Probe Before You Edit: Probing-Guided Molecular Optimization for LLM Agents in Structure-Based Drug Design

# Probing-Guided Molecular Optimization for LLM Agents in Structure-Based Drug Design
Source: [https://arxiv.org/html/2606.00555](https://arxiv.org/html/2606.00555)
Zaifei Yang1, Weiyu Chen2, Yaqing Wang3, James Kwok1 1The Hong Kong University of Science and Technology 2City University of Hong Kong 3Beijing Institute of Mathematical Sciences and Applications zyangea@connect\.ust\.hk

###### Abstract

Structure\-based drug design increasingly employs LLM agents to iteratively refine ligands against a target pocket, yet a viable ligand must satisfy two often\-conflicting objectives—binding affinity and druggability—which single optimization steps rarely improve together\. To quantify this difficulty, we introduce two diagnostic metrics: the first measures how often a single edit improves both objectives, and the second measures how often a gain on one objective comes with a loss on the other\. Applying these diagnostics to current LLM\-agent pipelines exposes a consistent failure mode: the agent performs molecular editing without knowing how the pocket\-ligand complex responds to local modifications, thus rarely achieving joint improvement\. Inspired by medicinal chemists, who probe the pocket\-ligand complex with controlled analog edits before choosing an optimization direction, we proposePROBE, an optimization framework built around edit–response probing\. PROBE first decomposes the ligand into editable sites and builds a pocket\-specificsite mapthat flags where joint gains are plausible, where the two objectives are likely in tension, and where liability substructures should be changed; it then performs controlled probe edits whose responses are distilled into anEditManual\. Guided by the site map and EditManual, PROBE runs an iterative multi\-agent loop in which an affinity agent, a druggability agent, and a co\-optimization agent jointly produce edits\. On the CrossDocked2020 benchmark, PROBE achieves state\-of\-the\-art performance and substantially mitigates the failure modes exposed by our diagnostics metrics\.

## 1Introduction

Structure\-based drug design \(SBDD\)\[[37](https://arxiv.org/html/2606.00555#bib.bib1),[22](https://arxiv.org/html/2606.00555#bib.bib2)\]aims to generate candidate ligands whose 3D geometries are complementary to a specific protein binding pocket, providing a cost\-effective alternative to high\-throughput screening in early\-stage drug discovery\. Recent deep 3D generative models \(including autoregressive\[[27](https://arxiv.org/html/2606.00555#bib.bib7),[29](https://arxiv.org/html/2606.00555#bib.bib8)\], diffusion\-based\[[15](https://arxiv.org/html/2606.00555#bib.bib3),[9](https://arxiv.org/html/2606.00555#bib.bib4),[16](https://arxiv.org/html/2606.00555#bib.bib6),[33](https://arxiv.org/html/2606.00555#bib.bib5),[24](https://arxiv.org/html/2606.00555#bib.bib13),[42](https://arxiv.org/html/2606.00555#bib.bib14)\], and language\-model\-based approaches\[[39](https://arxiv.org/html/2606.00555#bib.bib9),[6](https://arxiv.org/html/2606.00555#bib.bib10),[19](https://arxiv.org/html/2606.00555#bib.bib11)\]\) have achieved considerable progress in capturing structural complementarity\. However, these models cast SBDD as a one\-time conditional generation task: A candidate molecule is produced in a single forward pass, with no mechanism to iteratively inspect, critique, and repair it against pocket\-specific feedback\[[12](https://arxiv.org/html/2606.00555#bib.bib12),[18](https://arxiv.org/html/2606.00555#bib.bib33)\]\. In contrast, medicinal chemists typically optimize ligands through iterative design–make–test–analyze \(DMTA\) rounds\[[38](https://arxiv.org/html/2606.00555#bib.bib55),[30](https://arxiv.org/html/2606.00555#bib.bib56)\], where each round uses evidence from previous analogs to fix concrete issues\.

This mismatch motivates a growing line of work that employs LLMs as agents to drive iterative, post\-generation optimization of SBDD candidates\[[34](https://arxiv.org/html/2606.00555#bib.bib36),[12](https://arxiv.org/html/2606.00555#bib.bib12),[2](https://arxiv.org/html/2606.00555#bib.bib42)\]\. Existing pipelines mainly fall into two paradigms: \(i\) the LLM directly proposes edited molecules, relying on its learned chemical priors to decide which atoms or fragments should be modified\[[34](https://arxiv.org/html/2606.00555#bib.bib36),[12](https://arxiv.org/html/2606.00555#bib.bib12)\]; \(ii\) the LLM acts as a planner that selects an optimization objective and delegates the actual modification to an external algorithmic module\[[2](https://arxiv.org/html/2606.00555#bib.bib42)\]\. In both paradigms, the agent only sees feedback after an optimization edit has been finished, so exploration and optimization are entangled within the same step\. This is problematic because SBDD intrinsically involves competing multiple objectives\[[41](https://arxiv.org/html/2606.00555#bib.bib15),[24](https://arxiv.org/html/2606.00555#bib.bib13),[42](https://arxiv.org/html/2606.00555#bib.bib14)\]: a viable candidate must simultaneously have high binding affinity, typically measured by the AutoDock Vina scores \(Vina\)\[[36](https://arxiv.org/html/2606.00555#bib.bib20)\], and high druggability, typically measured by Quantitative Estimate of Drug\-likeness \(QED\)\[[5](https://arxiv.org/html/2606.00555#bib.bib21)\]and synthetic accessibility \(SA\)\[[10](https://arxiv.org/html/2606.00555#bib.bib22)\]\. These objectives frequently pull in opposite directions under any local edit: for example, attaching a hydrophobic group to fill a pocket subsite often improves the Vina score but lowers QED and SA by increasing molecular size and complexity\.

To characterize this challenge, we introduce two diagnostic metrics\. The first measures how often an optimization step improves both affinity and druggability at the same time\. The second metric measures how often achieving improvement in one objective leads to a degradation in another\. Across the two LLM\-agent paradigms\(See Section[3](https://arxiv.org/html/2606.00555#S3)\), these diagnostics reveal a consistent pattern: current agents rarely achieve joint improvement within a single edit, and their gains are often offset by objective interference\. This suggests that the main bottleneck is not simply the choice of agent paradigms\. Rather, current agents commit to edits before they have any pocket\-specific evidence about how the ligand will respond\. What is missing is a method to estimate edit responses before the iterative optimization process\.

Our key idea is to make this estimation explicit and conduct it before optimization\. Medicinal chemists often run small sets of controlled analog edits to see how a pocket responds before committing to a larger optimization direction\[[30](https://arxiv.org/html/2606.00555#bib.bib56)\]\. Inspired by this practice, we proposePROBE, a framework for LLM agent SBDD optimization that*probes before editing*, whereas prior agents*edit before knowing*\. Given an initial pocket–ligand complex, PROBE first decomposes the ligand into editable sites and builds a*site map*\. The site map marks where joint improvement is plausible, where the two objectives are likely to be in tension, and where liability substructures should be changed\. PROBE then performs controlled probe edits on the sites identified in the site map and records how affinity and druggability respond\. The responses are distilled into an*EditManual*, a site\-level guide that lists favorable edit directions, modifications to avoid, and chemical constraints for each site\.

During iterative optimization, PROBE uses the*site map*and*EditManual*to guide a multi\-agent optimization loop\. An affinity agent and a druggability agent propose edits from different priorities while grounded in the same guidance\. A co\-optimization agent then reconciles these proposals by combining compatible edits across different sites or resolving conflicts at the same site according to the*EditManual*\. This lets PROBE pursue joint improvement directly, rather than relying on independent single\-objective edits to align by chance\. Our contributions are summarized as follows:

- •We introduce diagnostic metrics for measuring joint improvement and objective interference in LLM\-agent\-based SBDD optimization, and use them to analyze the limitations of existing pipelines\.
- •We proposePROBE, a*probe\-before\-edit*framework that estimates pocket\-specific edit responses before optimization, distills them into a*site map*and an*EditManual*, and uses them to guide multi\-objective molecular editing\.
- •We evaluate PROBE on the CrossDocked2020 benchmark, where it achieves state\-of\-the\-art results on standard SBDD metrics and substantially reduces the failure modes revealed by our diagnostics\.

![Refer to caption](https://arxiv.org/html/2606.00555v1/x1.png)Figure 1:Overview of the PROBE\. \(a\) Site Map Construction\. \(b\) Pocket\-Specific Edit Manual Induction\. \(c\) Multi\-Agent Optimization\.
## 2Related Work

Deep generative models for structure\-based drug design\.Deep generative models for SBDD generate ligands structurally complementary to a given protein pocket, spanning autoregressive\[[27](https://arxiv.org/html/2606.00555#bib.bib7),[29](https://arxiv.org/html/2606.00555#bib.bib8)\], diffusion\-based\[[15](https://arxiv.org/html/2606.00555#bib.bib3),[16](https://arxiv.org/html/2606.00555#bib.bib6),[33](https://arxiv.org/html/2606.00555#bib.bib5),[9](https://arxiv.org/html/2606.00555#bib.bib4),[24](https://arxiv.org/html/2606.00555#bib.bib13),[42](https://arxiv.org/html/2606.00555#bib.bib14)\], and language\-model approaches\[[39](https://arxiv.org/html/2606.00555#bib.bib9),[6](https://arxiv.org/html/2606.00555#bib.bib10),[19](https://arxiv.org/html/2606.00555#bib.bib11)\]\. These methods share a one\-shot conditional\-generation formulation that produces a candidate in a single forward pass\. We treat their outputs as initial candidates and study what should happen*after*generation: how to refine a given pocket–ligand complex with pocket\-specific evidence\.

LLM agents for structure\-based drug design\.Recent works use LLM agents to iteratively refine SBDD ligands\[[34](https://arxiv.org/html/2606.00555#bib.bib36),[12](https://arxiv.org/html/2606.00555#bib.bib12),[2](https://arxiv.org/html/2606.00555#bib.bib42)\]in two paradigms\. \(i\) MoLLM\[[34](https://arxiv.org/html/2606.00555#bib.bib36)\]and CIDD\[[12](https://arxiv.org/html/2606.00555#bib.bib12)\]let LLM agents directly generate edited molecules\. \(ii\) LIDDIA\[[2](https://arxiv.org/html/2606.00555#bib.bib42)\]uses the LLM as a planner that, at each step, picks one target from \{Vina, QED, SA\} and dispatches it to an algorithmic executor for optimization\.111✓\*\*\* rewrite\. it’s not clear what obj refers to hereAll these paradigms only see feedback*after*an edit is finished, mixing exploration with optimization\. So they are all based on a blindly pocket\-agnostic multi\-objective trade\-off\. MoLLM weighs all objectives jointly, CIDD treats affinity as a constraint on druggability, and LIDDIA cycles one objective on the executor per step, so gains on one objective often hurt another\. In contrast, PROBE*probes before editing*: a probing stage first produces a pocket\-specific*site map*and*EditManual*marking where joint improvement is plausible and where objectives conflict, then guides an affinity\-druggability co\-optimization agent loop that grounds the multi\-objective optimization in measured evidence\.

222✓\*\*\* since u put quite some emphasize on the MOO nature of SBDD\. good to add a short review on that here \(and why current methods dont quite work\)
## 3Limitations of Existing LLM Agents for SBDD

To characterize the limitations of current LLM agents on multi\-objective SBDD, we experiment with the two paradigms in Section[2](https://arxiv.org/html/2606.00555#S2)\. For agents that use the LLM to generate edited molecules, we have \(i\) MOLLM\[[34](https://arxiv.org/html/2606.00555#bib.bib36)\], which performs multi\-objective global edits\. We also augment it with PLIP\[[35](https://arxiv.org/html/2606.00555#bib.bib41)\]text reports so that it can perceive 3D interactions; and \(ii\) CIDD\[[12](https://arxiv.org/html/2606.00555#bib.bib12)\], which targets modifications on low\-druggability substructures while maintaining or improving its affinity\. Moreover, to study whether textual guidance is enough to induce multi\-objective trade\-off behavior, we add a variant \(iii\) CIDD\+MOO, which injects explicit multi\-objective trade\-off instructions into the CIDD prompt \(full prompt in Appendix[A](https://arxiv.org/html/2606.00555#A1)\)\. For planner\-plus\-executor agents, we have \(iv\) LIDDIA\[[2](https://arxiv.org/html/2606.00555#bib.bib42)\], whose LLM reasoner only decides which single objective in \{Vina, QED, SA\} is the current bottleneck at the planning stage, and then dispatches optimization to a single\-objective graph\-based genetic algorithm executor GraphGA\[[23](https://arxiv.org/html/2606.00555#bib.bib43)\]\. All four methods are run under the same pocket set, starting molecules, and number of optimization rounds\. Details on the experimental setup are in Section[5](https://arxiv.org/html/2606.00555#S5)\.

Table 1:Statistics for change on affinity \(Δ​A\\Delta A\) and druggability \(Δ​D\\Delta D\) after one optimization step\. Positive values indicate improvement\. For each method, the largest outcome is marked inbold\.333✓I write more details in the introduction about the objectives and metrics\. \*\*\* yeah, but it’s too short \*\*\* we have some short intro of ’binding affinity represent by vina score and druggability represent by QED, SA’ in the introduction\. The details of these metrics are in appendix[D](https://arxiv.org/html/2606.00555#A4)\. \*\*\* first, talk about the objs\. it’s better to intro things like vina, QED, SA, etc somewhere \(maybe in appendix if u really run out of space, but a short desc in the main text will make the paper more readable\)Improvement statistics\.For a moleculemm, we define the affinity scoreA​\(m\)=−Vina​\(m\)A\(m\)=\-\\mathrm\{Vina\}\(m\)and druggability scoreD​\(m\)=QED​\(m\)\+SA​\(m\)D\(m\)=\\mathrm\{QED\}\(m\)\+\\mathrm\{SA\}\(m\)\. For both objectives, the larger the better\. We defineΔ​A=A​\(m′\)−A​\(m\)\\Delta A=A\(m^\{\\prime\}\)\-A\(m\)andΔ​D=D​\(m′\)−D​\(m\)\\Delta D=D\(m^\{\\prime\}\)\-D\(m\), wherem′m^\{\\prime\}is the optimized molecule after one optimization step\. We classify the steps to four categories: \(i\)*joint improvement*\(Δ​A\>0,Δ​D\>0\\Delta A\>0,\\Delta D\>0\), \(ii\)*affinity\-only improvement*\(Δ​A\>0,Δ​D≤0\\Delta A\>0,\\Delta D\\leq 0\), \(iii\)*druggability\-only improvement*\(Δ​A≤0,Δ​D\>0\\Delta A\\leq 0,\\Delta D\>0\), and \(iv\)*joint degradation*\(Δ​A≤0,Δ​D≤0\\Delta A\\leq 0,\\Delta D\\leq 0\)\. Table[1](https://arxiv.org/html/2606.00555#S3.T1)counts the number of steps in each category\. As can be seen, for all the baselines, fewer than one\-third of the steps lead to joint improvement, while most of the steps can only improve one objective at best\.

Table 2:Intent ratio, IOC rate, and OI rate of each pipeline\. “n/a” marks IOC cells where the intent is never declared\.Intent does not match outcome\.From the LLM’s reasoning text in each optimization step, we use Gemini\-3\.1\-Pro to see whether its intent is on improving the affinity, druggability, or both \(full prompt in Appendix[A](https://arxiv.org/html/2606.00555#A1)\)\. To measure whether the optimized molecule realizes the intent, we define three*Intent–Outcome Consistency*\(IOC\) measures: \(i\)IOCAffinity=Pr⁡\(Δ​A\>0∣intent=affinity\)\\operatorname\{IOC\}\_\{\\mathrm\{Affinity\}\}=\\Pr\(\\Delta A\>0\\mid\\text\{intent\}=\\mathrm\{affinity\}\), \(ii\)IOCDruggability=Pr⁡\(Δ​D\>0∣intent=druggability\)\\operatorname\{IOC\}\_\{\\mathrm\{Druggability\}\}=\\Pr\(\\Delta D\>0\\mid\\text\{intent\}=\\mathrm\{druggability\}\), and \(iii\)IOCJoint=Pr⁡\(Δ​A\>0∧Δ​D\>0∣intent=joint\)\\operatorname\{IOC\}\_\{\\mathrm\{Joint\}\}=\\Pr\(\\Delta A\>0\\wedge\\Delta D\>0\\mid\\text\{intent\}=\\mathrm\{joint\}\)\.

As shown in Table[2](https://arxiv.org/html/2606.00555#S3.T2), MOLLM, CIDD, and CIDD\+MOO achieve substantially higher IOC under single\-objective intents than under joint intent\. With the addition of multi\-objective instructions, CIDD\+MOO changes the intent distribution, but does not improve the joint\-intent IOC\. LIDDIA behaves differently\. Its planner must choose exactly one bottleneck objective at each step, and so cannot declare a joint intent explicitly\.

One objective improves while the other objective deteriorates\.For single\-objective intents, we define*Objective Interference \(OI\)*as the probability that the selected objective improves while the other objective does not:OIAffinity=Pr⁡\(Δ​D​<0∣Δ​A\>​0∧intent=Affinity\)\\operatorname\{OI\}\_\{\\mathrm\{Affinity\}\}=\\Pr\(\\Delta D<0\\mid\\Delta A\>0\\wedge\\text\{intent\}=\\mathrm\{Affinity\}\)andOIDruggability=Pr⁡\(Δ​A​<0∣Δ​D\>​0∧intent=Druggability\)\\operatorname\{OI\}\_\{\\mathrm\{Druggability\}\}=\\Pr\(\\Delta A<0\\mid\\Delta D\>0\\wedge\\text\{intent\}=\\mathrm\{Druggability\}\)\. As shown in Table[2](https://arxiv.org/html/2606.00555#S3.T2), for all the baselines, improvement in one objective frequently leads to deterioration of the other\. In particular, while LIDDIA can realize the intent reliably \(high IOC\), its OI is the highest because the unselected objective is not constrained in the genetic algorithm executor\.

## 4Method

For more effective SBDD optimization, the LLM agent needs to know which local sites can support joint improvement, rather than merely declaring a joint intent during LLM reasoning\. Moreover, on trying to improve one objective, it has to know which local edit directions can preserve the unselected objectives\. To address these requirements, in this section we introduce PROBE\. It first constructs a pocket\-specific site map \(Section[4\.1](https://arxiv.org/html/2606.00555#S4.SS1)\), then induces an EditManual from controlled probe edits \(Section[4\.2](https://arxiv.org/html/2606.00555#S4.SS2)\), and finally uses the resulting evidence in an iterative multi\-agent optimization loop \(Section[4\.3](https://arxiv.org/html/2606.00555#S4.SS3)\)\. The whole pipeline is shown in Figure[1](https://arxiv.org/html/2606.00555#S1.F1)\.

### 4\.1Site Map Construction

In this section, we construct a pocket\-specific*site map*before optimization\. The site map is a set of planning priors induced from PLIP, geometric, and property signals over the current pocket–ligand complex\. It marks where joint improvement is plausible, where the two objectives are likely in tension, and where liability substructures444✓change to liability, which is the usual call\.\*\*\* ”hazard substructures”? is this how they’re usually called?should be fixed\. The map is built once555✓It will be updated in the Multi\-Agent Optimization, this section only talks about site map\. About the update on edit manual, it is talked in Sec 4\.3 Analyze and iterate\.\*\*\* what about the edit manual?and reused across rounds, supplying the downstream planner with a structural prior for intent selection and edit implementation\.

Diagnosis signal aggregation\.Following the post\-hoc optimization pipeline\[[12](https://arxiv.org/html/2606.00555#bib.bib12)\], which starts from an initial ligand molecule666✓ligand is a molecule for protein pocket\. \*\*\* u dont start from the given molecule???produced by a deep generative model, we decompose the ligand with BRICS\[[8](https://arxiv.org/html/2606.00555#bib.bib32)\]into sub\-molecule fragments\{fj\}\\\{f\_\{j\}\\\}\. From each fragment, we aggregate three complementary streams of diagnosis signals:*interaction*signals from PLIP\[[35](https://arxiv.org/html/2606.00555#bib.bib41)\],*geometric*signals describing how the fragment fits the pocket, and*property*signals covering ligand efficiency\[[1](https://arxiv.org/html/2606.00555#bib.bib60)\]and known problematic groups\. Details of the signal extraction are in Appendix[B](https://arxiv.org/html/2606.00555#A2)\.777✓\*\*\* where are the details on how they are extracted?

Site selection and labeling\.A diagnosis LLM reads the fragment\-level signals888✓rewrited \*\*\* i dont know what this means\. will this be understandable to the reviewers?999✓rewrited \*\*\* u mean all the segments of the whole ligand?to select a subset of fragments as sites\{si\}\\\{s\_\{i\}\\\}\. It first writes a short holistic profile of the complex, summarizing the pocket context, anchor interactions, ligand efficiency, and druggability liabilities\. Conditioned on this profile, it marks each site assi=⟨fragment​\_​name,atom​\_​indices,local​\_​symptoms,type⟩,s\_\{i\}=\\langle\\mathrm\{fragment\\\_name\},\\mathrm\{atom\\\_indices\},\\mathrm\{local\\\_symptoms\},\\mathrm\{type\}\\rangle,where the type can be \(i\)Synergy, indicating a site where one edit can plausibly improve both affinity and druggability; \(ii\)Tension, where improving one objective is likely to hurt the other; or \(iii\)Liability, which contains a structural liability, such as a reactive alert, excessive chirality, or an overly complex spiro system, that should be repaired regardless of the trade\-off\. When multiple labels appear plausible, the LLM chooses the dominant one\. The full prompt is shown in Appendix[A](https://arxiv.org/html/2606.00555#A1)\.

### 4\.2Pocket\-Specific Edit Manual Induction

The site map gives the locations of high\-value edits, but does not determine which edit directions are safe for the specific pocket–ligand complex\. PROBE therefore probes the mapped sites before optimization\. The probing runs only once, before the iterative optimization loop begins\.101010✓\*\*\* desc clearly how this is diff from opt\. \(i suppose opt has multi steps while this one has only one\)The probe molecules are controlled perturbations used to observe how affinity and druggability respond\. The observations are distilled into a pocket\-specific EditManual\.

Probing planner\.Given the site map, a probing planner LLM proposes three strategies along the affinity\-druggability trade\-off:*Affinity\-first*,*Druggability\-first*, and*Co\-optimization*\. Each strategy specifies targeted sites, a chemically explicit edit prescription, and the trade\-off it accepts\. For example, an affinity\-first strategy may extend a donor toward an unfilled subpocket, while a druggability\-first strategy may replace a complex chiral group with an achiral bioisostere\. The full prompt is given in Appendix[A](https://arxiv.org/html/2606.00555#A1)\.

Probes\.For each strategy \(Affinity\-first, Druggability\-first, Co\-optimization\), a probing generator LLM generates four structured edits\. Each edit contains a target atom, an action, and semantic constraints over the fragment to be retrieved\. The first three probes apply the strategy at three edit magnitudes \(high, medium, low\),111111✓changed \*\*\* what do u mean by ”forward intensity”? as mentioned before, dont add words just to make it look fancy121212✓\*\*\* dont know what it means\. rewrite using simple plain englishso we can observe whether increasing the edit magnitude continues to improve the objectives or begins to degrade them, for example, whether filling a pocket void with a ring outperforms a small isosteric tweak\. The fourth probe is a counterfactual that reverses the strategy’s direction as a control: the strategy is credible only if the forward probes improve the response, while the counterfactual does not\.131313✓\*\*\* simplify this sentenceWith the three strategies, this gives a total of 12 probes per input molecule \(Table[3](https://arxiv.org/html/2606.00555#S4.T3)\)\. A141414✓removed \*\*\* why need the word ”deterministic”?fragment\-assembly engine instantiates these edits as valid molecules \(Appendix[C](https://arxiv.org/html/2606.00555#A3)\)\. All probes are then scored to yieldΔ​A\\Delta AandΔ​D\\Delta D\. The full prompt is shown in Appendix[A](https://arxiv.org/html/2606.00555#A1)\.

Table 3:The3×43\\times 4probing matrix\. Each row is a strategy used by the probing planner and each column is an edit\.Response summarization\.Before manual construction, an151515✓removed \*\*\* why do u always emphasize it’s deterministic?analyzer converts the 12 raw outcome scores into a response summary while retaining the exact deltas\. For each strategy, it reports the shape of the high–medium–low response, such as: monotone improvement, activity cliff, saturation, or flat/negative response\. It also reports the counterfactual signal: whether reversing the strategy makes the scores worsen, improve, or remain unchanged\. This gives the manual\-construction LLM both numerical evidence and a stable qualitative interpretation\. The full prompt is shown in Appendix[A](https://arxiv.org/html/2606.00555#A1)\.

EditManual construction\.The EditManual is built by distilling the scored probe outcomes and response summaries into structured constraints\. A manual\-construction LLM reads the site map, strategies, probe outcomes, and summarized response patterns, and then generates a structured EditManual\. For each site, the manual records: \(i\) a pharmacophore role \(anchor, linker, hydrophobic core, etc\.\); \(ii\) allowed and forbidden actions, each justified by the probe outcome, together with quantitative constraints on size, polarity, flexibility, and shape that any future fragment placed at that site must follow; and \(iii\) a blacklist of structural features that must not be introduced\. It also records cross\-site rules indicating which sites can be edited independently and which edit directions are mutually exclusive\. The EditManual therefore converts local probe responses into executable constraints for the optimization loop\. The full prompt is shown in Appendix[A](https://arxiv.org/html/2606.00555#A1)\.

### 4\.3Multi\-Agent Optimization

Next, PROBE refines the ligand molecule through a design\-make\-test\-analyze loop\[[38](https://arxiv.org/html/2606.00555#bib.bib55)\]\.161616✓\*\*\* refEach round is guided with the site map and the EditManual\. The site map supplies where the agents could act, while the EditManual constrains how each local edit should be designed\.

Design\.PROBE uses three role\-specialized agents \(affinity agent, druggability agent, and co\-optimization agent\) conditioned on the same site map, EditManual, and edit history\. The affinity agent proposes one localized edit under an affinity\-first prior, preferentially acting onSynergysites or the affinity\-favorable side ofTensionsites while preserving manual\-protected anchor interactions\. The druggability agent proposes one localized edit under a druggability\-first prior, preferentially acting onLiabilitysites or the property\-liability side ofTensionsites through simplification, pruning, or bioisosteric replacement when allowed by the manual\. Since the molecule changes across rounds while the site map is defined on the original ligand, both agents first resolve the selected site to its current atom indices before generating a structured edit\.

The two drafts are then cross\-reviewed171717✓no, only the revised drafts\. \*\*\* are the reviews sent to the co\-optimization agent?against the same EditManual\. Each agent checks whether the other draft targets the correct current site, violates semantic constraints, or threatens its own objective\. After revision, a co\-optimization agent receives the two revised drafts and generates one reconciled edit\. When the drafts are compatible, it preserves the constraints needed by both objectives; when they conflict on the same site, it designs a manual\-compliant hybrid constraint\. Thus, each round yields three designs—affinity\-oriented, druggability\-oriented, and co\-optimized—which are passed to the make\-and\-test stage\. The full prompt is shown in Appendix[A](https://arxiv.org/html/2606.00555#A1)\.

Make and test\.All three designs are instantiated by the same181818✓removed \*\*\* again… deterministic…fragment\-assembly engine used in the probing stage \(Appendix[C](https://arxiv.org/html/2606.00555#A3)\)\. The engine retrieves fragments satisfying the semantic constraints, applies the requested local action, and returns the candidate\. Each candidate is then evaluated, yielding itsΔ​A\\Delta AandΔ​D\\Delta Drelative to the current parent molecule\.191919✓changed \*\*\* lead?

Analyze and iterate\.The analysis stage updates the search state without discarding accumulated evidence\. First, each candidate is labeled as success or failure with respect to the intent that produced it\. Second, the edit history is updated with the targeted site, action, constraints, and observed outcome\. Third, the EditManual is revised incrementally when an outcome contradicts an existing rule: the affected entry is tightened or rewritten, and the failed action is recorded to prevent repeated attempts\. After updating the state, we select the best molecule for the next round from the three candidates produced in the current round\. The selection is based on the hypervolume\[[17](https://arxiv.org/html/2606.00555#bib.bib57)\]of the candidatemmwith respect to the initial moleculem0m\_\{0\}\(i\.e\., the input ligand at the start of the optimization loop\) in the\(A,D\)\(A,D\)space:

Sbal​\(m\)=\{\(A​\(m\)−A​\(m0\)\|A​\(m0\)\|\)​\(D​\(m\)−D​\(m0\)D​\(m0\)\)if​A​\(m\)\>A​\(m0\)​and​D​\(m\)\>D​\(m0\),0otherwise\.S\_\{\\mathrm\{bal\}\}\(m\)=\\begin\{cases\}\\left\(\\dfrac\{A\(m\)\-A\(m\_\{0\}\)\}\{\|A\(m\_\{0\}\)\|\}\\right\)\\left\(\\dfrac\{D\(m\)\-D\(m\_\{0\}\)\}\{D\(m\_\{0\}\)\}\\right\)&\\text\{if \}A\(m\)\>A\(m\_\{0\}\)\\text\{ and \}D\(m\)\>D\(m\_\{0\}\),\\\\\[6\.0pt\] 0&\\text\{otherwise\.\}\\end\{cases\}Usingm0m\_\{0\}as the reference point ensures that candidates failing to improve both objectives receive zero contribution\. The candidate with the largestSbalS\_\{\\mathrm\{bal\}\}is promoted as the start for the next round, and the iteration runs for a fixed number of optimization rounds\.

Final candidate selection\.After all rounds finish, PROBE returns the molecule with the largestSbalS\_\{\\mathrm\{bal\}\}from a pool that includes the initial molecule, all 12 probe molecules, and every molecule produced during iterative optimization\. The probes are counted as part of PROBE’s search budget, and the full resulting computation costs are reported in the compute budget analysis \(Section[5\.4](https://arxiv.org/html/2606.00555#S5.SS4)\)\.

## 5Experiments

Following CIDD\[[12](https://arxiv.org/html/2606.00555#bib.bib12)\], we conduct experiments on the CrossDocked2020 dataset\[[11](https://arxiv.org/html/2606.00555#bib.bib19)\]and evaluate models using the standard SBDD metrics\. Details of the dataset and metrics are in Appendix[D](https://arxiv.org/html/2606.00555#A4)\.

We compare PROBE with two categories of baselines\. The first category includes*De novo 3D\-generation models*: \(i\) autoregressive models: AR\[[27](https://arxiv.org/html/2606.00555#bib.bib7)\]and Pocket2Mol\[[29](https://arxiv.org/html/2606.00555#bib.bib8)\], \(ii\) LLM\-based: TamGen\[[39](https://arxiv.org/html/2606.00555#bib.bib9)\], LMLF\[[6](https://arxiv.org/html/2606.00555#bib.bib10)\], ELILLM\[[19](https://arxiv.org/html/2606.00555#bib.bib11)\]; and \(iii\) diffusion\-based: IDOLpro\[[24](https://arxiv.org/html/2606.00555#bib.bib13)\], TAGMol\[[9](https://arxiv.org/html/2606.00555#bib.bib4)\], DrugGPS\[[40](https://arxiv.org/html/2606.00555#bib.bib27)\], IPDiff\[[20](https://arxiv.org/html/2606.00555#bib.bib26)\], DecompDiff\[[16](https://arxiv.org/html/2606.00555#bib.bib6)\], MOC\[[42](https://arxiv.org/html/2606.00555#bib.bib14)\], MolCRAFT\[[33](https://arxiv.org/html/2606.00555#bib.bib5)\], MolPilot\[[31](https://arxiv.org/html/2606.00555#bib.bib29)\], DecompDPO\[[7](https://arxiv.org/html/2606.00555#bib.bib17)\], and MolJO\[[32](https://arxiv.org/html/2606.00555#bib.bib30)\]\.

The second category includes*LLM\-agent methods*, including MOLLM\[[34](https://arxiv.org/html/2606.00555#bib.bib36)\]\(paired with PLIP\[[35](https://arxiv.org/html/2606.00555#bib.bib41)\]reports\), LIDDIA\[[2](https://arxiv.org/html/2606.00555#bib.bib42)\], and CIDD\[[12](https://arxiv.org/html/2606.00555#bib.bib12)\]\. For every protein pocket in CrossDocked2020, each LLM\-agent method starts from1010initial molecules produced by a deep 3D generator \(TAGMol/MolCRAFT/MolJO trained on CrossDocked2020\) and refines each through55DMTA iterations, yielding1010optimized molecules per pocket\. The quantitative analysis in Section[3](https://arxiv.org/html/2606.00555#S3)is based on MolCRAFT initial molecules\. We use GPT\-4o\-mini as the backbone for all agent roles\. All the experiments are conducted on 64 parallel workers of a server with 96 Intel\(R\) Xeon\(R\) Gold 6342 CPUs\. The detailed prompts are in Appendix[A](https://arxiv.org/html/2606.00555#A1)\. PROBE has an extra probing stage whose overhead is separately analyzed in Section[5\.4](https://arxiv.org/html/2606.00555#S5.SS4)\. The experiment on the effect of different LLM Backbones is in Appendix[E](https://arxiv.org/html/2606.00555#A5)\.

Table 4:Performance on CrossDocked2020\. Performance ranking per column is color\-coded as follows:best,second\-best\\rowcolorheadercolorModelVina↓\\downarrowQED↑\\uparrowSA↑\\uparrowSuccess↑\\uparrowMRR↑\\uparrowQikProp↑\\uparrowLipinski↑\\uparrowLogPMWRank↓\\downarrow\\cellcolorsubheadercolorDe novo 3D\-generation modelsAR\-6\.6130\.5060\.6355\.98%64\.85%20\.31%4\.760\.45247\.6127Pocket2Mol\-7\.2450\.5670\.75522\.82%70\.97%29\.83%4\.831\.70243\.5415TamGen\-7\.4760\.5080\.5738\.01%24\.62%30\.33%4\.444\.16326\.4624LMLF\-7\.7760\.5150\.5728\.31%23\.52%25\.83%4\.434\.18327\.4923ELILLM\-7\.9300\.5280\.5799\.81%24\.02%27\.93%4\.424\.22328\.0922IDOLpro\-7\.6110\.5480\.5687\.60%38\.20%30\.10%4\.721\.81342\.7226TAGMol\-7\.5710\.5500\.5717\.80%40\.00%31\.10%4\.731\.86337\.4825DrugGPS\-7\.2950\.4650\.63613\.70%56\.70%26\.50%4\.451\.39328\.8221IPDiff\-7\.8060\.5210\.59314\.50%26\.30%23\.50%4\.533\.46331\.6020DecompDiff\-7\.9490\.4490\.60419\.80%64\.24%28\.89%4\.312\.45424\.1418MOC\-7\.6850\.5590\.64722\.60%25\.00%40\.60%4\.783\.09304\.2016MolCRAFT\-7\.6790\.5030\.68424\.80%63\.50%22\.20%4\.441\.18327\.3313MolPilot\-7\.8180\.5520\.73630\.00%77\.50%32\.30%4\.601\.80323\.9711DecompDPO\-8\.4270\.4710\.66230\.40%55\.10%15\.80%4\.205\.01435\.339MolJO\-8\.6630\.5600\.76346\.80%41\.30%14\.90%4\.554\.18308\.816\\cellcolorsubheadercolorLLM\-agent optimization methodsTAGMol \+ MoLLM\-8\.0520\.5660\.62116\.65%47\.31%29\.26%4\.812\.54357\.5319TAGMol \+ LIDDIA\-7\.7240\.5560\.67921\.52%55\.43%33\.71%4\.812\.83374\.8217TAGMol \+ CIDD\-7\.9440\.6410\.66524\.53%59\.19%47\.79%4\.743\.01333\.5114TAGMol \+ PROBE\-9\.168\\cellcolorsotacolor0\.7020\.69648\.43%62\.18%\\cellcolorsotacolor55\.63%\\cellcolorsecondcolor4\.852\.96365\.224MolCRAFT \+ MoLLM\-8\.1350\.5140\.69830\.32%65\.27%24\.06%4\.692\.26346\.7810MolCRAFT \+ LIDDIA\-7\.9340\.5240\.71228\.25%69\.22%27\.14%4\.512\.27362\.7312MolCRAFT \+ CIDD\-7\.9460\.5980\.73931\.40%\\cellcolorsecondcolor70\.60%37\.10%4\.692\.07331\.588MolCRAFT \+ PROBE\\cellcolorsecondcolor\-9\.203\\cellcolorsecondcolor0\.689\\cellcolorsecondcolor0\.766\\cellcolorsecondcolor63\.76%\\cellcolorsotacolor79\.58%\\cellcolorsecondcolor52\.61%\\cellcolorsotacolor4\.852\.97359\.532MolJO \+ MoLLM\-8\.7550\.5560\.74643\.61%42\.86%21\.31%4\.604\.38331\.617MolJO \+ LIDDIA\-8\.6930\.5670\.76548\.44%56\.93%29\.65%4\.593\.79354\.363MolJO \+ CIDD\-8\.5680\.6320\.76247\.80%57\.05%30\.16%4\.634\.18314\.595MolJO \+ PROBE\\cellcolorsotacolor\-9\.6260\.664\\cellcolorsotacolor0\.782\\cellcolorsotacolor67\.67%66\.94%34\.69%4\.674\.16344\.321

### 5\.1Performance Results

Table[4](https://arxiv.org/html/2606.00555#S5.T4)202020✓no, Log P should be with a optimal range, not higher is better\. \*\*\* for LogP, is it higher the better?212121✓\*\*\* really need 3 decimal places? others also do that? similarly for the percentages222222✓\*\*\* i’d prefer to provide a ranking for each col for easy comparison \(if some others are also doing that\)shows the multi\-objective optimization performance obtained\. The following can be observed\. \(i\)PROBE yields consistent improvements over the initial molecules\.Regardless of which deep generation model provides the initial molecule, attaching PROBE shows substantial gains on all metrics\.232323✓\*\*\* currently it’s hard to see\. if u hv a ranking, that’ll be easierBesides, the stronger the initial molecules, the stronger the final results\.242424✓\*\*\* also takes time to see\(ii\)PROBE outperforms LLM\-agent baselines\.PROBE consistently outperforms MOLLM, LIDDIA, and CIDD on all metrics\. \(iii\)PROBE closes the affinity\-druggability trade\-off\.252525✓\*\*\* again, this is also hard to seePROBE is the only method that significantly pushes both objectives upward across all three initializer settings\. These results stem from the PROBE’s*probes before editing*pipeline, which estimates how the specific pocket responds to local edits before optimization and turns these observations into guidance for the agents to propose edits improving affinity and druggability together\.

### 5\.2Closing the Bottlenecks

Tables[1](https://arxiv.org/html/2606.00555#S3.T1)and[2](https://arxiv.org/html/2606.00555#S3.T2)show the performance of PROBE alongside the four baselines\. The main findings are:\(i\) Joint progress becomes the dominant outcome\.PROBE is the only method with most of its steps achieving improvements on both affinity and druggability, and its joint\-intent IOC is roughly twice that of the best baseline \(CIDD\+MOO\)\. The*site map*and*EditManual*localize the edit to a site whose label supports both objectives, so the joint intent is reliably realized as joint improvement\.\(ii\) Single\-objective progress with limited collateral cost\.On single\-objective intents, PROBE keeps IOC high while keeping OI well below the baseline range on both axes\. Baselines all have an OI rate larger than 60%, meaning that when they realize the intended objective, they typically damage the other one\. PROBE overcomes this problem due to the explicit edit constraints carried by*EditManual*, so realizing one objective does not force the other down\.

### 5\.3Ablation Study

We ablate the two design choices that distinguish PROBE from other LLM\-agent paradigms in Table[5](https://arxiv.org/html/2606.00555#S5.T5)\. TheSite Mapaxis contrasts*Site*\(fragments organized by the site map\) against*Fragment*\(naive BRICS fragments\)\. TheManualaxis forms a three\-rung ladder:*Signals*forwards only diagnosis signals with no EditManual;*Prior*has the LLM write an EditManual from its own prior knowledge over the same signals, with no probing;*Probed*\(ours\) induces the EditManual from probing evidence\.

Table 5:Ablation study on CrossDocked2020\. We use MolCRAFT to generate initial molecules\.Site Mapcontrols whether fragments are organized by the site map \(Site\) or kept as flat BRICS fragments without site organization \(Fragment\)\.Manualcontrols how edit constraints are produced: raw diagnosis signals only, noEditManual\(Signal\);EditManualwritten by the LLM from its prior knowledge, no probing \(Prior\);EditManualinduced from probing evidence \(Probed, ours\)\.We have the following observations:\(i\) Site map consistently improves\.Comparing*Fragment*with*Site*\(rows 1 vs\. 2, 3 vs\. 4, 5 vs\. 6\), all metrics improve simultaneously\. Localized site map organization, rather than naive fragments, gives downstream edits a precise target for multi\-objective optimization\.\(ii\) Probing makes*EditManual*useful\.Asking the LLM to write an*EditManual*from its prior knowledge \(*Signals*→\\to*Prior*, rows 1 vs\. 3 and 2 vs\. 4\) gives only small and inconsistent changes\. Replacing prior knowledge with probing evidence \(*Prior*→\\to*Probed*, rows 3 vs\. 5 and 4 vs\. 6\) produces a clear jump on both affinity and druggability metrics\. The gain comes from the probing evidence distilled into the manual, not from having a manual\.\(iii\) The two components work together\.The best results on every metric appear when both are used \(row 6\)\. Site map alone \(row 2\) or probed manual alone \(row 5\) each improve over the bare baseline \(row 1\) but stay well below the whole model \(row 6\)\. The site map tells edits where to act, and the probed manual tells them what to do there\. Both are needed for joint affinity\-druggability progress\.

![Refer to caption](https://arxiv.org/html/2606.00555v1/Img/computation_cost_bar_chart.png)Figure 2:Computational budget and performance comparison\. CIDD \(27 Rounds\) is the baseline extended to match PROBE’s theoretical Vina budget\.
### 5\.4Compute Budget Analysis

In this experiment, we compare PROBE with CIDD, the best baseline in Table[4](https://arxiv.org/html/2606.00555#S5.T4)\.262626✓\*\*\* add272727✓yes they are aligned, I remove this duplicated parts and merge into the exp setting paragraph\. \*\*\* in tbl4, they should be based on an ”aligned” setup, right? otherwise, the comparison is not fair\!Since Vina docking takes significantly longer than an LLM call, it dominates the wall\-clock time\. Therefore, we use the per\-molecule Vina\-call budget for comparison\. A 5\-round CIDD run admits at most 5 Vina calls per molecule, whereas a 5\-round PROBE run admits up to 27 \(12 from probing, 15 from the multi\-agent DMTA cycles\)\. To282828✓\*\*\* can u just use simple non\-fancy english?ensure that the performance gap is not just caused by a larger budget, we additionally extend CIDD to 27 rounds so that its theoretical Vina budget matches PROBE’s\.

292929✓\*\*\* pls remove non\-essential wordsFigure[2](https://arxiv.org/html/2606.00555#S5.F2)shows the per\-molecule cost and performance\. Against 5\-round CIDD, PROBE uses more computation as expected\. The key comparison is with budget\-matched CIDD \(27 rounds\): it spends comparable Vina calls and more LLM tokens than PROBE, yet it yields very small improvements over 5\-round CIDD, showing that simply enlarging the CIDD cost is inefficient\. PROBE instead uses the site map and EditManual to constrain where and how each edit happens, reaching a much higher multi\-objective yield within a smaller actual budget\.

![Refer to caption](https://arxiv.org/html/2606.00555v1/x2.png)Figure 3:Comparison on target PLCD1\. Hydrogen bonds and hydrophobic contacts are marked by dashed lines\. Clashes are marked by red lines\. \(a\) initial molecule, \(b\-d\) CIDD pipeline: \(b\) CIDD analysis, \(c\) CIDD design, \(d\) CIDD optimized molecule\. \(e–j\) PROBE pipeline: \(e\) diagnosis signals, \(f,g\) site map, \(h\) EditManual, \(i\) PROBE Design, \(j\) PROBE optimized molecule\.
### 5\.5Qualitative Analysis

Figure[3](https://arxiv.org/html/2606.00555#S5.F3)visualizes the optimization trajectories of CIDD and PROBE on target PLCD1\. The initial molecule exhibits a clear affinity\-druggability trade\-off: it has baseline binding but its phosphate group and spiro system incur steric tension and synthetic penalties\. Without pocket\-specific evidence about how the ligand will respond to the edit, CIDD applies a generic heuristic and replaces the cyclic backbone with an aromatic scaffold\. The unconstrained topology change introduces new 3D clashes and forces a rigid\-body rotation of the full molecule that breaks the pre\-existing hydrogen\-bond and hydrophobic contacts, degrading affinity\. PROBE marks the molecule with explicit sites and induces an EditManual\. Guided by these constraints, PROBE performs small, rigid replacement edits that not only resolve the original steric clashes and reinforce the pre\-existing hydrogen bonds and hydrophobic contacts, but also fill the lower part of the pocket with a new fragment, substantially enhancing affinity without harming druggability\.

## 6Conclusion

In this paper, we studied why current LLM agents struggle to refine ligands in structure\-based drug design\. We showed that existing agents rarely improve binding affinity and druggability simultaneously, and gains on one objective often come at the cost of the other\. These agents decide how to edit a ligand before knowing how the pocket actually responds to local edits\. To address this, we proposed PROBE, which first probes the pocket with controlled edits, summarizes the observed responses into a site map and an EditManual, and then uses these to guide role\-specialized agents during optimization\. PROBE achieves state\-of\-the\-art performance on CrossDocked2020\. It also substantially reduces the failure modes of existing agents revealed by our diagnostics, with joint improvement becoming the dominant per\-edit outcome and objective interference greatly reduced\. We envision that*probe\-before\-edit*could serve as a general pipeline for LLM\-driven molecular optimization, where gathering task\-specific evidence before committing to edits may extend beyond SBDD to other multi\-objective design settings\.

## References

- \[1\]\(2005\)Ligand efficiency indices as guideposts for drug discovery\.Drug discovery today10\(7\),pp\. 464–469\.Cited by:[Appendix B](https://arxiv.org/html/2606.00555#A2.p4.5),[§4\.1](https://arxiv.org/html/2606.00555#S4.SS1.p2.1)\.
- \[2\]R\. Averly, F\. N\. Baker, I\. A\. Watson, and X\. Ning\(2025\)Liddia: language\-based intelligent drug discovery agent\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 12015–12039\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p2.1),[§2](https://arxiv.org/html/2606.00555#S2.p2.1),[§3](https://arxiv.org/html/2606.00555#S3.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p3.3)\.
- \[3\]J\. B\. Baell and G\. A\. Holloway\(2010\)New substructure filters for removal of pan assay interference compounds \(pains\) from screening libraries and for their exclusion in bioassays\.Journal of medicinal chemistry53\(7\),pp\. 2719–2740\.Cited by:[Appendix B](https://arxiv.org/html/2606.00555#A2.p4.5)\.
- \[4\]S\. H\. Bertz\(1981\)The first general index of molecular complexity\.Journal of the American Chemical Society103\(12\),pp\. 3599–3601\.Cited by:[Appendix B](https://arxiv.org/html/2606.00555#A2.p4.5)\.
- \[5\]G\. R\. Bickerton, G\. V\. Paolini, J\. Besnard, S\. Muresan, and A\. L\. Hopkins\(2012\)Quantifying the chemical beauty of drugs\.Nature chemistry4\(2\),pp\. 90–98\.Cited by:[Appendix B](https://arxiv.org/html/2606.00555#A2.p4.5),[Appendix D](https://arxiv.org/html/2606.00555#A4.p2.3),[§1](https://arxiv.org/html/2606.00555#S1.p2.1)\.
- \[6\]S\. B\. Brahmavar, A\. Srinivasan, T\. Dash, S\. R\. Krishnan, L\. Vig, A\. Roy, and R\. Aduri\(2024\)Generating novel leads for drug discovery using llms with logical feedback\.InProceedings of the AAAI conference on artificial intelligence,Vol\.38,pp\. 21–29\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[7\]X\. Cheng, X\. Zhou, Y\. Yang, Y\. Bao, and Q\. GuDecomposed direct preference optimization for structure\-based drug design\.Transactions on Machine Learning Research\.Cited by:[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[8\]J\. Degen, C\. Wegscheid\-Gerlach, A\. Zaliani, and M\. Rarey\(2008\)On the art of compiling and using’drug\-like’chemical fragment spaces\.ChemMedChem3\(10\),pp\. 1503\.Cited by:[Appendix C](https://arxiv.org/html/2606.00555#A3.p3.3),[§4\.1](https://arxiv.org/html/2606.00555#S4.SS1.p2.1)\.
- \[9\]V\. Dorna, D\. Subhalingam, K\. Kolluru, S\. Tuli, M\. Singh, S\. Singal, N\. A\. Krishnan, and S\. RanuTAGMol: target\-aware gradient\-guided molecule generation\.InICML’24 Workshop ML for Life and Material Science: From Theory to Industry Applications,Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[10\]P\. Ertl and A\. Schuffenhauer\(2009\)Estimation of synthetic accessibility score of drug\-like molecules based on molecular complexity and fragment contributions\.Journal of cheminformatics1\(1\),pp\. 8\.Cited by:[Appendix D](https://arxiv.org/html/2606.00555#A4.p2.3),[§1](https://arxiv.org/html/2606.00555#S1.p2.1)\.
- \[11\]P\. G\. Francoeur, T\. Masuda, J\. Sunseri, A\. Jia, R\. B\. Iovanisci, I\. Snyder, and D\. R\. Koes\(2020\)Three\-dimensional convolutional neural networks and a cross\-docked data set for structure\-based drug design\.Journal of chemical information and modeling60\(9\),pp\. 4200–4215\.Cited by:[Appendix D](https://arxiv.org/html/2606.00555#A4.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p1.1)\.
- \[12\]B\. Gao, Y\. Huang, Y\. Liu, W\. Xie, B\. He, H\. Tan, W\. Ma, Y\. Zhang, and Y\. Lan\(2025\)CIDD: collaborative intelligence for structure\-based drug design empowered by llms\.InThe Thirty\-ninth Annual Conference on Neural Information Processing Systems,Cited by:[Appendix D](https://arxiv.org/html/2606.00555#A4.p1.1),[Appendix D](https://arxiv.org/html/2606.00555#A4.p2.3),[Appendix D](https://arxiv.org/html/2606.00555#A4.p3.2),[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§1](https://arxiv.org/html/2606.00555#S1.p2.1),[§2](https://arxiv.org/html/2606.00555#S2.p2.1),[§3](https://arxiv.org/html/2606.00555#S3.p1.1),[§4\.1](https://arxiv.org/html/2606.00555#S4.SS1.p2.1),[§5](https://arxiv.org/html/2606.00555#S5.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p3.3)\.
- \[13\]J\. Gasteiger and M\. Marsili\(1980\)Iterative partial equalization of orbital electronegativity—a rapid access to atomic charges\.Tetrahedron36\(22\),pp\. 3219–3228\.Cited by:[Appendix B](https://arxiv.org/html/2606.00555#A2.p3.1)\.
- \[14\]A\. K\. Ghose, V\. N\. Viswanadhan, and J\. J\. Wendoloski\(1999\)A knowledge\-based approach in designing combinatorial or medicinal chemistry libraries for drug discovery\. 1\. a qualitative and quantitative characterization of known drug databases\.Journal of combinatorial chemistry1\(1\),pp\. 55–68\.Cited by:[Appendix D](https://arxiv.org/html/2606.00555#A4.p3.2)\.
- \[15\]J\. Guan, W\. W\. Qian, X\. Peng, Y\. Su, J\. Peng, and J\. Ma\(2023\)3D equivariant diffusion for target\-aware molecule generation and affinity prediction\.InThe Eleventh International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=kJqXEPXMsE0)Cited by:[Appendix D](https://arxiv.org/html/2606.00555#A4.p1.1),[Appendix D](https://arxiv.org/html/2606.00555#A4.p2.3),[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[footnote 37](https://arxiv.org/html/2606.00555#footnote37)\.
- \[16\]J\. Guan, X\. Zhou, Y\. Yang, Y\. Bao, J\. Peng, J\. Ma, Q\. Liu, L\. Wang, and Q\. Gu\(2023\)DecompDiff: diffusion models with decomposed priors for structure\-based drug design\.InInternational Conference on Machine Learning,pp\. 11827–11846\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[17\]A\. P\. Guerreiro, C\. M\. Fonseca, and L\. Paquete\(2021\)The hypervolume indicator: computational problems and algorithms\.ACM Computing Surveys \(CSUR\)54\(6\),pp\. 1–42\.Cited by:[§4\.3](https://arxiv.org/html/2606.00555#S4.SS3.p5.3)\.
- \[18\]C\. Harris, K\. Didi, A\. R\. Jamasb, C\. K\. Joshi, S\. V\. Mathis, P\. Lio, and T\. Blundell\(2023\)Benchmarking generated poses: how rational is structure\-based drug design with generative models?\.arXiv preprint arXiv:2308\.07413\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1)\.
- \[19\]X\. Hu, A\. Li, Q\. Xing, J\. Ji, H\. Tuo, and B\. Yang\(2026\)Empowering llms for structure\-based drug design via exploration\-augmented latent inference\.InProceedings of the ACM Web Conference 2026,pp\. 4244–4255\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[20\]Z\. Huang, L\. Yang, X\. Zhou, Z\. Zhang, W\. Zhang, X\. Zheng, J\. Chen, Y\. Wang, B\. Cui, and W\. Yang\(2024\)Protein\-ligand interaction prior for binding\-aware 3d molecule diffusion models\.InThe Twelfth International Conference on Learning Representations,Cited by:[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[21\]L\. Ioakimidis, L\. Thoukydidis, A\. Mirza, S\. Naeem, and J\. Reynisson\(2008\)Benchmarking the reliability of qikprop\. correlation between experimental and predicted values\.QSAR & Combinatorial Science27\(4\),pp\. 445–456\.Cited by:[Appendix D](https://arxiv.org/html/2606.00555#A4.p3.2)\.
- \[22\]C\. Isert, K\. Atz, and G\. Schneider\(2023\)Structure\-based drug design with geometric deep learning\.Current Opinion in Structural Biology79,pp\. 102548\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1)\.
- \[23\]J\. H\. Jensen\(2019\)A graph\-based genetic algorithm and generative model/monte carlo tree search for the exploration of chemical space\.Chemical science10\(12\),pp\. 3567–3572\.Cited by:[§3](https://arxiv.org/html/2606.00555#S3.p1.1)\.
- \[24\]A\. Kadan, K\. Ryczko, E\. Lloyd, A\. Roitberg, and T\. Yamazaki\(2025\)Guided multi\-objective generative ai to enhance structure\-based drug design\.Chemical Science16\(29\),pp\. 13196–13210\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§1](https://arxiv.org/html/2606.00555#S1.p2.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[25\]G\. Landrum\(2013\)Rdkit documentation\.Release1\(1\-79\),pp\. 4\.Cited by:[Appendix C](https://arxiv.org/html/2606.00555#A3.p2.4)\.
- \[26\]C\. A\. Lipinski, F\. Lombardo, B\. W\. Dominy, and P\. J\. Feeney\(2012\)Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings\.Advanced drug delivery reviews64,pp\. 4–17\.Cited by:[Appendix D](https://arxiv.org/html/2606.00555#A4.p3.2)\.
- \[27\]S\. Luo, J\. Guan, J\. Ma, and J\. Peng\(2021\)A 3d generative model for structure\-based drug design\.Advances in Neural Information Processing Systems34,pp\. 6229–6239\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[28\]N\. M\. O’Boyle, M\. Banck, C\. A\. James, C\. Morley, T\. Vandermeersch, and G\. R\. Hutchison\(2011\)Open babel: an open chemical toolbox\.Journal of cheminformatics3\(1\),pp\. 33\.Cited by:[Appendix B](https://arxiv.org/html/2606.00555#A2.p2.1)\.
- \[29\]X\. Peng, S\. Luo, J\. Guan, Q\. Xie, J\. Peng, and J\. Ma\(2022\)Pocket2mol: efficient molecular sampling based on 3d protein pockets\.InInternational conference on machine learning,pp\. 17644–17655\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[30\]A\. T\. Plowright, C\. Johnstone, J\. Kihlberg, J\. Pettersson, G\. Robb, and R\. A\. Thompson\(2012\)Hypothesis driven drug design: improving quality and effectiveness of the design\-make\-test\-analyse cycle\.Drug discovery today17\(1\-2\),pp\. 56–62\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§1](https://arxiv.org/html/2606.00555#S1.p4.1)\.
- \[31\]K\. Qiu, Y\. Song, Z\. Fan, P\. Liu, Z\. Zhang, M\. Zheng, H\. Zhou, and W\. Ma\(2025\)Piloting structure\-based drug design via modality\-specific optimal schedule\.InInternational Conference on Machine Learning,pp\. 50619–50644\.Cited by:[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[32\]K\. Qiu, Y\. Song, J\. Yu, H\. Ma, Z\. Cao, Z\. Zhang, Y\. Wu, M\. Zheng, H\. Zhou, and W\. Ma\(2025\)Empower structure\-based molecule optimization with gradient guided bayesian flow networks\.InInternational Conference on Machine Learning,pp\. 50645–50671\.Cited by:[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[33\]Y\. Qu, K\. Qiu, Y\. Song, J\. Gong, J\. Han, M\. Zheng, H\. Zhou, and W\. Ma\(2024\)MolCRAFT: structure\-based drug design in continuous parameter space\.InInternational Conference on Machine Learning,pp\. 41749–41768\.Cited by:[Appendix D](https://arxiv.org/html/2606.00555#A4.p2.3),[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p2.1),[footnote 37](https://arxiv.org/html/2606.00555#footnote37)\.
- \[34\]N\. Ran, Y\. Wang, and R\. Allmendinger\(2025\)Mollm: multi\-objective large language model for molecular design–optimizing with experts\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p2.1),[§2](https://arxiv.org/html/2606.00555#S2.p2.1),[§3](https://arxiv.org/html/2606.00555#S3.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p3.3)\.
- \[35\]S\. Salentin, S\. Schreiber, V\. J\. Haupt, M\. F\. Adasme, and M\. Schroeder\(2015\)PLIP: fully automated protein–ligand interaction profiler\.Nucleic acids research43\(W1\),pp\. W443–W447\.Cited by:[Appendix B](https://arxiv.org/html/2606.00555#A2.p2.1),[§3](https://arxiv.org/html/2606.00555#S3.p1.1),[§4\.1](https://arxiv.org/html/2606.00555#S4.SS1.p2.1),[§5](https://arxiv.org/html/2606.00555#S5.p3.3)\.
- \[36\]O\. Trott and A\. J\. Olson\(2010\)AutoDock vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading\.Journal of computational chemistry31\(2\),pp\. 455–461\.Cited by:[Appendix B](https://arxiv.org/html/2606.00555#A2.p2.1),[Appendix D](https://arxiv.org/html/2606.00555#A4.p2.3),[§1](https://arxiv.org/html/2606.00555#S1.p2.1)\.
- \[37\]M\. Wang, Z\. Wang, H\. Sun, J\. Wang, C\. Shen, G\. Weng, X\. Chai, H\. Li, D\. Cao, and T\. Hou\(2022\)Deep learning approaches for de novo drug design: an overview\.Current opinion in structural biology72,pp\. 135–144\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1)\.
- \[38\]S\. S\. Wesolowski and D\. G\. Brown\(2016\)The strategies and politics of successful design, make, test, and analyze \(dmta\) cycles in lead generation\.Lead Generation,pp\. 487–512\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§4\.3](https://arxiv.org/html/2606.00555#S4.SS3.p1.1)\.
- \[39\]K\. Wu, Y\. Xia, P\. Deng, R\. Liu, Y\. Zhang, H\. Guo, Y\. Cui, Q\. Pei, L\. Wu, S\. Xie,et al\.\(2024\)TamGen: drug design with target\-aware molecule generation through a chemical language model\.Nature Communications15\(1\),pp\. 9360\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[40\]Z\. Zhang and Q\. Liu\(2023\)Learning subpocket prototypes for generalizable structure\-based drug design\.InInternational Conference on Machine Learning,pp\. 41382–41398\.Cited by:[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.
- \[41\]Z\. Zhang, J\. Yan, Y\. Huang, Q\. Liu, E\. Chen, M\. Wang, and M\. Zitnik\(2025\)Structure\-based drug design with geometric deep learning: a comprehensive survey\.ACM Computing Surveys58\(5\),pp\. 1–35\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p2.1)\.
- \[42\]J\. Zhou, D\. Zhao, H\. Qian, S\. Tu, and L\. Xu\(2025\)Multi\-objective structure\-based drug design using causal discovery\.IEEE Transactions on Computational Biology and Bioinformatics\.Cited by:[§1](https://arxiv.org/html/2606.00555#S1.p1.1),[§1](https://arxiv.org/html/2606.00555#S1.p2.1),[§2](https://arxiv.org/html/2606.00555#S2.p1.1),[§5](https://arxiv.org/html/2606.00555#S5.p2.1)\.

## Appendix APrompts

`Prompt A\.1: MOO\-awareness prompt injected into the design stage of CIDD\. Prompt A\.2: Intent extraction prompt\. Prompt A\.3: Prompt for PROBE in site map construction\. Prompt A\.4: Prompt for PROBE in Probing Planner\. Prompt A\.5: Prompt for PROBE in Probing matrix\. Prompt A\.6: Prompt for PROBE in response summarization\. Prompt A\.7: Prompt for PROBE in EditManual construction\. Prompt A\.8: Prompt for PROBE in Multi\-Agent Optimization\.`

`Appendix B Details of Fragment\-Level Signal Extraction This appendix specifies how the diagnosis module extracts three signal types: \{interaction, geometric, and property\} from a docked protein–ligand complex\. Interaction signals We dock the ligand in the protein pocket with AutoDock Vina \[36\], convert the docked ligand pose to PDB with Open Babel \[28\], and merge it with the protein pocket into a single complex\. Based on this single complex, PLIP \[35\] then enumerates non\-covalent contacts, including hydrogen bonds, hydrophobic contacts, π\\pi\-stacking, salt bridges, halogen bonds, and water bridges; each contact is recorded with the participating ligand atom indices and the protein residue label\. Geometric signals\. Geometric signals are computed from heavy\-atom coordinates only, and hydrogens are excluded\. Each receptor atom is first annotated with four context tags: its element\-specific van der Waals \(VDW\) radius; a polarity flag \(N, O, S, P\); a hydrophobicity flag \(C\); and a formal\-charge class assigned by a residue\-level rule that labels the side\-chain oxygens of ASP/GLU as negative and the side\-chain nitrogens of LYS/ARG/HIS as positive\. A single ligand\-to\-receptor distance matrix is computed once and reused by the four checks below\. \(i\) Steric clash\. A ligand–receptor pair whose distance falls below a fixed fraction of the VDW sum is flagged as a clash\. A softer threshold is used when both atoms are N or O, so that tight polar contacts are not mislabeled\. \(ii\) Solvent exposure\. A ligand heavy atom whose nearest receptor atom is farther than a short\-range cutoff is reported as exposed, meaning it sticks out into solvent rather than engaging the pocket\. \(iii\) Electrostatic repulsion\. We assign Gasteiger partial charges \[13\] to the ligand\. If this step fails, the molecule is topologically invalid \(e\.g\., hypervalent atoms or kekulization failure\); we then abort the remaining geometric checks and mark the molecule with a liability flag so that it is repaired before any interaction\-level optimization\. When charges are available, a ligand atom with an appreciable partial charge that sits close to a same\-sign charged receptor atom is reported as a repulsion\. \(iv\) Buried unsatisfied polar atoms\. For each ligand N or O, we count its short\-range receptor contacts\. The atom is satisfied only if a receptor N or O lies within H\-bond range\. Buried\-but\-unsatisfied atoms are split into hydrophobic burial\(no polar receptor atom nearby\) and geometry mismatch \(polar atoms are nearby, but none within H\-bond range\)\. Property signals Property signals summarize the 2D molecule together with its docking outcome along four axes\. \(i\) Drug\-likeness and ligand efficiency\. We record molecular weight \(MW\), log⁡P\\log P, topological polar surface area \(TPSA\), and QED \[5\], together with the ligand efficiency \[1\]\. \(ii\) Synthetic complexity and topology alerts\. We count chiral centers \(including unassigned ones\) and raise an alert whenever the molecule contains a spiro atom, bridgehead atom, or macrocycle \(ring of more than eight atoms\)\. The Bertz complexity index \[4\] is retained as a continuous proxy for overall topological complexity\. \(iii\) Ring hybridization consistency\. Ring systems are built by merging rings that share at least one atom, yielding the connected components of fused rings\. Exocyclic C=O and C=N carbons are excluded from hybridization counting so that carbonyls and imines do not cause false positives\. Two rules are then enforced: Fused systems rule: a multi\-ring system must contain at least one s​p2sp^\{2\} carbon, otherwise an all\-s​p3sp^\{3\} fused ring alert is raised; Single ring rule: the non\-excluded carbons of a single ring should share one hybridization, and a ring mixing s​p2sp^\{2\} and non\-s​p2sp^\{2\} carbons triggers a mixed\-hybridization alert\. \(iv\) Toxicophore alerts\. We additionally match the molecule against a small set of common structural\-alert SMARTS patterns \[3\], including aniline, Michael acceptor, hydrazine/azo, and aldehyde motifs\. Any match is appended to the alert list returned alongside the numerical signals\. Appendix C Fragment\-Assembly Engine PROBE realizes molecular edits through a fragment\-assembly engine, shared by the probing stage \(Section 4\.2\) and the multi\-agent optimization stage \(Section 4\.3\)\. Sharing one engine ensures that rules accumulated in the EditManual are applied consistently at generation time\. Edit execution\. Each structured edit specifies a target atom in the molecule, an action from a set of fragment\-level operations \{\\\{Delete\_Group, Grow, Replace\_Terminal\_Group, Replace\_Sidechain\_or\_Ring\}\\\}, and a set of semantic constraints 𝒞\\mathcal\{C\} over the desired fragment\. Delete\_Group removes the fragment rooted at the target atom and caps the resulting open valence\. The other three actions attach a new303030✓\*\*\* or replacement? fragment or replace the current fragment at the target atom and therefore require retrieving a fragment that satisfies 𝒞\\mathcal\{C\}\. For these actions, the engine performs the retrieval step described below, then uses RDKit \[25\]313131✓\*\*\* ref to attach the retrieved fragment at the target atom according to the requested action and to sanitize the resulting molecule\. Retrieval in clinically\-grounded fragment library\. For the three retrieval\-based actions, new fragments or replacement323232✓\*\*\* or new? fragments come from a pre\-built library ℒ\\mathcal\{L\} of 8,5058\{,\}505 BRICS fragments \[8\] obtained by decomposing ChEMBL Phase\-1 compounds, providing a prior over fragments that have already survived early ADMET attrition\. Each fragment is precomputed with multiple properties, e\.g\., QED, SA, chiral\-center count, a discrete shape category, and rotatable\-bond count, so that the semantic constraints 𝒞\\mathcal\{C\} in Sections 4\.1\-4\.3 map directly onto a database query\. Given a constraint set 𝒞\\mathcal\{C\}, the engine first restricts ℒ\\mathcal\{L\} to the matching subset ℱmatch\\mathcal\{F\}\_\{\\text\{match\}\} and then ranks the matches by a composite score that biases retrieval toward the best druggability:333333✓removed \*\*\* what’s ”druggability Pareto frontier”? pls, remove all fancy but meaningless words from the paper343434✓because vina cannot be calculated on a single fragment, it is a score between a protein and a molecule\. \*\*\* why no vina? u used that in sec3 ℱ∗=arg​top​\-​Kf∈ℱmatch⁡\[λQED⋅QED​\(f\)\+λSA⋅SA​\(f\)−λchiral⋅Nchiral​\(f\)\]\.\\mathcal\{F\}^\{\*\}\\;=\\;\\operatorname\*\{arg\\,top\\text\{\-\}K\}\_\{f\\in\\mathcal\{F\}\_\{\\text\{match\}\}\}\\Big\[\\,\\lambda\_\{\\text\{QED\}\}\\\!\\cdot\\\!\\mathrm\{QED\}\(f\)\\;\+\\;\\lambda\_\{\\text\{SA\}\}\\\!\\cdot\\\!\\mathrm\{SA\}\(f\)\\;\-\\;\\lambda\_\{\\text\{chiral\}\}\\\!\\cdot\\\!N\_\{\\text\{chiral\}\}\(f\)\\,\\Big\]\. Nchiral​\(f\)N\_\{\\text\{chiral\}\}\(f\) is the number of chiral center of the fragment ff\.353535✓\*\*\* what’s Nchiral​\(f\)N\_\{\\text\{chiral\}\}\(f\)? In the experiments, we set λQED=λSA=0\.5,λchiral=0\.15\\lambda\_\{\\text\{QED\}\}=\\lambda\_\{\\text\{SA\}\}=0\.5,\\lambda\_\{\\text\{chiral\}\}=0\.15\. The top\-ranked363636✓it should be ’top\-ranked’ here\. \*\*\* u only need the top\-ranked? then why u hv top\-K above? fragment f∈ℱ∗f\\in\\mathcal\{F\}^\{\*\} is then passed to the executor\. Appendix D Dataset and Evaluation Metrics Dataset\. Following CIDD \[12\], we conduct experiments on the CrossDocked2020 dataset \[11\]\. To ensure a fair comparison, we adopt the classic data splitting strategy proposed by TargetDiff \[15\], yielding a test set of 100 protein pockets\. Evaluation Metrics\. We evaluate models using the following standard SBDD metrics\. \(i\) Vina docking score \(Vina\) \[36\], which measures binding affinity\. Following CIDD \[12\], since our pipeline does not output 3D conformations, we report only Vina Dock \(re\-docking based\) and exclude Vina Score and Vina Min\. \(ii\) QED \[5\], which is for drug\-likeness, \(iii\) SA score \[10\], which is for synthetic accessibility\. \(iv\) Success ratio \[15, 33\],373737✓\*\*\* refs\. should be \[15, 33\]? defined as the percentage of molecules satisfying all the following criteria as in \[15, 33\]: Vina docking score <−8\.18<\-8\.18, QED \>0\.25\>0\.25, and SA \>0\.59\>0\.59\. We also report \(v\) Molecular Reasonability Ratio \(MRR\) \[12\], which captures rule\-based structural plausibility\. \(vi\) QikProp pass ratio \[21\], which assesses a wide range of physicochemical and pharmacokinetic properties to predict molecular drug\-likeness properties\. \(vii\) Lipinski \[26\], which calculates the number of rules the molecule obeys in Lipinski’s rule of five\. \(viii\) LogP,383838✓I think LogP and MW don’t need ref, they are very well\-known terms in molecule research areas\. \*\*\* ref which is the octanol\-water partition coefficient, with values between −0\.4\-0\.4 and 5\.65\.6 are considered favorable for drug candidates \[14\]\. \(ix\) MW,393939✓I think LogP and MW don’t need ref, they are very well\-known terms in molecule research areas\.\*\*\* ref which is the molecular weight\. \(x\) Rank\. For all methods, we rank them based on their Success ratio\. Appendix E Effect of the LLM Backbone Our main results in Table 4 use GPT\-4o\-mini as the backbone for all agent roles\. To verify that the gains of PROBE are not tied to a specific LLM, we re\-run PROBE and the strongest LLM\-agent baseline, CIDD, on MolCRAFT initial molecules with two additional LLM backbones, GPT\-4o and Gemini\-3\-flash\. Results are reported in Table 6\. Table 6: Effect of the LLM backbone\. PROBE and CIDD are evaluated on MolCRAFT seeds with three different LLMs\. Per\-column best and second\-best are color\-coded\. PROBE outperforms CIDD under every backbone\. Across all three backbones, PROBE outperforms CIDD on all the metrics with a wide margin\. The gap is preserved when both methods are changed to use a stronger LLM, indicating that the advantage stems from the probes before editing pipeline with site map and EditManual, rather than from any particular capacity of the backbone\. Consistent gains from stronger backbones\. Moving from GPT\-4o\-mini to GPT\-4o, and further to Gemini\-3\-flash, PROBE yields a mild but consistent improvement on the affinity and druggability, which is consistent with the intuition that stronger LLMs help refine pattern interpretation and constraint synthesis but do not alter the underlying evidence on which the agent reasons\. Appendix F Limitations We discuss several limitations of PROBE from two aspects\. First, the evaluation relies on in silico proxies\. Following standard practice in SBDD, we assess affinity with Vina and druggability with QED and SA; these are widely adopted surrogates but cannot fully substitute for wet\-lab validation, and prospective experimental confirmation is left to future work\. Second, the benchmark coverage is limited to CrossDocked2020\. While it is the de facto benchmark for SBDD, it covers a finite set of pocket families, and extending PROBE to broader target classes \(e\.g\., membrane proteins, protein–protein interfaces\) is an interesting direction\. Appendix G Broader Impacts Positive impacts\. PROBE makes LLM\-agent\-based ligand optimization more effective, which can lower the cost and time of early\-stage drug discovery\. This is useful for under\-resourced areas such as rare and neglected diseases, where large\-scale screening is often infeasible\. Better in silico tools also reduce the number of compounds that need to be synthesized and tested, cutting chemical waste in early discovery\. Negative impacts\. Like other molecular generative models, PROBE could, in principle, be misused to design harmful compounds\. We view this risk as limited: PROBE only optimizes ligands for a user\-specified pocket and does not choose targets on its own, and its outputs are computational candidates that cannot have a real\-world effect without synthesis\.`