PaperlayerView exact revision on arXiv

Source-linked AI summary

DrugSAGE:Self-evolving Agent Experience for Efficient State-of-the-Art Drug Discovery

Yikun Zhang, Xiwei Cheng, Tianyu Liu, Yuanqi Du, Wengong Jin

arXiv:2605.15461v1cs.LGcs.AI

TL;DR

Drug-discovery model development requires costly searches, and current agents typically discard what they learn before the next task. DRUGSAGE reuses verified cross-task experience to guide model search or transfer a solution directly, with strong reported results in single-task and cross-task evaluations.

  • Problem

    Building state-of-the-art drug-discovery models requires costly searches, while current agents repeat that cost on each task because they do not retain prior search experience.

    Source: §1 · p. 1

  • Method

    DRUGSAGE accumulates cross-task memory of verified skills, effective strategies, and recurring errors with fixes to guide search or transfer working solutions.

    Source: §1 · p. 2

  • Results

    With memory from 16 smaller tasks, DRUGSAGE scores 0.935 on 17 held-out tasks and DRUGSAGE-ZERO outperforms all baseline agents by more than 10-30% without test-time search.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors report that DRUGSAGE improves algorithm performance over baselines for in-distribution and out-of-distribution drug-discovery problems.

    Source: §5 · p. 9

    Biomni results were provided by the Biomni team from its beta platform and were not independently reproduced.

    Source: §4.2 · p. 7

Abstract

from arXiv · show

Building state-of-the-art (SOTA) predictive models for drug discovery requires expensive search over tools, architectures, and training strategies. Current LLM-based agents can find SOTA solutions through extensive trial and error, but they do not retain the experience accumulated along the way and therefore pay the full search cost on every new task. We propose \method (Self-evolving Agent Experience), a framework that accumulates and reuses experience across tasks to build SOTA drug discovery models efficiently. \method maintains a cross-task memory of verified skills, statistical evidence about effective strategies, and a record of recurring errors and their fixes. In some cases, \method transfers a working solution directly without test-time search. In 33 molecular property prediction tasks, \method ranks first among nine SOTA agents in a single-task setting. With memory accumulated from 16 smaller tasks, \method achieves an averaged normalized score of 0.935 on 17 held-out tasks in a cross-task evaluation setting and outperforms all baseline agents by 10-30\% in a zero-test-time search regime. In summary, our work shows the advantage of cross-task memory for efficient SOTA model development in drug discovery.

1 Introduction

Building state-of-the-art drug-discovery models requires costly searches that current agents repeat for every task. DRUGSAGE accumulates cross-task experience to guide that search or transfer a verified solution directly, and the paper reports strong results across single-task and cross-task evaluations.

1 Introduction

  • SOTA model development is costly because agents must search tools, adapt code, tune configurations, and repeat experiments, while most discard that experience before each new task.Large datasets, expensive foundation-model fine-tuning, and runs lasting hours or days further increase costs.Source: §1 · p. 1
  • DRUGSAGE stores verified skills, evidence about effective strategies, and recurring failures with fixes so experience from one task can guide later tasks.It can prioritize promising approaches or transfer a working solution without additional search.Source: §1 · p. 2
  • With memory from 16 smaller tasks, DRUGSAGE scores 0.935 on 17 held-out tasks and DRUGSAGE-ZERO outperforms all baselines by more than 10-30% without test-time search.In the single-task setting, DRUGSAGE ranks first among eight baseline agents.Source: §1 · p. 2
  • DRUGSAGE integrates cross-task memory into an MCTS-based search loop, with a formal guarantee that its selection policy preserves standard UCB's regret bound.The memory-augmented policy is described as preserving the regret bound of standard UCB.Source: §1 · p. 2
  • Unlike systems that refine user-provided or leaderboard models, DRUGSAGE builds its skill library by searching literature and GitHub repositories.This broadens the search space and removes the need for a human-curated starting point.Source: §1 · p. 2

2 Related Work

Prior work develops agents for algorithm search, scientific workflows, and persistent memory, but DRUGSAGE emphasizes structured experience accumulated across tasks. It pairs an expanding executable skill library with performance-grounded memory that can guide later searches or enable direct reuse.

Automated algorithm discovery agents

  • Unlike agents that search algorithms or improve them through agent structures, search methods, or external knowledge, DRUGSAGE accumulates structured cross-task experience.Later tasks can retrieve verified solutions at zero search cost or warm-start from empirically validated starting points.Source: §2 · p. 2

Agent experience

  • DRUGSAGE combines an executable skill library that expands across tasks with performance-grounded memory that deepens through execution.The combination supports broader search coverage and more targeted reuse over time.Source: §2 · p. 2

Scientific discovery agents

  • DRUGSAGE frames biomedical modeling as cross-task search, reusing empirically validated strategies to seed experience-conditioned MCTS and avoid redundant exploration.This contrasts with Agentomics, which explores modeling strategies for a given biomedical dataset.Source: §2 · p. 3
  • DRUGSAGE differs from scientific workflow agents such as Coscientist and ChemCrow by focusing on reusable modeling strategies across tasks.Earlier work also includes computational workflow agents [3] and SAGA [35], which evolves workflow objectives.Source: §2 · p. 3

3 Methodology

DRUGSAGE constructs executable skills, records experience across tasks, and uses that memory to guide budgeted model search or route a verified solution when no new search is allowed.

3 Methodology

  • DRUGSAGE targets the best executable solution within budget B, first constructing a skill library and maintaining cross-task memory before choosing search or direct routing.For B > 0 it runs memory-enhanced MCTS; for B = 0 it transfers a verified solution without additional search.Source: §3 · p. 3

3.1 From Literature to Executable Skills

  • The Explore Agent expands the skill library from literature and GitHub methods, grounds proposed skills in repository APIs, and validates them in sandboxes before inclusion.Validation ranges from syntax and import checks to end-to-end execution, with automatic repair on failure.Source: §3.1 · p. 4

3.3 Experience-Memory-Enhanced MCTS

  • Memory-enhanced MCTS searches a forest of executable solutions, using historical family performance for UCB selection and memory-grounded edits and execution repair; its regret remains O(|F| log T).R guides family selection, H grounds child proposals, and Q supports sandbox repair and resource-aware execution; execution outcomes update persistent memory.Source: §3.3 · pp. 4–6

3.2 Experience Memory

  • Experience memory Z = (R, H, Q) stores solution performance, edit rationales and effects, and execution failures, fixes, and resource profiles for reuse across tasks.After execution, outcomes and traces are written back to memory as reusable cross-task evidence.Source: Figure 2 · p. 5

3.4 DRUGSAGE-ZERO: Zero-Test-Time Search by Memory Routing

  • With B = 0, DRUGSAGE-ZERO ranks analogous tasks by task type, metric, training-set size, and description similarity, then deploys the best solution from the closest analog.The task-description embedding contributes to the similarity comparison.Source: §3.4 · p. 6

4 Experiments

The experiments test DRUGSAGE both as a from-scratch model-search agent and as a cross-task system that reuses experience to reduce or bypass search. Across these settings, the reported results show strong performance on single-task benchmarks and held-out tasks, with ablations examining the skill library, memory, and analog-task effects.

4 Experiments

  • The experiments compare from-scratch search with cross-task reuse, including direct transfer without test-time search on held-out tasks.The cross-task experience pool comprises the 16 smallest TDC tasks, with six larger TDC tasks and 11 Polaris tasks held out for evaluation.Source: §4 · p. 6, §4.2 · pp. 7–8

4.1 Evaluation of DRUGSAGE in the single-task setting

  • Full DRUGSAGE achieved the best average rank (1.95) and most wins (10/22) across 22 from-scratch TDC tasks.DRUGSAGE-anchor also outperformed Autoresearch and ShinkaEvolve under a matched starting-point setting.Source: §4.1 · p. 7, Table 1 · p. 7

4.2 Evaluation of DRUGSAGE in the cross-task setting

  • 0.90 on six held-out TDC-ADMET tasks and 0.97 on 11 held-out Polaris tasks: DRUGSAGE-ZERO beat the strongest baselines, including those using their full search budgets.The reported gain over baselines was more than 10–30% in both benchmarks.Source: Figure 3 · p. 8, §4.2 · p. 8

4.3 Ablation Study

  • DRUGSAGE outperformed DRUGSAGE-anchor, which disables the Explore Agent, with normalized scores of 0.929 versus 0.875.The authors attribute the difference to the skill library automatically constructed by the Explore Agent.Source: §4.3 · p. 8
  • From without Z to with Q to full memory, normalized score increased while LLM cost decreased; DRUGSAGE-ZERO had near-zero task-level LLM API cost.DRUGSAGE-ZERO uses no LLM generation at test time and calls only a lightweight text-embedding model for memory routing.Source: §4.3 · p. 9
  • After same-category experience tasks were removed, DRUGSAGE-ZERO remained close to leaderboard performance on three CYP targets, with reductions of 0.03–0.05.The authors conclude that performance on these targets is not merely copying exact same-category analogs.Source: §4.3 · p. 9

5 Discussion

DRUGSAGE combines agent-driven exploration, cross-task memory, and automatic experiment refinement, and the authors report improved performance over baselines on in- and out-of-distribution problems. They identify skill prioritization and evaluation beyond drug discovery as directions for further work.

5 Discussion

  • DRUGSAGE combines agent-driven exploration, cross-task memory, and automatic experiment refinement for drug discovery.The authors report improved algorithm performance compared with baselines on both in-distribution and out-of-distribution problems.Source: §5 · p. 9
  • Skill selection based on both prior signals and dataset experiments is proposed as a direction for further investigation.The current framework prioritizes skills using prior knowledge such as citation and GitHub numbers.Source: §5 · p. 9
  • The framework is tested mainly on drug discovery tasks, while possible generalization to other scientific research areas remains future work.Source: §5 · p. 9

A Benchmark Dataset Summary

The benchmark dataset summary covers 22 Therapeutics Data Commons datasets and 11 Polaris Hub held-out test datasets. The table identifies each dataset's description, size, and metric.

A Benchmark Dataset Summary

B Dataset Partition

The appendix partitions datasets by training-set size and documents evaluation protocols, prompts, baseline approaches, and budget mapping. These details specify how the benchmark comparisons were run.

B Dataset Partition

  • The target set contains the six largest datasets with at least 5,000 training samples, while the remaining 16 datasets form the experience pool.The next-smallest dataset is Lipophilicity at 4,200 training samples; Table 4 partitions 33 datasets into 16 experience-pool and 17 target tasks, including 11 Polaris datasets.Source: B Dataset Partition · p. 16, Table 4 · p. 16

C Per-Baseline Protocols and Prompts

  • Except for Biomni, baselines use five-seed train-validation splits and the fixed prepare.py evaluation pipeline; the shared prompt specifies task, metric, data sizes, API, output format, and constraints.For Polaris, the prompt uses prepare_polaris.py and the Polaris-specific metric.Source: C Per-Baseline Protocols and Prompts · p. 17

C.1 Autoresearch

  • Autoresearch starts from the top three open-sourced TDC leaderboard models or, for Polaris, Chemelon.Source: §C.1 · p. 17

C.2 ShinkaEvolve

  • ShinkaEvolve evolves candidate programs by sampling archive parents and mutating them with LLM-generated code; prompts provide parent and archive code and performance metrics.It uses diff patches 70% of the time and full rewrites 30% of the time.Source: §C.2 · p. 18

C.3 MLEvolve

  • MLEvolve uses tree search with draft-debug-improve cycles, selecting promising nodes and executing proposed code improvements.Source: §C.3 · p. 18

C.4 AIRA-Dojo

  • AIRA-Dojo applies MCTS to draft, improve, debug, and analyze operators, while varying draft complexity across rounds.Source: §C.4 · p. 19

C.5 Claude Code; C.6 STELLA

  • Claude Code uses the Autoresearch system prompt except for its anchor methods, while STELLA runs in its default self-evolve mode with the shared task prompt.Source: §C.5 · p. 19, §C.6 · p. 19

C.7 Agentomics; C.8 Per-Baseline Budget Mapping

  • Agentomics follows a fixed multi-step sequence from iteration planning and data exploration through model training, inference, and validation evaluation.The budget mapping converts each baseline's native iteration unit into B, where one unit represents a complete target-task train-and-evaluate cycle, making comparisons approximately compute-equalized.Source: §C.7 · pp. 20–21, §C.8 · p. 21, Table 5 · p. 21

D Per-Task Scores for the TDC-ADMET Benchmark

Table 6 presents per-task scores on the 22 TDC-ADMET benchmark datasets. Most systems were evaluated under a unified protocol, while Biomni scores come from its beta platform without reported standard deviations.

D Per-Task Scores for the TDC-ADMET Benchmark

E Per-Task Scores on the Polaris Benchmark

On the 11 Polaris hold-out tasks, DRUGSAGE leads budgeted-search agents on most tasks, while DRUGSAGE-Zero remains competitive without experiment budget.

E Per-Task Scores on the Polaris Benchmark

  • DRUGSAGE ranks first on eight of eleven tasks among agents with budgeted search (B = 20); DRUGSAGE-Zero ranks first on one and second on six.DRUGSAGE-Zero achieves the top score on adme-fang-hclint.Source: E Per-Task Scores on the Polaris… · p. 23

F Per-Task Budget Trajectories on the 17 Held-Out Tasks

The section presents per-task budget trajectories across 17 held-out tasks and describes the zero-test-time routing results reported for those tasks.

F Per-Task Budget Trajectories on the 17 Held-Out Tasks

G Per-Task Zero-Test-Time Search Results

G.1 TDC-ADMET Benchmark Held-Out Tasks

The appendix details zero-shot task routing, skill construction, and cross-task transfer scoring, then states boundedness and regret guarantees for memory-augmented selection.

G.1 TDC-ADMET Benchmark Held-Out Tasks

  • Across three ADMET routing patterns, the selected historical task’s best solution transfers without modification, including matches based on shared metrics despite differing biological contexts.CYP inhibition tasks match to CYP2C9_Substrate based on AUPRC and CYP domain alignment despite the substrate-versus-inhibition task-type difference.Source: §G.1 · p. 23

G.2 Polaris Benchmark Held-Out Tasks

  • For Polaris, no historical task has Pearson or MSE for the nine regression tasks, while both AUPRC kinase tasks match directly to CYP substrate tasks.The AUPRC matches rely on shared metric and task type despite a kinase-versus-enzyme domain gap.Source: §G.2 · p. 25

G.3 Zero-Test-Time Routing Case Study

  • For Solubility, routing transfers a wide-descriptor 10-member homogeneous ensemble, while scratch search finds a heterogeneous ensemble with MAE = 0.722.The authors state that the transferred Lipophilicity solution is more effective than the solution found by searching Solubility from scratch.Source: §G.3 · p. 25

H Explore Agent Workflow Details

  • The offline Explore Agent turns task-relevant literature and repositories into skills, then validates and repairs each skill through three escalating checks before library admission.LiteratureScout filters and ranks candidates; SkillBuilder uses repository AST snapshots to constrain executable integration instructions.Source: H Explore Agent Workflow Details · pp. 25–26

I.1 Per-Task Standardized Score

  • Historical scores are oriented higher-is-better, robustly normalized around each task’s median and MAD, then sigmoid-mapped to [−1, 1].A small ε prevents division by zero, and median/MAD normalization is described as robust to outlier solutions.Source: §I.1 · pp. 26–27

I.2 Task-Similarity Weights

  • Historical-task relevance is weighted by metric match, task-type match, dataset-size similarity, and task-description embedding similarity.The factors are combined multiplicatively.Source: §I.2 · p. 27

I.3 Formal Definition of Transfer Scores

  • Family-level transfer aggregates standardized scores for solutions in a model family, while node-level transfer matches solutions by model family and modification type.Both transfer terms are weighted averages over historical tasks.Source: §I.3 · p. 27

I.4 Boundedness Guarantee; J.1 Problem Setup; J.2.2 Lemma: Causal Bonus Is Order-Compatible with UCB Exploration

  • Bounded transfer scores preserve the standard UCB regret order under the stated assumptions; the causal bonus changes only the leading constant.The order-compatibility lemma likewise states that the causal bonus introduces no new asymptotic order.Source: §I.4 · p. 28, §J.1 · p. 28, §J.1.1 · p. 28

J.2.3 Theorem: Family-Level Regret Bound

The family-level regret bound retains standard UCB’s asymptotic order, with the causal factor affecting the leading constant rather than the order.

J.2.3 Theorem: Family-Level Regret Bound

  • The family-level regret bound is O(|F| ln T), identical in asymptotic order to standard UCB, with the causal factor changing only the leading constant.The bound follows under the stated assumptions.Source: §J.2.3 · pp. 29–30
  • The proof follows standard UCB analysis while tracking the causal error term and bounding the number of pulls of suboptimal families.It proceeds through high-probability bounds and a cumulative-regret argument.Source: §J.2.3 · pp. 29–30
  • The node-level bonus leaves the original O(T ln T) asymptotic regret order unchanged and may improve the leading constant.It increases sampling probability only for nodes with positive CEG signal.Source: §J.2.3 · p. 30

Related papers

Loading 2605.15461v1…