acceptodds
Under review as a conference paper at ICLR 2027

SpecOpt: Contact-Diff Reasoning for Agentic Molecule Optimization Toward Binding Specificity

Abstract

Off-target protein binding is a major source of adverse effects for small-molecule drugs, yet most structure-based molecular design methods focus on generating selective compounds de novo rather than improving the selectivity of existing, well-characterized drugs. We introduce specificity optimization (SpecOpt), a molecular design task that seeks constrained structural modifications to an existing compound that increase its binding preference for an intended target over known off-targets while preserving its structural identity and drug-like properties. To enable systematic evaluation of this task, we construct a ChEMBL-derived benchmark from compound-target interaction data, identifying intended targets through curated drug-mechanism annotations and off-targets through measured activities. We then develop an agentic framework for SpecOpt that docks each compound against its intended target and off-targets, compares the resulting poses through residue-aware atom-protein contacts, and provides these differential interactions to a large language model to propose targeted structural modifications. Candidates are retained only if they satisfy molecular similarity, ADMET, and target-off-target docking selectivity criteria. On 915 compounds, the agent improves the target-off-target binding gap for 84.8% of compounds, shifting the mean gap from -0.72 to +0.47 kcal/mol while maintaining a mean Tanimoto similarity of 0.72 to the starting compounds. Ablation studies identify residue-specific contact information as the critical optimization signal: replacing residue identities with binary contact indicators eliminates improvement on all 29/29 ablation compounds. These results establish SpecOpt as a distinct molecular design problem and demonstrate residue-aware differential interactions as an effective signal for improving the specificity of existing compounds.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Drug specificity, LLM agents, Structure-based drug design

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CDeYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces specificity optimization, a task that aims to improve the binding preference of existing molecules toward intended targets over known off-targets while preserving their structural similarity and drug-like properties. To support this task, the authors construct a ChEMBL-derived benchmark of 1,848 compounds with annotated targets and off-targets. They then propose SpecOpt, an LLM-based framework that uses residue-aware contact differences to guide molecular edits, followed by similarity, ADMET, and docking-based filtering. Experiments on 915 compounds show improvements in docking-predicted specificity gaps for 84.8% of the compounds.

Critical strengths
  1. The paper formulates a practically meaningful task of improving the binding specificity of existing molecules and proposes a dedicated agentic framework to address it.
  2. The residue-aware contact difference is an intuitive and interpretable signal for guiding molecular edits, and the overall framework is clearly presented.
  3. The authors construct a ChEMBL-derived benchmark and demonstrate promising results on 915 compounds, improving the docking-predicted specificity gap for 84.8% of them.
  4. The paper provides useful case studies, ablation experiments, and implementation details that support reproducibility.
Critical weaknesses
  1. Missing AI use statement. The paper does not include the mandatory AI use statement required by ICLR 2027. This should be flagged for consideration of a potential desk rejection.
  2. Reliance on Vina for both optimization and evaluation. The method uses Vina-predicted poses to guide molecular edits and Vina scores to select and evaluate candidates, raising concerns about overfitting to docking artifacts. Independent validation using additional scoring functions (e.g., GNINA) and pose-validation tools (e.g., PoseBusters) would strengthen the reported improvements.
  3. Insufficient baseline comparisons. The paper lacks direct comparisons with existing or simpler docking-guided optimization methods under matched conditions, such as fragment enumeration or LLM editing without contact guidance. It is therefore unclear how much improvement comes specifically from the proposed reasoning mechanism.
  4. Limitations of the key ablation. Removing residue identities causes the system to terminate without making edits, leading to zero improvement on all 29 compounds. Since the ablated system does not perform a comparable search, this result alone does not fully establish the benefit of residue-aware contact guidance.
Questions for authors

My main questions are covered in the weaknesses above.

Reviewer_JpVjOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents SpecOpt, an agentic framework for improving the binding specificity of existing compounds. The method compares molecular docking interactions between intended targets and off-target proteins, uses residue-aware contact differences to guide LLM-based structural modifications, and filters candidates based on predicted binding selectivity, structural similarity, and ADMET properties. Experiments on 915 compounds report improvements in docking-based specificity scores.

Critical strengths

The paper addresses a relevant problem in molecular optimization and attempts to incorporate protein–ligand interaction information into LLM-guided molecular editing. The experiments cover a relatively large number of compounds and include several ablation studies.

Critical weaknesses
  1. Limited novelty and insufficient methodological contributions.

The overall contribution appears incremental. Improving the properties of existing compounds is already a well-established objective in drug discovery and lead optimization. Numerous existing methods perform structure-guided molecular modification rather than de novo generation. The paper does not sufficiently distinguish SpecOpt from these approaches or demonstrate substantial methodological innovation beyond combining docking analysis, LLM-based editing, and property filtering.

  1. The research motivation and problem formulation are not sufficiently convincing.

The abstract emphasizes that off-target protein binding is a major source of adverse drug effects and argues that existing methods primarily focus on de novo molecular generation. However, off-target binding does not necessarily lead to adverse effects, and the relationship between binding selectivity and drug safety requires more careful justification. Furthermore, improving the selectivity of existing compounds is not a fundamentally new drug design problem. The authors should better clarify the motivation and explain what distinguishes the proposed task from conventional lead optimization.

  1. Insufficient evidence supporting the effectiveness of the proposed method.

The main improvements are evaluated using docking scores rather than experimentally measured binding affinities or selectivity. It remains unclear whether these improvements translate into meaningful biological outcomes. Moreover, the paper lacks sufficiently competitive comparisons with conventional molecular optimization methods under matched evaluation budgets, making it difficult to establish the actual benefits of the proposed agentic framework.

  1. The presentation and organization need substantial improvement.

The manuscript is difficult to follow in several places. Frequent use of emphasized or italicized text, along with excessive parenthetical explanations and numerical details, disrupts the flow of the discussion. The authors should improve the organization, simplify the presentation, and more clearly communicate the methodological contributions and experimental findings.

Questions for authors

No, see weak

Reviewer_ye9kOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new and practical task of binding specificity optimization. The task starts with an existing compound and tries to maximize the gap between the binding affinities (estimates of affinities) of the designed compound with intended-target and that with off-targets. The authors developed an agentic framework to make it work with agents and show improved binding selectivity for compounds on average. The work also identifies target residue identity as one important feature for this optimization task by ablation.

Critical strengths
  1. This is a meaningfully important task to be proposed and a sensible and rational way to organize molecular editing is also present.
  2. Reproducible improvements and candid reporting of negative findings. This is an important and in some perspective critical part of a research that the community should value.
Critical weaknesses
  1. From the released trace, the simplified variant in ablation generates zero candidate proposals on the same 29 compounds for which the full system records 620 proposals. It exits before proposed edits by looking at the code. So the LLM does not make any attempts across those round since it think there's no modifications or whatnot. This needs further explanation or further pushing to the agent to make it at least try to propose modifications.
  2. Still for the proposed system: a simplified optimization methods could also improve the performance, but the proposed system's edge over a simple one is not present.
Questions for authors

Have you considered other metrics than Vina?

2 comments

I have read the authors’ response and appreciate the efforts that went into improving the paper. As the manuscript stands, I do think maintaining the score is more proper than not.

We sincerely thank the reviewer for taking the time to read our rebuttal and consider our additional experiments and clarifications. We respect your decision to maintain your rating and greatly appreciate your constructive feedback, which has helped us strengthen the manuscript.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.