Reflective AI edits help protein prediction with limited data labels

Does reflective LLM feedback survive a wet-lab-scale label budget?

Artificial Intelligence

Summary

The paper studies whether using reflective feedback from a large language model (LLM) can improve how well a computer predicts biological data when only a small amount of measurement data can be collected at a time. The researchers replayed multiple biological tests using limited, phased data gathering to see if letting the AI suggest model changes or guide data collection makes predictions better. They found that AI-suggested model edits can improve prediction quality, but using the AI to choose which data to gather next or to repeatedly revise the model did not clearly help. This suggests focused AI edits have value, while recursive feedback needs more evidence with this limited data setup.

What this means in practice

  • For biotechnology data teams: Enhance biological prediction models by applying AI-driven focused edits under constrained experimental data budgets.
  • For drug discovery researchers: Improve predictive modeling of protein assays through selective AI-guided model adjustments during early-stage experimentation.

Tested on one dataset.

Authors

Haoxuan Zeng, Xin Luo

Abstract

Can reflective language-model feedback improve a biological predictor when each cycle buys only one plate of labels? We separate three possible sources of benefit: proposing model edits, choosing measurements, and revising decisions through repeated feedback. We replay $36$ ProteinGym assays, starting with $96$ labels and acquiring $96$ per round for six rounds, with a sealed test set and a shared edit schema. The pre-specified comparison of full LLM control against random edits detects no advantage in test rank correlation ($Δz=-0.008$, $95\%$ CI $[-0.034,+0.018]$); equivalence is also unestablished. Exploratory ablations nevertheless identify useful model editing. With random acquisition fixed, LLM edits improve Fisher-transformed test Spearman by $0.043$ over no edits; with acquisition fixed to upper confidence bound, they improve normalized top-variant recovery by $0.023$. LLM acquisition alone does not detectably outperform fixed adaptive rules, and repeated feedback does not detectably outperform one initial recommendation under full control. Thus, the observed gains support constrained specialist editing, while the incremental value of recursive reflection remains unestablished at this label budget.
View PDFOpen arXiv