Salt AI Showcases Recursive AI Workflow for Small-Molecule Optimization

According to a recent LinkedIn post from Salt AI, the company is emphasizing workflow architecture over pure model quality in AI-driven drug discovery. The post describes a small-molecule optimization pipeline built on Salt’s visual platform that uses Anthropic’s Claude Opus 4.6 to generate candidate molecular structures and iteratively refine them.

The system reportedly scores each candidate in real time using Boltz docking, Lipinski drug-likeness rules, ADMET properties, and synthetic accessibility, applying these physical and chemical constraints at every step. The workflow is described as recursive, reusing the same model across multiple rounds while accumulating context to improve output quality over time.

As an illustrative test, the post notes that Salt AI’s team started from the structure of bacampicillin and ran 15 optimization rounds before searching resulting molecules in PubMed and PubChem. The pipeline is said to have independently converged on known antibiotics ampicillin and amoxicillin, along with 11 additional molecules that did not appear in those databases, each accompanied by a scored property profile.

The post argues that this outcome indicates the system is guided by physics-based scoring and constraints rather than simple memorization from model training data, given that docking scores and Lipinski compliance derive from external calculations. According to the post, this suggests that recursive feedback loops combined with multi-model scoring can enhance output quality largely independent of the specific underlying large language model.

Salt AI highlights that the workflow was built on its visual platform, where each node is visible and auditable, emphasizing transparency in how drug candidates are generated and evaluated. The company explicitly notes that this is not yet a fully fledged drug-discovery engine, positioning the work instead as an example that constrained, recursive workflows can yield pharmacologically interesting results with clear traceability.

For investors, the described approach points to a potential competitive angle in focusing on system architecture and constraint-driven optimization rather than solely on frontier models. If validated and adopted by pharmaceutical or biotech partners, such transparent, auditable workflows could support platform-licensing opportunities, deeper collaborations in early-stage discovery, and potentially usage-based revenue tied to computational screening at scale.

The indication that results may “compound as models get better” suggests that Salt AI’s value proposition could benefit from ongoing improvements in foundation models without requiring the company itself to develop those models. This may position Salt AI as an infrastructure or tooling layer within AI-driven drug discovery, potentially improving its strategic relevance in the healthtech and pharma-tech ecosystems as demand grows for explainable, reproducible AI workflows in R&D pipelines.