[Paper Review] AptaFind: A lightweight local interface for automated aptamer curation from scientific literature

Geoffrey J. Taghon|arXiv (Cornell University)|Jan 12, 2026

Biomedical Text Mining and Ontologies0 citations

TL;DR

AptaFind combines local language models with deterministic regex methods to curate aptamer data from literature via a three-tier, on-device pipeline, achieving ~84% coverage across tiers and ~953 targets/hour processing speed.

ABSTRACT

Aptamer researchers face a literature landscape scattered across publications, supplements, and databases, with each search consuming hours that could be spent at the bench. AptaFind transforms this navigation problem through a three-tier intelligence architecture that recognizes research mining is a spectrum, not a binary success or failure. The system delivers direct sequence extraction when possible, curated research leads when extraction fails, and exhaustive literature discovery for additional confidence. By combining local language models for semantic understanding with deterministic algorithms for reliability, AptaFind operates without cloud dependencies or subscription barriers. Validation across 300 University of Texas Aptamer Database targets demonstrates 84 % with some literature found, 84 % with curated research leads, and 79 % with a direct sequence extraction, at a laptop-compute rate of over 900 targets an hour. The platform proves that even when direct sequence extraction fails, automation can still deliver the actionable intelligence researchers need by rapidly narrowing the search to high quality references.

Motivation & Objective

• Address the scattered aptamer literature problem and reduce manual curation effort.
• Develop a local, cloud-free pipeline combining language models with deterministic parsing for reliability.
• Provide three-tier outputs (direct sequences, curated leads, exhaustive literature) to balance precision and coverage.
• Validate the approach on UT database targets and quantify performance across tiers.
• Deliver open-source software with an emphasis on privacy and reproducibility.

Proposed method

• Integrate a three-tier intelligence architecture: direct sequence extraction (Tier 1), curated leads (Tier 2), and exhaustive literature discovery (Tier 3).
• Use local 1B parameter Llama3.2 model for semantic understanding with on-device processing.
• Combine language model guidance with deterministic regex pipelines (Minimum Agentic Flow, MAF) for validation, deduplication, and formatting.
• Extract nucleotide sequences (20–100 nt) and binding data (Kd, Ki) with unit preservation via regex, and harmonize data with LM context.
• Perform multi-source discovery from local PDFs, PubMed/PMC, and bioRxiv, with supplement harvesting and browser automation when needed.
• Validate sequences against biological constraints (length 20–100 nt, GC 20–80%, 5’→3’ orientation) and deduplicate at 100% identity across sources.

Figure 1: AptaFind implements a three-tier research intelligence approach ensuring every search delivers actionable value. (A) System Architecture: Multi-source literature discovery combines PubMed, PMC, and bioRxiv searches with supplement harvesting and browser automation for comprehensive coverag

Experimental results

Research questions

• RQ1 How effective is AptaFind’s three-tier output in capturing aptamer data across diverse literature sources?
• RQ2 What are the per-tier recovery rates and processing speed on typical hardware?
• RQ3 Does the Minimum Agentic Flow principle improve reliability over LM-centric or regex-centric approaches?
• RQ4 What are the limitations regarding paywalled content, image-based sequence extraction, and complex tables?
• RQ5 Can the approach generalize to other literature-mining domains beyond aptamers?

Key findings

• Tier 3 (Literature Discovery) achieves 84.0% ± 3.5% coverage across 100-target samples.
• Tier 2 (Research Leads) achieves 84.0% ± 3.5% coverage across 100-target samples.
• Tier 1 (Direct Extraction) achieves 79.3% ± 0.6% coverage across 100-target samples.
• Processing speed is about 954 ± 43 targets per hour on a Mac Studio (M2 Max).
• Validation shows the MAF approach improves reliability by separating semantic understanding from deterministic data handling.
• The methodology remains local with no cloud dependencies, delivering ~1000 target queries per hour and enabling rapid literature screening.