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 molecular structures and iteratively refine them.

The workflow reportedly scores each candidate in real time using Boltz docking, Lipinski drug-likeness, ADMET assessments, and synthesizability metrics, and runs recursively over at least 15 rounds while reusing the same base model. According to the post, performance improves as context accumulates and physical constraints are applied at each step, suggesting the architecture may enhance model outputs without changing the underlying foundation model.

As an illustrative test, the post indicates the team began with the known structure of bacampicillin and carried out 15 optimization rounds, then compared the results against PubMed and PubChem. The workflow is said to have independently converged on ampicillin and amoxicillin, two widely prescribed antibiotics, alongside 11 additional molecules that the team reports did not appear in the referenced literature and databases.

The LinkedIn post argues that convergence on known antibiotics is driven by physics-based scoring and constraint systems rather than training-data memorization, noting that Boltz docking, Lipinski criteria, and synthetic accessibility are grounded in external computation and chemistry. The post further suggests that recursive feedback and multi-model or multi-metric scoring could systematically improve molecular design outputs and that such effects may compound as core AI models advance.

Salt AI highlights that the pipeline was created on its visual, node-based platform, where each component of the workflow is visible and auditable to users. While the post explicitly stops short of positioning this setup as a complete drug discovery engine, it portrays the experiment as evidence that constrained, transparent workflows may yield pharmacologically interesting candidates and richer data outputs for evaluation.

For investors, this update points to Salt AI’s strategic focus on tooling and workflow orchestration rather than owning primary drug assets, potentially positioning the company as an enabling technology provider within pharma and biotech. If adopted by industry partners, such capabilities could support recurring software or platform revenues, though regulatory validation, integration into existing R&D pipelines, and real-world hit-to-lead success will be critical determinants of long-term commercial impact.