TRPV4–Gemma–3B: A Method for Discovering Transient Receptor Potential Vanilloid 4 Inhibitors Based on Causal Language Model

Abstract

Transient receptor potential vanilloid 4 (TRPV4) is an ion channel sensitive to temperature, mechanical force and osmotic pressure, which can serve as a programmable cellular sensor and regulatory element in bioengineering for tissue regeneration, disease modelling and smart biomaterial design. This study aims to harness the outstanding knowledge transfer and downstream task performance of next generation generative language models to mine limited experimental data on small‐molecule TRPV4 inhibitors, and thereby extend their potential for designing novel inhibitor candidates. The pre‐train–fine‐tune pipeline of such models can loosely emulate a ‘degenerate’ drug repurposing simulator; yet, unlike experienced medicinal chemists, the model cannot deterministically or transparently vouch for the physicochemical plausibility of the molecules it proposes. Grounding model interpretability in target–ligand interaction energies is therefore pivotal for endowing the otherwise black‐box fine‐tuned model with reliable pharmacological knowledge. To this end, we captured structural fingerprints of known TRPV4 inhibitors and combined them with molecular dynamics analyses to quantify the contribution of individual sub‐structures to the TRPV4–inhibitor binding affinity. These efforts verified that, when generating inhibitor candidates, our model statistically favours key pharmacophoric groups while filtering out ineffective fragments. Integrating deeper pharmacological insights and more sophisticated agent‐based models into this framework will deliver even more direct and tangible benefits to domain researchers.