← All proceedings

TheIIER International Conference

Proceedings cover: TheIIER International Conference

Marrakesh,Morocco 12-08-2026

2 papers

  1. Putative Molecular Mechanism of BHQ-Mediated Modulation of CaV2.1 Channels: Insights from Functional Studies and Recent Cryo-EM Structures

    Akira Inagaki

    Paper ID II-MPSMARRA-120826-34370 · Pages 1-4

    Abstract

    2, 5-Di-tert-butylhydroquinone (BHQ), originally identified as a SERCA inhibitor, also directly modulates the gating of the P/Q-type calcium channel CaV2.1. BHQ suppresses voltage-dependent activation, slows deactivation, and enhances Ca²⁺-dependent facilitation, indicating an allosteric effect on channel gating rather than pore blockade. Based on recent cryo-EM structures of human CaV2.1, the hydrophobic cavity at the interface between the voltage-sensing domain and pore domain, particularly around the S4–S5 linker, represents the most plausible BHQ-binding site. Binding at this interface may alter electromechanical coupling, providing a unified mechanism for the observed changes in activation, deactivation, and Ca²⁺-dependent facilitation. Future structural and functional studies are required to validate this proposed mechanism.

    Keywords - CaV2.1, 2, 5-Di-tert-Butyl Hydroquinone (BHQ), Gating Modulation, Cryo-EM Structure

    Download PDF
  2. Development and Implementation of a SNOMED CT-Specialized Large Language Model via QLoRA Fine-Tuning and Local Deployment with Ollama

    Daniel Antokoletz Huerta, Fernando Mario La Rosa

    Paper ID IRES-ICANSMHSRABAT-120826-33204 · Pages 5-10

    Abstract

    This article describes the complete development of a large language model (LLM) specialized in the clinical terminology of the SNOMED CT standard, covering three phases: training corpus construction, fine-tuning using Quantized Low-Rank Adaptation (QLoRA), and local deployment with Ollama. The final corpus (v5_final) comprises 2,186,183 examples across five semantic types generated from the SNOMED CT Release Format 2 (RF2) files and presents a statistically uniform distribution with a P95/mean ratio of 1.78 tokens — comparable to standard NLP corpora. The complete fine tuning on this corpus is currently in progress, with approximately 60 days of processing remaining. Initial tests with a reduced corpus of 50,000 examples, carried out on Mistral 7B Instruct v0.2 using consumer grade hardware (NVIDIA Turing GPU CC 7.5, 6 GB VRAM, float32), achieved a minimum loss of 0.370 and a token accuracy of 91.5%. We document the technical implementation errors, the numerical instability observed during checkpoint resumptions when processing the full corpus, and the formulation of the Cold Start Optimizer Resumption (CSOR) framework [1] as a solution to the incompatibility between the AdamW optimizer state and the restarted learning rate scheduler. The specialized model is integrated into the Ollama ecosystem for local inference without reliance on cloud services.

    Keywords - SNOMED CT, QLoRA, LoRA, Mistral-7B, fine-tuning, Ollama, CSOR, AdamW, clinical NLP, medical terminology, consumer GPU, fp32, local deployment.

    Download PDF