An Evidence-Grounded Research Assistant for Functional Genomics and Drug Target Assessment

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Biology
• Computational Biology

Background:

• Biological data resources are expanding, but their effective use is hindered by fragmentation and the need for domain expertise.
• Current large language models in biomedicine often produce unsupported or incorrect claims, lacking provenance.
• Synthesizing reliable conclusions from diverse biological data is labor-intensive and challenging.

Purpose of the Study:

• To introduce Alvessa, an evidence-grounded agentic research assistant designed for verifiability in biological research.
• To address limitations of general-purpose language models in handling complex biomedical data and ensuring claim accuracy.
• To support reproducible and verifiable AI-assisted biological research.

Main Methods:

• Alvessa integrates entity recognition, pre-validated biological tools, and data-constrained answer generation.
• It performs statement-level verification against retrieved records, flagging unsupported claims.
• Evaluation was conducted on dbQA (LAB-Bench) and GenomeArena benchmarks covering diverse biological questions.

Main Results:

• Alvessa significantly improves accuracy compared to general-purpose language models on biomedical QA tasks.
• It performs comparably to coding-centric agents while ensuring fully traceable outputs.
• Detection of fabricated statements critically depends on access to retrieved evidence, as shown by adversarial testing.

Conclusions:

• Alvessa offers a verifiable and accurate approach to AI-assisted biological research, overcoming key limitations of existing LLMs.
• The system demonstrates practical application in drug discovery, identifying candidate targets missed by literature-centered methods.
• Alvessa and GenomeArena are released to foster reproducible and reliable AI-driven scientific discovery.