider a hypothetical small-molecule program against an intracellular CNS enzyme:

• Target Kd = 10 nM (biochemical assay, pH 7.4, 25 °C)
• Required occupancy for efficacy = 70 percent (from pharmacological modeling)
• Brain Kp,uu = 0.1 (in vitro prediction from brain-slice binding and plasma protein binding)
• Maximum tolerated total plasma Cmax = 2000 nM (from toxicology)
• Plasma protein binding = 95 percent (free fraction 0.05)
• Step 1. From Equation 1, 70 percent occupancy at K_d = 10 nM requires free brain concentration = (0.70 × 10)/(1–0.70) = 23.3 nM.
• Step 2. With K_p,uu = 0.1, required free plasma concentration = 23.3/0.1 = 233 nM.
• Step 3. Applying the free fraction of 0.05 to the tolerated total plasma Cmax of 2000 nM gives a tolerated free plasma concentration of 100 nM. The required 233 nM exceeds this tolerated free concentration by about 2.3-fold, flagging a physical feasibility risk under these assumptions.
• Step 4, options: (a) improve affinity to Kd ≤ about 4.3 nM; (b) improve brain penetration to Kp,uu ≥ about 0.23; (c) accept reduced occupancy (about 50 percent at the exposure limit, since a free brain concentration of 10 nM against a Kd of 10 nM yields half-maximal occupancy), whose adequacy depends on mechanism; or (d) pursue alternative delivery such as intrathecal administration to bypass the barrier.

Given that each input entails uncertainty, the appropriate outcome is not a definitive verdict but rather a flagged risk accompanied by quantified targets for remediation. With a two-fold uncertainty concerning affinity and Kp,uu, the 2.3-fold difference resides within a range where the decision may be reversible, which is precisely why the gate output should be regarded as a risk to be mitigated through measurement, rather than a declaration of impossibility. This example demonstrates four fundamental principles: feasibility is assessed quantitatively based on early parameters; potency, exposure, and distribution are interconnected; specific remediation targets are identified; and the analysis does not address the validity of the biological hypothesis.

4 The feasibility-gating framework: operational implementation

This section organizes the constraints into a decision aid with explicit criteria, minimum evidence, and decision rules, and clarifies for each gate whether its threshold is a hard boundary, a practical heuristic, or a context-dependent template.

4.1 Nature of the thresholds

A central clarification concerns the status of the numerical criteria. The framework deliberately mixes three kinds of threshold, and labels them as such (Figure 1):

• Hard boundaries are set by physical law and are not adjustable; for example, fractional occupancy cannot exceed the value dictated by Equation 1 for a given free concentration and Kd.
• Practical heuristics are conventional decision triggers chosen to balance false-positive and false-negative triage errors; the 2 to 3-fold gap at the thermodynamic gate is a heuristic, not a law, and an organization may tighten or loosen it according to its risk tolerance and the cost of the next experiment.
• Context-dependent templates require a target-class-specific value to be supplied before the gate can be evaluated; the occupancy requirement, the relevant biological timescale for the kinetic gate, and the disease-relevant pH range are templates of this kind.

Qualitative expressions such as the residence time being considerably shorter than the biological window, or properties degrading significantly under pathological conditions, are heuristic in nature. Where applicable, they can be assigned order-of-magnitude thresholds, for example, flagging when the residence time is less than ten times the dosing interval or when a key parameter varies by more than twofold across the disease-relevant condition spectrum. These thresholds serve as initial reference points for calibration rather than definitive constants. In all instances, the gate output should be regarded as an indicator of risk accompanied by an associated degree of uncertainty, rather than a binary decision. Whenever inputs are uncertain, the resulting determination should be expressed as a probability of feasibility, obtained by propagating the input distributions.

4.2 Gating logic and decision rules

• Gate 1: Thermodynamic feasibility involves applying a heuristic boundary to a strict relationship. It is necessary to determine whether the required occupancy can be attained at pharmacokinetically feasible free drug concentrations. Calculate the necessary free concentration using Equation 1 and the measured Kd, then compare this with the achievable site-of-action concentrations. If the required concentration exceeds the achievable value by more than the heuristic factor (default set to 2–3 times), this indicates a potential risk, warranting subsequent steps such as affinity optimization with explicit targets or exploring alternative delivery methods.
• Gate 2: Kinetic Feasibility (Conditional, Heuristic). When the mechanism necessitates sustained engagement, compare residence time (1/k_off) with the pertinent biological timescale. If the residence time is significantly shorter than required, the equilibrium potency will overestimate the in vivo effect; accordingly, flag this issue and either optimize the residence time or design dosing regimens for continuous engagement. If the biological system permits intermittent or partial occupancy, this criterion may not be applicable.
• Gate 3: Transport feasibility, representing a heuristic boundary on a rigid relationship. Estimate K_p,uu for the designated compartment and calculate attainable site-of-action concentrations at feasible systemic exposures. Should transport inhibit the attainment of adequate target-site concentrations, solely enhancing potency will not suffice to salvage the project. In such cases, consider modifying physicochemical properties, pursuing localized or targeted delivery strategies, or altering the modality.
• Gate 4, microenvironmental robustness (template). Do key parameters remain favorable under disease-relevant pH, ionic strength, redox, and crowding? Test under conditions approximating the disease microenvironment rather than standard buffers. If properties degrade beyond the chosen template tolerance, flag elevated translation risk and optimize for robustness, design a prodrug strategy, or document and accept the risk.

4.3 Gate ordering, parameter interactions, and tradeoffs

The four-gate sequence is a default, not a fixed pipeline, and the ordering should be set by the target class. For CNS and intracellular targets, transport is frequently the first and most decisive constraint and should be screened first; for a high-affinity extracellular target with straightforward access, the thermodynamic gate may dominate; for mechanisms requiring durable suppression, the kinetic gate may take precedence. The framework is therefore best read as a set of parallel constraints that any program must satisfy, prioritized in the order most likely to be decisive for that target class, rather than as an obligatory four-step March. Figure 2 shows the default order, explicitly noting this flexibility.

More importantly, the gates are interconnected, and enhancing one often compromises another. A prime example is found centrally within medicinal chemistry: efforts to reduce polar surface area and hydrogen-bond count to enhance brain penetration (facilitating Gate 3) typically weaken the polar interactions that contribute to binding affinity (hindering Gate 1), thereby trading increased transport capacity for diminished potency. Similarly, structural modifications aimed at increasing the dissociation barrier to prolong residence time (Gate 2) may lead to an increase in molecular weight or lipophilicity, adversely affecting permeability or solubility (Gates 3 and 4). Additionally, the introduction of ionizable groups to improve solubility can influence microenvironmental partitioning through ion trapping (Gate 4), while concurrently altering protein binding and the free fraction (Gate 1). Although this framework does not resolve these tradeoffs, which are an inherent aspect of lead optimization, it renders them explicit by requiring the evaluation of all gates collectively rather than in isolation. Consequently, any advantage claimed at one gate must be weighed against its potential costs at the others. Treating these gates as an interconnected system, rather than as separate filters, is a key feature that distinguishes an explicit framework from informal approaches.

4.4 Failure-mode taxonomy

A principal contribution of this work lies in the explicit categorization of failure modes based on their addressability through early physical analysis. The categorization is structured as a hierarchical tree (see Figure 3), a form designated here as a taxonomy. The root of the tree signifies failure to attain approval; the three primary branches represent failures that are physics-addressable, partially physics-addressable, and biology-addressable, respectively. Each branch culminates in specific failure modes represented as leaves. This tree visually clarifies the scope of the framework: it applies solely to the left and middle branches.

5 Modality-specific constraint mapping

Each therapeutic modality presents characteristic physical constraints that can be evaluated independently of biological validity, as well as unresolved questions that are inherently experimental. This section covers small molecules, covalent inhibitors, targeted protein degraders, photopharmacological agents, biologics, and nucleic-acid therapeutics, with Table 3 summarizing the assessable physical constraints and the irreducibly experimental questions for each.

5.1 Covalent inhibitors

Covalent drugs establish chemical bonds with their targets, incorporating considerations of electronic structure beyond traditional force fields (Baillie, 2016). The primary criterion for feasibility is the equilibrium between warhead reactivity and selectivity: a reactive warhead must selectively modify the intended nucleophile rather than abundant cellular thiols such as glutathione. This balance is determined by both the intrinsic reactivity and the binding affinity that aligns the warhead. Quantum Mechanics/Molecular Mechanics (QM/MM) investigations of covalent kinase inhibitors, including ibrutinib, have elucidated how activation barriers influence both kinetics and selectivity (Voice et al., 2021). Excessive intrinsic reactivity may result in apparent biochemical potency while leading to cellular failure due to nonspecific reactivity, a risk that can be assessed through early reactivity profiling.

5.2 Targeted protein degraders

Degraders, including PROTACs and molecular glues, must engage both the target and an E3 ligase to establish a productive ternary complex (Bekes et al., 2022). Their physical constraints encompass binary binding, as well as the geometry, cooperativity, and dynamics of the ternary complex, in addition to permeability despite high molecular weight. Minor modifications in linker length, attachment site, or exit-vector geometry can significantly influence degradation efficiency, highlighting the sensitivity of ternary complex formation to three-dimensional spatial arrangement. Presently, due to limitations in predictive methodologies, extensive experimental iteration remains indispensable.

5.3 Photopharmacology

Photopharmacological agents incorporate light-switchable moieties to toggle activity through illumination (Liu et al., 2025), thereby enabling both spatial and temporal control that are not achievable through molecular recognition alone. Their distinctive physical constraints include tissue light-penetration depth (measured in millimeters for visible light and centimeters for near-infrared light), photoswitch quantum yield and kinetics, photofatigue resistance, and the thermal stability of the metastable state. These factors are quantifiable and can be integrated into feasibility analyses. The question of whether spatial control confers a therapeutic advantage over systemic exposure remains to be determined through experimental investigation.

5.4 Biologics

Biologics, primarily comprising monoclonal antibodies and engineered proteins, are subject to a distinct yet equally physical set of constraints. Their distribution is predominantly driven by convective and lymphatic transport mechanisms rather than passive diffusion; consequently, tissue penetration, particularly into solid tumors, is constrained by factors such as molecular size, interstitial pressure, and the binding-site barrier, wherein high affinity can paradoxically inhibit deeper tissue penetration. The disposition of biologics frequently exhibits nonlinear characteristics due to target-mediated drug disposition, wherein binding to a highly available target forms a saturable clearance pathway, rendering drug exposure a complex function of dosage (Datta-Mannan, 2019; Mager and Jusko, 2001). Constraints regarding developability, specifically conformational and colloidal stability, aggregation propensity, viscosity at elevated concentrations, and chemical degradation hotspots, are physical properties that can be measured early and are critical for manufacturability and subcutaneous administration. The risk of immunogenicity involves physically assessable factors, including aggregate load and sequence or structural liabilities, although actual clinical immunogenicity rates remain an experimental endpoint. For biologics, the transport barrier is redefined in terms of macromolecular transport and target-mediated disposition, while a microenvironmental-robustness barrier relates to conformational and colloidal stability under formulation and physiological stress conditions (Figure 4).

5.5 Nucleic-acid therapeutics

Nucleic-acid therapeutics, including antisense oligonucleotides, small interfering RNA, and messenger RNA, are governed predominantly by physical constraints that align naturally with the present framework (Hammond et al., 2021). The principal challenge lies in delivery: these molecules are large, highly negatively charged, and incapable of permeating membranes, rendering intracellular delivery, as well as, in the case of mRNA, endosomal escape, the rate-limiting physical steps. Delivery is facilitated chemically through conjugation methods such as triantennary N-acetylgalactosamine for hepatocyte targeting or via formulation within lipid nanoparticles, where factors such as composition, size, surface charge, and apparent pKa influence cellular uptake, endosomal escape, and biodistribution. Early assessments of stability against nucleases, including the effects of chemical modifications such as 2′-O-methyl, phosphorothioate linkages, and modified nucleosides on stability and innate immune recognition, as well as the physical attributes of the carrier, are all measurable. Within this framework, the transport gate becomes the predominant barrier for this modality, reframed around cellular uptake and endosomal escape rather than passive permeability. Thermodynamic and kinetic gates are applicable to hybridization or loading steps, while the microenvironmental robustness corresponds to nuclease stability and carrier integrity under physiological conditions. The ultimate question, as always, remains whether modulating the selected transcript confers clinical benefit, which is inherently a biological inquiry (Tables 2, 3).

TABLE 2. Failure-mode taxonomy in tabular form, corresponding to the branches of Figure 3.

TABLE 3. Physical constraints and irreducibly experimental questions across therapeutic modalities.

6 Retrospective illustration: applying the logic to documented histories

It is reasonable to ask whether the framework accurately reflects the behavior of both successful and unsuccessful programs rather than remaining speculative. A comprehensive validation in the statistical sense, applying the gating logic blindly to a large, representative candidate set and assessing predictive performance, transcends the scope of a conceptual analysis and would require a curated dataset containing early-stage parameters and known outcomes; this is identified below as the necessary next step. What can be offered here is a retrospective illustration: a limited set of documented histories examined to determine whether the gating logic, applied to the physical parameters known at an early stage, would have identified the physically predictable risks, and equally whether it would have remained silent in cases where failure was due to biological factors. The cases below were selected because their physical causes are unusually well documented, which both aids exposition and limits the inferential weight that can be drawn from them. The inclusion of both successes and failures is intended to show that the gating mechanism does not simply exclude all candidates; rather, a useful filter must permit the passage of candidates that ultimately proved successful.

These are illustrative representations, not blindly made predictions, and they are susceptible to hindsight bias; we only draw the limited inference that the logic aligns with documented physical outcomes and appropriately refrain from claiming credit for biological phenomena.

6.1 A transport-gate failure the framework would have flagged: early CNS programs limited by efflux

Numerous historical CNS programs have advanced potent compounds that failed to achieve adequate brain exposure due to P-glycoprotein efflux mechanisms and unfavorable physicochemical properties. This pattern has driven the now-standard focus on Kp,uu and CNS property guidelines (Hitchcock and Pennington, 2006; Loryan et al., 2022). A clear illustration of this principle is provided by loperamide, a high-affinity opioid agonist that lacks central opioid activity at therapeutic doses because P-glycoprotein at the blood-brain barrier maintains its unbound brain concentration well below the level necessary for receptor occupancy. In mice lacking the mdr1a transporter, brain penetration increases significantly, resulting in observable central opioid effects (Schinkel et al., 1996). A candidate compound with efflux liability similar to loperamide possesses substantial intrinsic potency but is unable to reach a central target due to this transport mechanism, exemplifying the physically predictable mismatch that the transport gate is designed to reveal. When applied early, the transport gate could have identified such candidates based on efflux ratios and low predicted Kp,uu values prior to conducting in vivo efficacy studies. This would guide efforts toward reducing polar surface area and P-gp liability, preventing resource expenditure on ineffective compounds. This framework is specifically designed to detect physically predictable failures of this nature, and emphasizing initial screening through the transport gate is essential.

6.2 A thermodynamic and transport success the framework would have passed: HIV protease and the value of explicit exposure coupling

The development of HIV protease inhibitors offers valuable insights, as early agents demonstrated high potency but suffered from inadequate and inconsistent oral bioavailability. Progress was only achieved when exposure was intentionally enhanced, such as through pharmacokinetic boosting with low-dose ritonavir to inhibit cytochrome P450 3A4 metabolism (Zeldin and Petruschke, 2004). Analyzing this history within the framework, it becomes evident that the thermodynamic barrier relates to binding affinity; however, the critical factor distinguishing viable candidates from non-viable ones was the coupling of potency with achievable free exposure. The ultimate resolution involved an intervention to improve exposure rather than further increasing potency. The framework would have approved the chemotype while identifying exposure as the primary binding constraint to be addressed, aligning with the approach ultimately adopted by the field.

6.3 A biology-addressable failure the framework would correctly have stayed silent on: CETP inhibition

The CETP inhibitors successfully elevated HDL cholesterol levels as intended; however, they failed to improve, and in some instances worsened, cardiovascular outcomes. Notably, torcetrapib in the ILLUMINATE trial was associated with an increase in both cardiovascular events and all-cause mortality, despite a significant increase in HDL cholesterol (Barter et al., 2007). All physical assessments would have confirmed that these compounds engaged their targets, achieved sufficient exposure, and modulated the intended biomarkers. The framework appropriately refrains from making claims regarding their clinical efficacy, as the failure was rooted in biological hypothesis failure, and in the case of torcetrapib, was additionally attributable to compound-specific off-target effects, aspects that lie beyond physical analysis. This discipline is precisely what the framework is designed to enforce: it must not attribute predictive success to failures that are inherently unobservable. Its lack of comment in this context is deliberate and not indicative of a deficiency.

6.4 What a proper prospective validation would require

The illustrations provided above align with the conceptual framework; however, they do not constitute formal validation. A credible validation study would involve assembling a representative dataset of historical programs, containing documented early-stage physical parameters, such as Kd or Ki with associated uncertainty, free fraction, an early Kp,uu estimate, an occupancy target, and a tolerated-exposure estimate, together with resultant development outcomes. This study would apply the decision gates solely based on information available at each decision point, and would evaluate the frequency with which physical indicators correlated with physically attributable failures, the proportion of passing candidates devoid of physically predictable failures, and the rate of false alarms among drugs that ultimately achieved success. This dataset construction and blinded analysis represents the essential next step, and is a substantial effort partly because early-stage parameters are reported inconsistently in the historical record. The present work establishes the underlying logic and scope; the quantification of its predictive performance remains a subject for future investigation.

7 Limitations of the framework

The framework has explicit limitations that define its domain.

7.1 Necessary, not sufficient; flags, not verdicts

The most important limitation is interpretive. A physical flag identifies a major feasibility risk early; it does not declare impossibility. Programs are routinely rescued by dosing strategy, formulation, alternative or local delivery, pharmacokinetic boosting, or the discovery that the true occupancy requirement is lower than assumed. The framework’s outputs are therefore risks to be reduced by measurement and engineering, and any wording suggesting that gate failure is terminal should be read in this light.

7.2 Relationship to richer models

Where data justify it, QSP and PBPK models supersede the gating calculation, computing the same quantities more rigorously and handling interactions the gates only flag. The framework is a triage layer upstream of those models, not a substitute for them.

7.3 Nonlinear pharmacokinetics and TMDD

The simplistic occupancy calculations presume approximately linear dose-exposure relationships. Saturable absorption or metabolism, along with target-mediated drug disposition, in which target binding constitutes a significant clearance pathway, necessitate more comprehensive modeling (Mager and Jusko, 2001). These phenomena are especially prevalent in high-affinity biologics and targets with high abundance.

Active transport is distinguished from passive diffusion, which is the predominant mode of transport discussed. Uptake and efflux transporters can significantly influence tissue distribution in ways that are not foreseeable solely based on physicochemical properties and necessitate dedicated characterization (Giacomini et al., 2010).

Computational prediction methods, including free-energy perturbation, machine learning, and physics-based simulation, are progressively informing optimization processes within established chemical series (Cournia et al., 2017). However, forecasting de novo activity for novel scaffolds continues to pose significant challenges. The framework described does not rely on first-principles structure-activity relationship (SAR) predictions.

7.4 Assay dependence and uncertainty

All inputs (Kd, koff, permeability, solubility, Kp, uu) are assay-dependent and uncertain. Regime validation is essential but inconsistently performed (Jarmoskaite et al., 2020). Because gate decisions can be sensitive to two-fold input uncertainty, the framework should be used with explicit uncertainty propagation rather than point estimates.

8 Conclusion

Drug discovery functions within the parameters established by physical constraints that delineate the operational boundaries for effective therapeutics. Thermodynamics restricts binding at feasible concentrations; kinetics determines the duration of engagement over pertinent timescales; transport processes influence whether pharmaceuticals reach their designated sites of action; and microenvironmental factors adjust properties under pathological conditions. These limitations are inherent to physical laws, and biological systems cannot circumvent them.

Physical constraints, however, do not determine therapeutic success. The primary cause of clinical failure, namely, inadequate efficacy despite confirmed engagement, reflects biological hypothesis errors that physical analysis cannot anticipate. Most genuine failures are multi-causal. The feasibility-gating framework accepts this reality and clarifies what physical analysis can legitimately accomplish: identify early and explicitly the candidates likely to fail on physical grounds, thereby reserving the more challenging experimental work of validating biological hypotheses for candidates that are at least testable. Its contribution is organizational and procedural, advancing established questions to the lowest feasible cost, recording the responses in an auditable manner, treating the gates as a coupled system, and passing qualified candidates to the more comprehensive QSP and PBPK models that supersede it.

The framework does not claim to predict which compounds will succeed; rather, it only identifies those that carry predictable physical risks. This distinction delineates the boundary between legitimate application and overreach. Physical reasoning serves as a filter, not a navigation instrument: it highlights certain foreseeable risks but cannot assure success for candidates that meet the criteria. The biological hypothesis must still undergo experimental testing, ultimately involving patient studies. The blinded retrospective dataset identified here as the means of assessing the framework’s predictive performance is the essential step from a coherent theoretical approach toward a validated decision-making tool.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. For grammar correction, figure drawing, and checking that the article does not read as written by AI.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Ferdinand Molnár, Nazarbayev University, Kazakhstan

Reviewed by: Jesús Valdiviezo, Pontifical Catholic University of Peru, Peru

Nathaniel Charest, Chemical Insights Research Institute, United States