Why do animal models fail to translate into clinical osteoarthritis pain therapies?

Only 7–10% of animal successes carry over

The paper frames the problem with a benchmark: about 64% of treatments that show efficacy in animal models advance to clinical trials, but only 7–10% demonstrate comparable benefit in humans. Osteoarthritis is presented as a case that concentrates this gap, because pain is its most debilitating symptom yet remains the hardest endpoint to model.

The authors' question is therefore narrow and practical. Rather than asking whether animal models reproduce cartilage damage, they ask which models and pain-assessment methods predict drug efficacy for OA pain, treating pain as the primary endpoint on both the animal and human sides. The answer they build is a comparison of trial outcomes with the corresponding preclinical studies for the same compounds.

Human OA pain is chronic and mixed in origin

In the paper's account, OA pain is not a simple sensory readout of joint damage. Pain severity in knee OA often does not correspond closely with radiographic findings; synovitis and bone marrow lesions show stronger associations with pain intensity in nociceptive pain. Sustained inflammation can drive aberrant nerve sprouting and peripheral sensitization, and persistent peripheral input can contribute to neuropathic-like symptoms.

The scale of that neuropathic component matters for model design. The paper estimates that only 23% of people with OA show neuropathic pain features, consistent with the definition of neuropathic pain as pain caused by a lesion or disease of the somatosensory nervous system. Sensitization, both peripheral and central, can make pain more intense, widespread, and persistent than structural pathology alone would predict.

Comorbidities and ageing add further layers. OA is associated with up to 30 coexisting conditions, including depression, anxiety, and diabetes, all of which can influence pain perception; anxiety can heighten sensitivity through limbic modulation of descending pain pathways, and diabetes can alter afferent input through peripheral neuropathy. In older cohorts, pain severity may therefore reflect joint pathology, peripheral sensory change, central sensitization, and cognitive modulation together.

Common models capture evoked, early-phase pain

The review divides OA animal models into spontaneous models, where disease develops through ageing or genetic predisposition, and induced models triggered by mechanical, surgical, chemical, or genetic intervention; such models have been established in at least 18 species. Surgical models such as destabilization of the medial meniscus in mice or anterior cruciate ligament transection in rats and rabbits resemble post-traumatic OA and give rapid, reproducible degeneration. The chemical monoiodoacetate (MIA) model produces early immune activation over days 0–7 followed by OA-related pain around days 14–28.

The measurement side is where the mismatch sharpens. Animal assays commonly measure evoked pain: von Frey and knee bending tests for mechanical allodynia, Randall–Selitto paw pressure for mechanical hyperalgesia, acetone for thermal allodynia, and Hargreaves or hot and cold plate tests for thermal hyperalgesia. The paper notes that evoked pain is reported in roughly 15–50% of patients with chronic pain. Spontaneous pain in animals is instead inferred from static or dynamic weight bearing, gait analysis, locomotor activity, wheel running, and burrowing.

Clinical studies rely on patient-reported instruments: visual analogue and numeric rating scales for intensity, the Patient Global Assessment for disease activity, and multidimensional questionnaires such as WOMAC and KOOS for pain, stiffness, and function, with quantitative sensory testing available for mechanistic assessment of pressure and thermal thresholds. The two sides are thus not measuring the same construct by default.

NSAIDs and corticosteroids mostly agree across species

Not every drug class exposes the gap. NSAIDs showed good concordance between clinical and preclinical efficacy independently of dose and route, with efficacy in nearly all cases except paracetamol. Sustained-release paracetamol failed to improve WOMAC pain scores versus placebo in knee OA patients, and MIA rat studies reported only transient reductions in pain-related behaviors with little efficacy during chronic disease stages.

Extended-release triamcinolone acetonide illustrates a similar pattern with a caveat. Single intra-articular injections significantly improved pain outcomes in knee OA patients, with reductions in weekly numeric rating scale scores maintained up to 12 weeks, and preclinical studies in MIA rats generally supported analgesic efficacy, in some cases for up to 60 days. Efficacy was less consistent in surgically induced models such as meniscal transection, where improvements in mechanical allodynia and weight-bearing deficits were not observed, while positive effects were reported in rabbits and C57BL/6 mice.

A timing discrepancy qualifies this concordance. In clinical trials NSAID therapy generally starts after at least 3–6 months of symptoms, whereas animal studies often begin treatment at OA induction or early disease; 36 of the 45 preclinical NSAID studies analyzed used the MIA model. Lack of efficacy appeared mainly when NSAIDs were given at later MIA stages or in the single primate study, which the authors read as evidence that species differences and treatment timing both shape outcomes.

Lacosamide and capsaicin split between species

The clearest failures come from compounds whose preclinical testing used short, early, evoked-pain designs. The anticonvulsant lacosamide reduced tactile allodynia and mechanical hyperalgesia modestly in MIA rat models, particularly at the higher dose at 7 and 14 days after induction, but oral lacosamide at 400 mg/day failed to improve OA pain versus placebo after 8 weeks in a clinical trial. All preclinical lacosamide studies used MIA rats, and efficacy was often assessed from minutes to hours after dosing; only one study evaluated 28-day chronic administration, and it began treatment at OA induction, unlike patient enrollment timing.

Topical and intra-articular capsaicin show a similarly uneven picture. In one trial, 5% topical capsaicin but not the 1% formulation met the primary efficacy endpoint by reducing WOMAC pain scores at day 35 relative to baseline, and both formulations outperformed vehicle at later follow-up on days 64 and 94. Two studies of intra-articular transcapsaicin, one with repeated injections over 12 weeks and one with a single injection over 52 weeks, showed no superiority over placebo.

Preclinical capsaicin results were heterogeneous in matching ways. Intra-articular capsaicin at 30–100 micrograms rapidly improved mechanical hyperalgesia within 20 minutes in the Randall–Selitto test; prophylactic 0.5% capsaicin before MIA induction attenuated pain behaviors over four weeks; but 1 microgram per microliter given after OA was established failed to improve mechanical allodynia or weight-bearing deficits. One positive study also lacked a vehicle-treated control, comparing only against untreated MIA animals, which limits interpretation.

Duloxetine sits on the concordant side: oral 60–120 mg/day improved pain and function over 8 to 52 weeks in placebo-controlled trials, and MIA rat studies consistently reduced mechanical hypersensitivity, tactile allodynia, and functional impairment after oral, intraperitoneal, or subcutaneous dosing. Yet even here the preclinical studies assessed efficacy from minutes to hours after dosing, in contrast to the chronic treatment in trials.

Biologics limit direct animal-to-human comparison

NGF blockade is the strongest analgesic story in the review. Fasinumab, given subcutaneously at 1–6 mg every 4–8 weeks over 4–24 weeks, significantly improved pain and functional outcomes versus placebo across several studies, and a single intravenous dose also showed efficacy at 8 weeks. Tanezumab showed robust efficacy across multiple trials with intravenous or subcutaneous administration over treatment periods from 8 to 72 weeks, and preclinical studies using MIA-induced OA or medial meniscal models supported the analgesic efficacy of NGF blockade.

The authors flag a structural limit on this comparison: because biologics are species-specific, human-based biologics are needed to avoid immune reactions, which restricts direct comparison between human and animal studies. That makes the biologic class a place where the usual model-to-trial inference is hardest to evaluate, independent of how well pain is measured.

Match treatment timing and spontaneous pain readouts

The review's central recommendation is not a single model but a design principle. Because OA is too complex to reproduce every patient characteristic in one preclinical system, experimental design should be guided by clearly defined target populations and clinically relevant phenotypes. The authors state that a clear correspondence between preclinical model characteristics and human endotypes is not currently possible given limited information on endotype-specific features, but that consensus on correlations with clinical phenotypes is feasible and necessary.

Two concrete changes follow. First, treatment should begin after the structural and inflammatory changes that progressively produce pain and allodynia have been established, mimicking a therapeutic intervention rather than a prophylactic one. Second, pain assessment should include spontaneous pain-related behaviors such as gait analysis and static weight bearing, which more closely resemble the functional and pain-related outcome measures used clinically. The authors connect this refinement to the Replacement, Reduction, and Refinement principles, framing it as both an ethical requirement and a scientific necessity.

Concordance depends on design, not model label

The most useful takeaway is a ranking of where to look first. Across the classes reviewed, NSAIDs and corticosteroids showed good clinical–preclinical agreement, duloxetine agreed, lacosamide and intra-articular transcapsaicin did not, and NGF blockade was efficacious clinically but constrained by species-specific biologics. The dividing line the paper draws is not surgical versus chemical or rat versus mouse; it is whether the animal experiment started treatment after disease establishment and measured spontaneous, function-linked pain rather than only brief evoked responses.

That reframes how a preclinical OA pain result should be read. A positive finding in young, single-sex animals treated at induction and tested minutes to hours after dosing carries little predictive weight for a chronic therapy in older patients with comorbidities and mixed nociceptive–neuropathic pain. The review's own threshold for a translationally meaningful study is therefore a treated, established-disease model with spontaneous pain outcomes and a duration that matches the intended clinical regimen.

Sources used in this answer