ResistIA: Reasoning-Guided Agentic Evaluation of Synthetic Metal-Resistance Genes from Conditional Genomic Foundation Models

José Vásquez-Bastías

Track: long paper (up to 10 pages)

Keywords: Genomic foundation models, Agentic evaluation, Metal-resistance gene generation, Tool-augmented scientific reasoning

TL;DR: ResistIA is an agentic, validation-aware framework that refines synthetic metal-resistance genes from a LoRA-tuned genomic model using DIAMOND, ORF/physchem checks, and ChatNT evidence.

Abstract:

Genomic foundation models can generate plausible coding DNA, but turning raw generations into scientifically useful candidates requires reliable post hoc evaluation and robust model-selection criteria. We present ResistIA, a reasoning-guided agentic evaluation and refinement framework for synthetic metal-resistance genes generated from a LoRA-adapted GenomeOcean-500M model.

ResistIA combines three complementary evaluators—(i) DIAMOND blastx protein homology, (ii) ORF/translation/physicochemical checks, and (iii) ChatNT-based semantic/regulatory Q&A with embedding similarity—under an auditable decision policy that selects tools based on runtime availability and evidence needs. We study three progressively stronger versions of the framework: V1 (baseline multi-tool evaluation), V2 (closed-loop reweighting over generated batches), and V3 (validation-aware closed-loop optimization with score smoothing and early stopping).

Across internal benchmark runs, V3 improves reliability of iteration selection and achieves the strongest validation-selected performance, with generated-set DIAMOND pass rate up to 0.283 and ORF success up to 0.975, while preserving high ChatNT semantic similarity to natural references (centroid cosine ~0.97–0.99). We also observe a practical trade-off: V2 can produce sharper peak improvements in some physicochemical gaps, whereas V3 yields more reproducible selection under iteration-to-iteration variability.

These results position ResistIA as a compact, reproducible testbed for reasoning over heterogeneous biological evidence in agentic scientific workflows, and a practical framework for evaluator-guided refinement of genomic sequence generators.

Presenter: ~José_Vásquez-Bastías1

Format: Yes, the presenting author will attend in person if this work is accepted to the workshop.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.

Submission Number: 126