acceptodds
Under review as a conference paper at ICLR 2027

Prediction-Reasoning Fusion for Molecular Hit Discovery

Abstract

Molecular hit discovery requires turning imperfect activity predictions into useful assay choices. We study prediction-reasoning fusion, in which a predictive model estimates molecular activity, a language model proposes interpretable acquisition policies, and a policy evaluator measures their utility on development data. The language model requests numerical summaries relating predictions and molecular features to observed outcomes, proposes decision trees and uses evaluation feedback to guide subsequent proposals. A tree guides the next assay batch only when it improves development utility over model-based ranking. Before each subsequent acquisition round, the predictive model is refitted using the initial and acquired labels, and a new tree-development session begins. This automated, agentic procedure produces decision rules that scientists can examine alongside their development support. A theoretical bound relates the transfer of development gains to adaptive data reuse, differences between candidate populations and the extent of policy changes. Experiments on HIV and ALDH1 show higher mean acquisition hits and early discovery than ranking by a multilayer perceptron ensemble and random-tree search, although higher development scores do not always lead to better acquisition outcomes. A supplementary exploratory held-out assessment finds significant hit gains on HIV.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular hit discovery, agentic AI, large language models, sequential acquisition, interpretable decision trees, biochemical knowledge, automated policy development

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_i4pcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a prediction-reasoning fusion framework for molecular hit discovery, combining an MLP ensemble for activity prediction with LLM-generated decision trees for compound selection. The LLM iteratively develops acquisition policies using molecular features, numerical summaries, and evaluation feedback, while the predictive model is retrained after each acquisition round. The authors also provide a theoretical analysis of how improvements in development utility may transfer to actual discovery performance. Experiments on HIV and ALDH1 benchmarks report improvements in hit discovery and early acquisition efficiency compared with MLP-based ranking and random-tree search, although development gains do not consistently translate into better acquisition outcomes.

Critical strengths

The paper presents an interesting integration of predictive modelling and LLM-guided policy optimisation for molecular hit discovery. A notable strength is the use of interpretable decision trees, which allows scientists to examine and understand compound selection strategies. The iterative evaluation and feedback mechanism provides a systematic approach to refining acquisition policies, while the theoretical analysis offers insight into the relationship between development utility and actual discovery performance. The comparisons with MLP ranking and random-tree search further support the potential effectiveness of the proposed framework.

Critical weaknesses

The experimental evaluation is limited to two molecular screening benchmarks, which restricts the evidence for generalisability across different chemical and biological settings. Although the proposed approach improves average hit discovery performance, higher development utility does not consistently translate into better acquisition outcomes, raising concerns about policy stability and reliability. Furthermore, the iterative LLM-guided search introduces additional computational complexity, but its efficiency and scalability are not sufficiently established. Broader experimental validation and a more detailed analysis of the contribution of LLM-guided reasoning would strengthen the paper.

Questions for authors
  1. How consistently do improvements in development utility translate into better acquisition outcomes, and what factors contribute to cases where this relationship does not hold?

  2. How much of the performance improvement can be attributed specifically to LLM-guided reasoning, rather than the decision-tree representation and iterative policy evaluation?

  3. How well does the proposed framework scale to larger molecular libraries and more diverse screening tasks, particularly considering the computational cost of repeated LLM-guided policy searches?

Reviewer_o4wYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper examines the integration of prediction and reasoning for molecular hit discovery. An ensemble of MLPs scores candidate molecules, a LLM proposes decision trees that assign ordinal acquisition priorities based on these scores and an evaluator assesses the utility of each tree on a separate held-out development partition. Experiments on HIV and ALDH1 compare MLP-based ranking, random tree search and two LLM workflows across ten paired campaign initializations.

Critical strengths
  1. Within each replicate, all methods start from the same labeled set. The campaign, not individual rounds or molecules, is the unit of analysis. The null hypothesis for the sign flip test is stated explicitly. In Figure 1, gray lines connect results from the same initialization across methods. This directly visualizes the paired design and shows how much of the variation comes from initialization rather than from the method itself.

  2. The baseline is strong, and the paper quantifies it clearly. MLP ranking reaches a 62.7% hit rate on HIV and 36.0% on ALDH1, compared with base rates of 3.51% and 4.99%. The authors also note that these numbers are not directly comparable to published benchmark scores under other partitions.

Critical weaknesses

The result is a clean application of Russo & Zou to relative acquisition utility, and the proof is sound. My concern is what it contributes to the paper. By the authors’ own account in A.3, n corresponds to no quantity in the experiments; I and δ are not estimated; ρ=1 for their priority trees, since a priority change can reorder the entire pool and the bound does not cover the sequential campaign, because later predictive models depend on earlier development-guided acquisitions. With ρ=1 and no estimates of I or δ, inequality (1) reduces to the claim that development gain overstates acquisition gain by an unquantified amount. This is the motivation for the experimental design, not a consequence of the theorem.

A constructive path is available: constrain a deployed tree to replace at most ρ⋅b of the MLP’s top-b selections, sweep ρ∈{0.25,0.5,1} and report acquisition utility, since the bound implies that limiting ρ trades attainable gain for tighter transfer. If such a replacement cap improves transfer, the theory earns its place; otherwise, Section 3.3 and Appendix A could be removed without changing a single empirical conclusion.

Questions for authors
  1. How often is a tree actually deployed rather than falling back to MLP ranking, by method and dataset? What is the acquisition outcome conditional on deployment versus fallback?

  2. What happens if random-tree search is given 100 or 1,000 candidates per checkpoint rather than four? Alternatively, how does a greedy or exhaustive optimizer over the same depth-≤3 representation compare?

  3. On HIV, Astra outperforms Qwen on the development and held-out sets but underperforms on acquisition. How should this reversal be reconciled with the overfitting interpretation offered for the earlier one?

  4. Would a second held-out partition be feasible, so that the HIV result does not rest on a single molecular split?

  5. When was the decision to reduce from twenty to ten replicates made relative to inspecting the outcomes, and were the remaining seeds ever run?

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

This paper proposes Prediction-Reasoning Fusion, a framework that combines MLP-based molecular activity prediction with LLM-proposed decision-tree acquisition policies. The framework formulates molecular hit discovery as a sequential acquisition process, where the LLM iteratively generates and refines decision trees using feedback from the development set at each acquisition round. Experiments on two molecular screening datasets, HIV and ALDH1, show higher average hit discovery than MLP ranking and random-tree search.

Critical strengths
  1. The paper addresses the problem of improving molecular hit discovery beyond ranking molecules based on predicted activity. The proposed framework provides a useful way to combine molecular activity prediction with LLM-guided acquisition policy optimization.

  2. The use of LLM-generated decision trees provides interpretable acquisition policies, allowing the explanation of the decision rules used to prioritize molecules for experimental testing.

Critical weaknesses
  1. In the Experiments section, the authors compare the proposed method only with MLP ranking and random-tree search. It is unclear whether LLM-generated decision trees provide advantages over more advanced machine learning methods. Additionally, a direct acquisition decision proposed by LLM without decision-tree generation would be a useful baseline to justify the proposed policy representation.

  2. In the sequential acquisition process, feedback from previous rounds is only incorporated through data updates. The LLM restarts policy proposal without reusing previous decision trees or evaluation feedback. Comparisons with policy search initialized from previous decision trees would help assess whether reusing experience could improve performance.

  3. The proposed framework consists of evidence querying, iterative decision-tree generation, and evaluation feedback. However, the experiments do not isolate the contributions of these components. Ablation studies comparing one-shot tree generation with iterative refinement, etc., would be helpful.

  4. The presentation could be clearer. First, the paper lacks a clear problem definition and formulation in the Framework section. Second, several notations are not clearly defined. Third, the roles of different sets and the overall sequential acquisition workflow are not clearly explained.

  5. The figures could also be clearer, the meanings of different colors and visual elements are not explicitly illustrated.

Questions for authors

Please see the Weaknesses section for detailed questions and concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.