Author: Lucia Urcelay Ganzabal, ML Research Scientist in the Centre for Genomic Regulation (CRG)

Link to full report: https://docs.google.com/document/d/1aKT_qfJaBFO3UVmnm54ez34dEvlMkEZz/edit?usp=sharing&ouid=104967016980658516197&rtpof=true&sd=true

This submission looks into the feasibility of using an autonomous AI agent to carry out a complete de novo protein binder design campaign against RBX1 under time constraints. With only 15 hours available before submission, the primary objective was to test how far such a system could progress independently, from literature review to a submission-ready FASTA file, based on a single prompt, rather than to optimize a specific methodology. The agent selected for this experiment was Biomni, and the focus was placed on observing its end-to-end performance with minimal user intervention. The workflow included structure selection (favoring NMR structure 2LGV), hotspot identification, backbone generation using RFdiffusion, and sequence design with ProteinMPNN. Three parallel campaigns targeting different functional regions of RBX1 produced 600 backbones and 9,600 sequences, which were filtered and clustered to yield 100 diverse candidates meeting all competition requirements.

The final designs show strong in silico metrics, including favorable ProteinMPNN scores, high diversity, and clear compliance with novelty constraints. Throughout the process, the agent independently addressed a range of technical and methodological issues, such as execution errors, resource limitations, and refinement of filtering criteria to preserve valid designs. While these results suggest that autonomous systems can meaningfully accelerate complex computational workflows, the experiment also highlights a trade-off in reduced direct engagement with the scientific process. Overall, this work serves as a practical assessment of current capabilities in autonomous protein design, with experimental validation needed to determine the true effectiveness of the generated binders.