AI Benchmarking Effort Highlights Model Performance in Ophthalmic Drug Discovery

According to a recent LinkedIn post from Insilico Medicine, the company’s ongoing “ScienceAIBench” series is now evaluating AI models in ophthalmic drug discovery, an area characterized as high-stakes and technically complex. The benchmark focuses on the ability of multiple large models to identify valid clinical-stage targets across retinal, corneal, and optic nerve diseases.

The post describes a metric based on Mean Average Precision at TopK for indications including macular degeneration, corneal dystrophy, diabetic retinopathy, glaucoma, pterygium, retinitis pigmentosa, and thyroid-associated ophthalmopathy. Models assessed include DeepSeek R1, Gemini 2.5 Pro, Opus 4.1, Grok 4, and GPT-5, with performance reported at the indication level.

According to the shared results, GPT-5 appears to lead in several complex and multi-factorial retinal and systemic conditions, with relatively high scores in macular degeneration, diabetic retinopathy, and thyroid-associated ophthalmopathy. DeepSeek R1 is described as strongest in hereditary indications, achieving a perfect score in corneal dystrophy and leading in retinitis pigmentosa.

The post also highlights performance dispersion among models, including a pronounced gap in corneal dystrophy where GPT-5 scores zero while DeepSeek R1 scores perfectly, suggesting a potential training-data blind spot for some large general models. Other models show more targeted strengths, with Gemini 2.5 Pro leading in glaucoma and Grok 4 slightly ahead in pterygium.

For investors, this benchmarking activity suggests that Insilico Medicine is positioning its platform and datasets as a reference environment to stress-test frontier AI models in disease-relevant tasks rather than generic benchmarks. Demonstrated differentiation across indications may support the company’s value proposition in precision target discovery and could strengthen its attractiveness as a partner to AI model developers and pharmaceutical firms.

If Insilico can monetize such benchmarking as a service or embed it into its own drug discovery workflows, the insights into model strengths and blind spots could translate into more efficient target selection and reduced R&D risk. Over time, consistent publication of granular, domain-specific AI performance data may enhance Insilico’s standing in AI-native biotech, potentially supporting future capital-raising, partnerships, or platform-licensing opportunities.