acceptodds
Under review as a conference paper at ICLR 2027

AUTOMATICALLY BUILDING MOLECULAR PROPERTY PREDICTION MODELS WITH AN EXPERT-GUIDED AI AGENT

Abstract

Molecular property prediction guides compound prioritisation in drug discovery, but turning domain expertise into an effective model takes machine-learning engineering and repeated cycles of design, implementation, training and evaluation. We present a hierarchical agent pipeline that automates this workflow. Given a task description, the data and optional written expert guidance, it runs an agent loop that designs a model, writes and debugs the code, trains and tunes it, evaluates the result and, when the results call for it, revises the design, then selects and delivers a final model, so a domain expert steers development in prose rather than code. A controlling language-model agent assigns these steps to specialised sub-agents and decides each next step from their reports and training results; the pipeline code is the same for every task. We evaluate it retrospectively on the OpenADMET pregnane X receptor (PXR) induction challenge, a recent blind competition, using guidance we wrote from published challenge reports and earlier runs of this pipeline after the competition ended. Across three runs, the guided pipeline achieves a mean relative absolute error of 0.561, compared with 0.627 without guidance; the published winning score of 0.563, obtained using proprietary data, provides an external reference. On two further benchmarks, hepatocyte clearance and plasma protein binding, the same pipeline with task-specific guidance scores better on average than the published leaderboard bests. Given guidance that names no model, the pipeline's run fell within the range of three AIDE runs and scored better than three MLEvolve runs.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
AI for Science, LLM agents, automated machine learning, expert guidance, scientific prediction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_WvtHOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper builds a molecular property prediction pipeline on Claude Code’s agent SDK. A controller delegates setup, design, implementation, training, and selection to specialised agent sessions supplied with expert guidance. Experiments on PXR induction, hepatocyte clearance, and plasma protein binding evaluate guidance effects and compare performance with other coding agents.

Critical strengths
  1. The paper presents a functional pipeline covering environment setup, model design, implementation, training, selection, and prediction delivery, demonstrated on several molecular property prediction tasks.
  2. The evaluation includes comparisons with a single coding agent, AIDE, and MLEvolve, alongside predictive-model ablations and a control for additional training data. The authors openly report experimental deviations and limitations.
Critical weaknesses
  1. The methodological novelty is limited. The framework combines written expert guidance with a controller coordinating role-specific coding agents. Although the authors explain the design rationale, they do not clearly identify a new technical mechanism beyond existing agent orchestration practices. The contribution appears primarily to be an application of these practices to molecular property prediction.
  2. The effectiveness of the proposed framework is not convincingly demonstrated. The evaluation does not establish a clear benefit over existing coding agents supplied with the same guidance, or reliably isolate the contribution of expert guidance. The authors need stronger evidence that the proposed framework improves accuracy, reliability, or efficiency beyond executing a prescribed modelling solution.
Questions for authors
  1. Could the authors demonstrate what the proposed hierarchy adds over a single coding agent given the same guidance, data, and tools?
  2. Could the authors repeat the agent comparisons with consistent run counts and matched resource limits to establish whether the reported differences are reliable?
  3. Could the authors isolate the effect of expert guidance under otherwise identical conditions, using guidance fixed before access to evaluation outcomes and an untouched test set?
Reviewer_4DhaOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a hierarchical language-model agent system for automating molecular property prediction.

Critical strengths

The system addresses a practical modeling problem and demonstrates a workflow across three molecular prediction tasks.

Critical weaknesses
  1. Writing. This paper does not look like a scientific article. Particularly Section 3, where no concept is formally defined. Far from an ICLR-standard paper. Take one sentence as an example: "The task description defines the task, namely the data and their columns, the target, the metric, the fold assignment and the deliverable, and adds working rules such as fixing the numbers of epochs and folds before training." What are "task", "data", "column", "target", "metric", "fold assignment", "working rules"?

  2. Methodology. I don't see any contributions or novelty.

  3. Experiments. Incomplete. No baselines, no competing methods. Unclear experimental settings. No results to support the design of the "hierarchical agent pipeline".

Questions for authors

No questions. This should not be an ICLR submission.

Reviewer_AYcSOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents an LLM agent framework for automated molecular property prediction. A controlling agent dispatches four specialized sub-agents (setup, design, implementation, selection), and the framework code is task-agnostic, with all task-specific knowledge supplied as natural language guidance. The authors evaluate their agent on the OpenADMET PXR blind challenge and two TDC ADMET regression benchmarks, ablating the guidance document, and compare against the leaderboard solution, a coding agent, AIDE, and MLEvolve.

Critical strengths
  1. The paper is straightforward. It discloses that the reported PXR batch replaced an earlier one after both were scored, that the two arms used different fold files and ran eleven days apart, that one unguided run retrieved a summary of the winning method, and that a clearance run failed an isolation check and was kept. All 21 runs are listed with no outlier rule. Most of my concerns below come from the authors’ own text, which is rare and worth crediting.
  2. The single coding agent comparison, whatever it is, is the most useful experiment in the paper, since it matches the full system closely and the authors report it rather than hide it.
Critical weaknesses
  1. The writing is loose and the paper is hard to follow. The system is called a hierarchical pipeline when what is described is a fairly standard multi-agent setup with one dispatcher and four workers. Calling it hierarchical sets an expectation of depth or recursion that the method does not deliver. The single coding agent, which carries a lot of weight in the results, is never properly defined. I could not tell from the text what it is, how it is prompted, or why it is the right baseline. The conclusion opens the discussion with a sentence like “the evidence is small,” which is not a statement a reader can do anything with. More broadly the paper reads as a sequence of facts without an argument connecting them: it is often unclear which claim a given paragraph is supporting.
  2. The proposed pipeline seems very straightforward but I cannot see what is special about it. Why not just write a skill that calls a coding agent to do this? Many coding agents now have subagent functionality that can do exactly this. What is the key contribution of this paper?
  3. The experimental setup is confusing. Why is the pipeline run three times while the single coding agent is run only once? The evaluation is also limited, with only three task datasets. The Prescription column in Table 2 is confusing that I can not understand what these mean here.
  4. The performance gain seems very marginal.
Questions for authors
  1. What exactly is the single coding agent? Please specify its prompt, its stopping rule, and why one run is sufficient when the pipeline gets three.
  2. What does the orchestration provide that a subagent-capable coding agent with a well-written skill file would not? If the answer is adaptive redesign, please show a run where it fired and changed the outcome.
  3. Can you rule out that generated code fetched the labeled TDC test splits at runtime? They are public, the design prompt encourages web search on named benchmarks, and step logs do not capture requests issued inside running programs.
  4. Could you run one guided PXR batch on the same fold file as the unguided arm? That is the single largest confound in the main result.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.