AutoNovo: Agentic Program Search Discovers Transferable Nucleic Acid Design Algorithms

Joel Shor; Subhashini Venugopalan

cs.LGOct 12, 2026

AutoNovo: Agentic Program Search Discovers Transferable Nucleic Acid Design Algorithms

Authors: Joel Shor, Subhashini Venugopalan

Organizations: AI Bio Design, Allen Institute & Move37 Labs · Google Research

Abstract

In-silico optimization of nucleic acids relies on predictive oracles that imperfectly capture biological function, so sequences optimized against one model may exploit model-specific artifacts and fail to generalize to independent models. Recent studies have shown that not all designers generalize equally well to independent models. In this work, we use agentic program search to iteratively propose, execute, evaluate, and revise code for nucleic-acid designers. We describe the successful strategies used to prompt the search. AutoNovo designers achieved the best observed mean among the evaluated methods on 3/3 DNA-design tasks in terms of in-silico performance, and 6/6 transfer tasks in terms of independent models. The designers come from a family of discovered nucleic acid designers rather than a single universal best designer. We plan to release the resulting designers. These results are the first steps in creating nucleic acid designers that maximize cross-model generalization, and thus improving the odds that in-silico designed sequences reflect real biology.

Figures & tables

Optimization oracle $f_{\mathrm{opt}}$	Orthogonal oracle $f_{\mathrm{orth}}$	$\boldsymbol{\lambda}$
BPNet ATAC	ChromBPNet	$1$
ChromBPNet	BPNet ATAC	$1$
BPNet ATAC	Enformer	$0.01$
ChromBPNet	Enformer	$0.01$
BPNet ATAC + ChromBPNet	none	$1$

Table 1: Oracle configurations used during ERA program search. The designer can query

f_{\mathrm{opt}}

, the sum of the listed optimization-oracle scores. The orthogonal oracle

f_{\mathrm{orth}}

is hidden from the designer when present.

Figure 1: Pairwise oracle scores on the same 100 NucleoBench start sequences across the DNA panel (BPNet ATAC, Enformer K562 DNase, ChromBPNet K562).

Designer	bpnet	chrombpnet	enformer
Baselines
adabeam	$-21.90$ $\pm\,3.53$	$-37.71$ $\pm\,6.91$	$-61.28$ $\pm\,73.78$
fastseqprop	$-23.27$ $\pm\,5.64$	$-46.26$ $\pm\,6.24$	OOM
gradabeam	$-27.85$ $\pm\,3.26$	$-38.42$ $\pm\,4.86$	$-62.67$ $\pm\,75.32$
ledidi	$-40.88$ $\pm\,2.58$	$-60.08$ $\pm\,4.53$	$-124.53$ $\pm\,148.27$
AutoNovo-discovered designers

Table 2: Direct performance of sequence design after 30 minutes on CPU. Each cell reports the mean on the first line and standard deviation on the second across

N=20

start sequences; lower is better. AutoNovo designers are grouped by the optimization oracle

f_{\mathrm{opt}}

and orthogonal oracle

f_{\mathrm{orth}}

used during ERA search. The FSP, Ledidi, and PEX prefixes identify the seed algorithm. Bold indicates the best observed mean in each column.

Designer	bpnet $\boldsymbol{\rightarrow}$ chrom	bpnet $\boldsymbol{\rightarrow}$ enformer	chrom $\boldsymbol{\rightarrow}$ bpnet	chrom $\boldsymbol{\rightarrow}$ enformer	enformer $\boldsymbol{\rightarrow}$ bpnet	enformer $\boldsymbol{\rightarrow}$ chrom
Baselines
adabeam	$-10.05$ $\pm\,1.25$	$-803.77$ $\pm\,142.81$	$-4.45$ $\pm\,0.39$	$-502.41$ $\pm\,110.40$	$-1.52$ $\pm\,1.95$	$-5.37$ $\pm\,1.24$
fastseqprop	$-10.18$ $\pm\,0.90$	$-711.98$ $\pm\,321.24$	$-6.48$ $\pm\,0.61$	$-701.38$ $\pm\,183.28$	OOM	OOM
gradabeam	$-10.47$ $\pm\,1.18$	$-931.81$ $\pm\,147.39$	$-4.57$ $\pm\,0.43$	$-483.59$ $\pm\,174.38$	$-1.53$ $\pm\,1.93$	$-5.38$ $\pm\,1.23$
ledidi	$-9.99$ $\pm\,0.83$	$-1054.55$ $\pm\,184.66$	$-7.00$ $\pm\,0.81$	$-937.19$ $\pm\,233.70$	$-1.73$ $\pm\,1.83$	$-5.52$ $\pm\,1.29$
AutoNovo-discovered designers

Table 3: Cross-oracle transfer performance after 30 minutes of sequence design.

A\rightarrow B

denotes sequences designed against oracle

A

and evaluated by oracle

B

at test time. Each cell reports the mean on the first line and standard deviation on the second across

N=20

start sequences; lower is better. AutoNovo designers are grouped by the optimization oracle

f_{\mathrm{opt}}

and orthogonal oracle

f_{\mathrm{orth}}

used during ERA search. The FSP, Ledidi, and PEX prefixes identify the seed algorithm. Bold indicates the best observed mean in each column, not statistical significance.

Molecule Design

See all 43 papers

What an Autonomous Agent Discovers About Molecular Transformer Design: Does It Transfer?
Mar 30, 2026Edward WijayaNeural Architecture Search Transfer Learning
StemRIM, Inc.
General Multimodal Protein Design Enables DNA-Encoding of Chemistry
Apr 6, 2026Jarrid Rector-Brooks, Théophile Lambert, Marta Skreta +15Protein Design Molecular Generation
California Institute of Technology · Mila – Québec AI Institute · Université de Montréal +9
The Montparnasse Algorithm for RNA Design
May 25, 2026Tristan CazenaveMonte Carlo Tree Search Molecular Generation
LAMSADE, Universit´e Paris Dauphine - PSL, Paris, France

Mar 30, 2026J/K move · Enter open · S save

What an Autonomous Agent Discovers About Molecular Transformer Design: Does It Transfer?

Edward Wijaya

StemRIM, Inc.

Neural Architecture Search Transfer Learning Molecular Representation Learning Protein Representation Learning

Open paper AskPDF arXiv

Deep learning models for drug-like molecules and proteins overwhelmingly reuse transformer architectures designed for natural language, yet whether molecular sequences benefit from different designs has not been systematically tested. We deploy autonomous architecture search via an agent across three sequence types (SMILES, protein, and English text as control), running 3,106 experiments on a single GPU. For SMILES, architecture search is counterproductive: tuning learning rates and schedules alone outperforms the full search (p = 0.001). For natural language, architecture changes drive 81% of improvement (p = 0.009). Proteins fall between the two. Surprisingly, although the agent discovers distinct architectures per domain (p = 0.004), every innovation transfers across all three domains with <1% degradation, indicating that the differences reflect search-path dependence rather than fundamental biological requirements. We release a decision framework and open-source toolkit for molecular modeling teams to choose between autonomous architecture search and simple hyperparameter tuning.

Abstract

Molecule Design

Explore similar work

What an Autonomous Agent Discovers About Molecular Transformer Design: Does It Transfer?

General Multimodal Protein Design Enables DNA-Encoding of Chemistry

The Montparnasse Algorithm for RNA Design