Insilico Medicine Highlights Limitations of Frontier AI Models in 3D Drug Discovery Benchmarks

A LinkedIn post from Insilico Medicine describes new benchmark results from its ongoing #ScienceAIBench series, focusing on how large language models handle 3D protein‑ligand interactions. The post outlines a task that tests whether models can move beyond simple counts of non‑covalent interactions to accurately localize and characterize them in 3D space.

According to the post, Insilico used its Chemistry42 platform as the ground‑truth engine for protein pharmacophores and drew on the LP‑PDBBind dataset to evaluate several leading models. The benchmark assessed interaction types, residues, and interaction power, with metrics including fraction of restored interactions and powers, as well as the rate of spurious or “fake” interactions.

The post indicates that format validity was a significant hurdle, with GPT 5.1 producing fully valid answers while some newer models, such as GPT 5.2 and Nemotron 3 Nano, struggled to generate usable outputs. Even among models that produced valid responses, the results reportedly showed limited spatial accuracy, with only modest success in restoring exact interactions or their strengths.

Insilico’s write‑up highlights a high incidence of hallucinated or non‑existent bonds, particularly in models like Grok 4.1 and GPT 5.2, which allegedly predicted many interactions that do not appear in the underlying 3D structures. The post suggests that current LLMs face a “3D translation barrier,” indicating that while they may handle bulk molecular properties, they remain unreliable for precise spatial physics relevant to drug discovery.

For investors, these benchmark observations point to both a technical bottleneck and a potential opportunity in AI‑driven drug design. If Insilico’s Chemistry42 platform can serve as a robust reference standard where general‑purpose models falter, the company could strengthen its competitive position in structure‑based discovery workflows and justify continued investment in proprietary cheminformatics and AI infrastructure.

The series also underscores that the broader AI ecosystem may require specialized models or hybrid architectures to close the gap between textual reasoning and 3D molecular understanding. This dynamic could support demand for domain‑specific tools and datasets, potentially reinforcing the value of Insilico’s assets and know‑how in biopharma collaborations, platform licensing, or downstream pipeline development, though commercial impacts are not quantified in the post.