Insilico Medicine Highlights AI Model Performance on Drug Potency Benchmark

According to a recent LinkedIn post from Insilico Medicine, the company’s ScienceAIBench series has moved to a core drug discovery challenge focused on bioactivity prediction. The post describes a benchmark that evaluates how well leading AI models predict IC50, a key measure of drug potency used to prioritize lead candidates in optimization.

The post highlights testing on the BindingDB Cold Drug split from the Therapeutics Data Commons, a setting designed to assess generalization to unseen chemical space. Models evaluated include Opus 4.5, Sonnet 4.5, GPT 5.2, Grok 4.1, GPT 5.1, and Deepseek 3.2, with performance measured by Spearman and Pearson correlation.

According to the benchmark results cited, Opus 4.5 led performance with Spearman 0.347 and Pearson 0.349, while GPT 5.2 followed and showed fewer extreme outliers on this difficult task. Deepseek 3.2 lagged with a Spearman score of 0.128, and the post notes that all models produced negative R² values, underscoring the challenge of predicting activity for completely novel drug scaffolds.

For investors, this benchmarking activity suggests Insilico is positioning itself as a rigorous evaluator of foundation models for drug discovery tasks rather than relying solely on headline performance metrics. Demonstrating systematic assessment of model generalization in realistic “cold start” settings could support the company’s credibility with biopharma partners, potentially enhancing its ability to monetize AI platforms and collaborate on high‑value discovery programs.

The focus on modest but consistent predictive signal also indicates that, despite rapid AI progress, bioactivity prediction for novel chemistry remains an unsolved problem, leaving room for proprietary model development and data assets. If Insilico can demonstrate internally superior performance to the public models it evaluates, the gap highlighted in this series may translate into competitive differentiation and pricing power in AI‑enabled drug discovery services.