Fibrodysplasia ossificans progressiva (FOP), often called stone man syndrome, is a rare genetic disorder in which soft connective tissue progressively ossifies into bone, leading to total immobility, severe pain, and premature death. It is caused by gain-of-function mutations in ACVR1 (ALK2), a receptor kinase. Until now, no effective disease-modifying therapies exist; mainstream pharmaceutical pipelines have largely ignored it because the patient population is tiny and return on investment uncertain.

A new preprint posted to arXiv (2604.20886, April 2026, not peer-reviewed) introduces KinetiDiff, a structure-based de novo design framework that fuses a geometry-complete diffusion model with real-time gradient guidance from AutoDock Vina docking scores. Researchers generated 10,000 molecules; 9,997 were chemically valid. The best-scoring candidate achieved a docking score of -11.05 kcal/mol (predicted pKd 8.10), 19.2% better than the crystallographic reference ligand. All top 100 candidates outperformed the reference, maintained 100% Lipinski compliance, showed a median synthetic accessibility score of 2.67, and exhibited high internal diversity (0.79).

The methodology is notable: rather than post-hoc filtering, the system injects physics-based docking gradients directly into each denoising step of the diffusion process, steering generation toward high-affinity poses. Systematic ablation of four guidance strategies (Vina-Direct, a neural HNN proxy, multi-objective, and unguided) demonstrated that real-time physics guidance dominated every performance metric. The neural proxy offered a 60-fold speedup but correlated poorly with Vina (r = 0.224), exposing a domain-shift limitation.

This work goes well beyond incremental improvement. It builds on DiffDock (arxiv.org/abs/2210.01776), which first showed diffusion models could outperform traditional docking, and on earlier kinase inhibitor generation efforts such as those from Insilico Medicine. Yet the KinetiDiff paper underplays a critical gap that prior coverage of similar AI drug design stories has also missed: these molecules exist only in simulation. No synthesis, no binding assays, no cell-based ACVR1 inhibition data, and no testing in FOP patient-derived models were reported. Docking scores remain noisy approximations; historical benchmarks show only modest correlation with true binding affinities once compounds reach the bench.

Context reveals a pattern. After the 2006 discovery of the ACVR1 R206H mutation, several academic groups and small biotechs pursued small-molecule inhibitors, but clinical progress has been limited. Palovarotene, a retinoic acid receptor agonist from Ipsen/Clementia, received FDA rejection in 2023 due to side effects despite some functional benefit. Big Pharma has repeatedly deprioritized such ultra-rare kinase targets. KinetiDiff therefore exemplifies how compute-heavy AI tools can democratize lead discovery for neglected diseases: a small academic team can now generate thousands of diverse, drug-like candidates without maintaining massive high-throughput screening libraries.

Synthesizing three sources clarifies both promise and pitfalls. The primary preprint establishes the technical advance. A 2022 NEJM review (nejm.org/doi/full/10.1056/NEJMra2204582) on FOP clinical features and failed therapies highlights the decades-long translational gap this AI work aims to close. A Nature Reviews Drug Discovery perspective on AI for rare diseases (nature.com/articles/s41573-023-00768-5) notes that while generative models cut design timelines from years to weeks, attrition from computation to clinic remains >95% across the industry. What the KinetiDiff preprint under-emphasizes, and what mainstream science journalism often glosses over, is selectivity: ACVR1 sits in a crowded kinase family. Without explicit counter-screens against off-targets, many generated molecules risk toxicity.

The genuine insight is structural. By privileging physics-based guidance over learned proxies, KinetiDiff exposes the current limits of purely neural approaches in domains where training data are sparse, exactly the case for rare-disease targets. This suggests a hybrid future where diffusion models act as intelligent proposal engines, but rigorous experimental triage must follow immediately. For FOP patients, the work is cause for cautious optimism: AI can shine a light on pharma's blind spots, yet the distance from a -11 kcal/mol docking hit to a safe, orally available drug remains vast. The preprint's greatest contribution may be methodological, offering a blueprint that other neglected kinase-driven conditions could rapidly adopt while the community insists on prompt wet-lab validation.