AttriLens-Mol: RL-based Molecular Prediction

• AttriLens-Mol is an RL framework that directs predictions by reasoning through a curated set of molecular attributes derived from SMILES strings.
• The framework enforces a strict XML-like output schema with rewards for format, correctness, count, and rationality to ensure interpretable and coherent predictions.
• It leverages state-of-the-art LLM backbones, RDKit descriptors, and GRPO optimization to outperform traditional prompting and chain-of-thought methods on both classification and regression tasks.

Searching arXiv for the cited papers and closely related work to ground the article. arxiv_search({"query":"AttriLens-Mol attribute guided reinforcement learning molecular property prediction LLMs", "max_results": 5}) Searching arXiv for AttriLens-Mol and related molecular reasoning papers. {"query":"AttriLens-Mol attribute-guided reinforcement learning molecular property prediction with LLMs", "max_results": 10} AttriLens-Mol is an attribute-guided reinforcement learning framework for molecular property prediction with LLMs, in which a molecule is given as a SMILES string and the model is trained to predict a target property by reasoning through relevant molecular attributes before emitting a final answer. The framework is designed for both classification and regression settings, including blood–brain barrier permeability, BACE inhibition, toxicity, side effects, and solubility-related values. Its central claim is that steering a policy model toward a small, relevant set of attributes yields predictions that are both more effective and more interpretable than direct prompting or generic chain-of-thought prompting. The method operationalizes this idea with a structured XML-like output schema and a reinforcement-learning objective built from format, correctness, count, and rationality rewards (Lin et al., 6 Aug 2025).

1. Problem formulation and conceptual basis

AttriLens-Mol is situated in the emerging use of LLMs for molecular property prediction, but it explicitly departs from approaches that depend on human-crafted prompting, in-context demonstrations, or manually authored chain-of-thought templates. The paper identifies three limitations in such prior methods: prompt quality is fragile and can introduce recency bias, performance instability, and poor transfer across tasks; generated reasoning traces do not necessarily correlate with correct predictions; and large reasoning models such as DeepSeek-R1 may exhibit an “overthinking” tendency in which long-form reasoning becomes verbose and irrelevant rather than chemically useful (Lin et al., 6 Aug 2025).

The framework therefore replaces externally scripted reasoning with RL-based self-exploration under verifiable rewards. In the paper’s terminology, the model is encouraged to reason through attributes, meaning chemically meaningful molecular properties or descriptors, typically drawn from RDKit-computable molecular descriptors. All 53 RDKit-computable molecular attributes are considered in the rationality verification pipeline. The required reasoning pattern is explicit: the model must identify relevant attributes, list them, state whether each promotes or inhibits the target property, and only then produce the final answer. This design is intended to elicit the model’s inherent knowledge of relevant molecular attributes rather than merely imitating an expert-written reasoning script (Lin et al., 6 Aug 2025).

A plausible implication is that AttriLens-Mol treats explanation not as a post hoc narrative attached to a prediction, but as an intermediate representation that the policy is optimized to produce. In that respect, its interpretability claim is narrower than full mechanistic attribution but stronger than unconstrained natural-language rationale generation.

2. Framework architecture and output schema

The base policies are two distilled reasoning backbones from DeepSeek-R1: R1-Distilled-Qwen2.5 (7B) and R1-Distilled-LLaMA3.1 (8B). The input consists of a system-style instruction and a user query specifying the task type (classification or regression), the molecule’s SMILES, and the target property. The output is required to follow a strict three-part schema with > ... for reasoning, <name>...</name> for the attribute list, and <answer>...</answer> for the final prediction (Lin et al., 6 Aug 2025).

The paper gives the template in explicit form. The system instruction requires “Step-by-step reasoning with consideration on relevant attributes can be calculated using RDKit,” asks that “for each attribute” the model provide its estimated value and explain whether it promotes (improve) or inhibits (not improve) the target property, and requires the <name> block to list attributes followed by “: promotes” or “: inhibits”. This schema creates three machine-checkable zones: a long-form reasoning trace, a structured attribute summary, and an answer field.

A rollout trajectory is therefore the full model response under this schema. The paper’s policy notation is the current LLM $\pi_\theta$, and for a query $q$ the old policy $\pi_{\theta_{\text{old}}}$ samples a group of candidate outputs. The optimizer is mainly GRPO (Group Relative Policy Optimization), with DAPO also evaluated as a variant. The training setup reported in the paper uses TRL, batch size 8, around 1000 RL steps, 8 candidate responses per input query, temperature 0.6, 4 × RTX L40 GPUs (48 GB each), and approximately 5 hours of training time (Lin et al., 6 Aug 2025).

The architecture is minimal in the sense that it does not add a separate neural module for chemistry-specific representation learning. The chemical grounding is instead pushed into the output format and the reward verifier. This distinguishes AttriLens-Mol from methods that encode molecules through graph neural operators or 3D-aware representations.

3. Reward design and reinforcement-learning objective

The technical core of AttriLens-Mol is its reward design. Two rewards are inherited from DeepSeek-R1-style RLVR, while two are introduced as attribute-specific controls. The format reward enforces the XML-like output structure:

$$\mathcal{R}^{\text{format}} = \begin{cases} 1, & \text{if the format is correct},\ -2, & \text{if the format is incorrect}. \end{cases}$$

The correctness reward ensures that the policy still solves the target task:

$$\mathcal{R}^{\text{correct}} = \begin{cases} 2, & \text{if the answer is correct},\ 0, & \text{if the answer is incorrect}. \end{cases}$$

The count reward is introduced to control overthinking by constraining the number of generated attributes to a moderate range:

$$\mathcal{R}^{\text{count}} = \begin{cases} 0, & \text{if } n_{\text{att}} \in [3,10],\ -1, & \text{otherwise}. \end{cases}$$

The rationality reward is the novel verifier. It compares the model’s promote/inhibit labels with verifier-derived labels obtained by combining fuzzy matching, RDKit descriptor computation, and advantageous value ranges supplied by advanced LLMs such as GPT-4o or DeepSeek-R1:

$$r = \{r_1, r_2, \ldots, r_{|\mathcal{A}|}\}, \qquad \hat{r} = \{\hat{r}_1, \hat{r}_2, \ldots, \hat{r}_{|\mathcal{A}|}\},$$

$\mathcal{R}^{\text{rational}} = \frac{1}{|\mathcal{A}|}\sum_{i=1}^{|\mathcal{A}|} \mathds{1}\{r_i=\hat{r}_i\}.$

Here, $r_i \in \{0,1\}$ is extracted from the model response, with $1/0$ corresponding to promotes or inhibits, and $q$0 indicates whether the RDKit-computed descriptor value is within or out of the advantageous range for the target property. The reward therefore lies in $q$1 and encourages the model to mention attributes that are both descriptor-groundable and directionally coherent (Lin et al., 6 Aug 2025).

The reinforcement-learning objective follows GRPO. For a query $q$2, given a group $q$3 sampled from the old policy and scalar rewards $q$4, the group-normalized advantage is

$q$5

The clipped GRPO objective is written as

$q$6

The paper also gives the KL approximation

$q$7

An important implementation detail is that the paper does not specify the exact scalar combination formula for the total reward, nor common RL hyperparameters such as learning rate, optimizer type, context length, max generation length, top-p, exact KL coefficient $q$8, or clip parameter $q$9 (Lin et al., 6 Aug 2025).

4. Data, tasks, and empirical performance

AttriLens-Mol is trained on 4,023 structured training samples drawn from the MoleculeNet training sets of BBBP, BACE, and ClinTox. These are the in-distribution tasks: BBBP has 1,631 / 204 / 204 train/valid/test samples, BACE has 1,210 / 151 / 152, and ClinTox has 1,182 / 148 / 148. The out-of-distribution evaluation uses tasks whose training sets are not used in RL training: SIDER with 1,141 / 143 / 143, ESOL with 902 / 113 / 113, and FreeSolv with 513 / 64 / 65. The table in the paper states that these are scaffold splits from MoleculeNet. Accuracy is used for BBBP, BACE, ClinTox, and SIDER, while RMSE is used for FreeSolv and ESOL (Lin et al., 6 Aug 2025).

The strongest reported AttriLens-Mol variants are heterogeneous across backbone and optimizer. R1-Distilled-Qwen2.5 (Ours, DAPO) achieves 56.9 on BBBP, 62.6 on BACE, 91.2 on ClinTox, 54.4 on SIDER, 7.82 on FreeSolv, 2.67 on ESOL, with Avg$\pi_{\theta_{\text{old}}}$0 66.3 and Avg$\pi_{\theta_{\text{old}}}$1 5.25. R1-Distilled-Qwen2.5 (Ours, GRPO) achieves 58.1, 60.8, 88.5, 56.9, 11.12, and 1.84, with Avg$\pi_{\theta_{\text{old}}}$2 66.1 and Avg$\pi_{\theta_{\text{old}}}$3 6.48. R1-Distilled-LLaMA3.1 (Ours, GRPO) gives the strongest classification aggregate with Avg$\pi_{\theta_{\text{old}}}$4 67.9, while R1-Distilled-LLaMA3.1 (Ours, DAPO) reaches ClinTox 93.7 but a weaker regression aggregate (Lin et al., 6 Aug 2025).

The improvements over the untuned R1-distilled bases are large. For R1-Distilled-Qwen2.5, BBBP improves from 51.8 to 56.9 / 58.1, BACE from 49.1 to 60.8 / 62.6, ClinTox from 34.0 to 88.5 / 91.2, SIDER from 48.9 to 54.4 / 56.9, FreeSolv from 9.36 to 7.82, and ESOL from 7.96 to 1.84 / 2.67. For R1-Distilled-LLaMA3.1, BBBP improves from 50.6 to 53.4 / 57.2, BACE from 55.7 to 58.6 / 67.6, ClinTox from 33.6 to 87.8 / 93.7, SIDER from 52.4 to 52.9 / 58.9, FreeSolv from 47.53 to 15.48 / 18.44, and ESOL from 5.06 to 3.26 / 7.87 (Lin et al., 6 Aug 2025).

The paper also positions AttriLens-Mol against prompting, CoT, supervised fine-tuning, and frontier prompted models. With Qwen2.5, direct prompting obtains Avg$\pi_{\theta_{\text{old}}}$5 42.8 and Avg$\pi_{\theta_{\text{old}}}$6 74.76, CoT prompting improves to 48.9 and 28.90, and AttriLens-Mol DAPO reaches 66.3 and 5.25. Against large prompted models, DeepSeek-R1 (671B) achieves Avg$\pi_{\theta_{\text{old}}}$7 58.6 and Avg$\pi_{\theta_{\text{old}}}$8 4.35, GPT-4o + CoT gives 54.8 and 5.11, and GPT-4o gives 53.8 and 7.13. The paper’s interpretation is that AttriLens-Mol clearly surpasses these baselines on classification, while remaining competitive on regression despite using no regression-specific reward in training (Lin et al., 6 Aug 2025).

A methodological point emphasized in the paper is that SFT baselines and AttriLens-Mol do not explicitly use the corresponding OOD training sets, whereas task-specific models such as ChemBERTa, MolBERT, and SPMM do use their standard training data. This is part of the paper’s argument for transfer via attribute-guided RL rather than task-specific supervision.

5. Interpretability, attribute extraction, and analysis

The interpretability claim of AttriLens-Mol is centered on the proposition that the generated attributes are not merely decorative rationales but predictive intermediate variables. The paper operationalizes this by extracting the top ten attributes from training samples for each task, computing their values with RDKit, and using these values as features for an interpretable decision-tree-style model, with the text and table specifically referring to random forest and evaluation by AUC-ROC (Lin et al., 6 Aug 2025).

The reported results show that AttriLens-Mol-derived attributes are more useful than attributes from prompting alone. In Table 6, R1-Distilled-Qwen2.5 (Ours, GRPO) + DT yields BBBP 0.7183, BACE 0.7982, and ClinTox 0.8330, while the base R1-Distilled-Qwen2.5 attributes + DT yields 0.6794, 0.7799, and 0.7232. The GRPO version is marked statistically significant relative to the base model with $\pi_{\theta_{\text{old}}}$9. The same table includes comparison points such as GCN at 0.6780 / 0.6893 / 0.8390, GIN at 0.6971 / 0.7346 / 0.8420, GraphMVP at 0.6780 / 0.7430 / 0.7900, and LLM4SD at 0.7135 / 0.7618 / 0.5000. The paper’s interpretation is that AttriLens-Mol extracts attributes that are more predictive, relevant, and more readily transferred into classical interpretable models (Lin et al., 6 Aug 2025).

The ablation studies further specify what makes the attribute-centric policy work. For R1-Distilled-Qwen2.5 under GRPO, the full model gives 58.1 / 60.8 / 88.5 / 56.9 / 11.12 / 1.84 on BBBP, BACE, ClinTox, SIDER, FreeSolv, and ESOL. Removing $$\mathcal{R}^{\text{format}} = \begin{cases} 1, & \text{if the format is correct},\ -2, & \text{if the format is incorrect}. \end{cases}$$0 drops performance to 55.7 / 52.4 / 81.7 / 53.8 / 12.32 / 7.16. Removing $$\mathcal{R}^{\text{format}} = \begin{cases} 1, & \text{if the format is correct},\ -2, & \text{if the format is incorrect}. \end{cases}$$1 gives 50.0 / 42.6 / 82.1 / 49.7 / 11.84 / 5.98. Removing both gives 51.0 / 51.4 / 78.9 / 47.6 / 15.79 / 7.80. Under DAPO, the drop from removing count reward is especially severe on BACE, from 62.6 to 30.9. The paper therefore argues that the count reward is crucial for preventing the model from drifting into irrelevant attribute sprawl, while the rationality reward improves both coherence and predictive quality (Lin et al., 6 Aug 2025).

The comparison between attribute CoT and attribute RL is particularly important because it isolates the benefit of training from the benefit of prompting. For R1-Distilled-Qwen2.5, Attr. CoT gives 55.2 / 59.6 / 88.4 / 53.6 / 8.52 / 4.65, whereas Attr. RL gives 56.9 / 62.6 / 91.2 / 54.4 / 7.82 / 2.67. For R1-Distilled-LLaMA3.1, Attr. CoT gives 52.5 / 55.8 / 90.4 / 50.3 / 20.72 / 9.19, while Attr. RL gives 53.4 / 58.6 / 93.7 / 52.9 / 15.48 / 7.87. This supports the narrower claim that the gain is not simply due to adding an attribute prompt, but to training the model to internalize attribute-guided reasoning (Lin et al., 6 Aug 2025).

Figure 1 is also used to argue that AttriLens-Mol uses significantly fewer tokens during inference while achieving high accuracy across classification tasks. This suggests that the method is not only structured but also more concise than unconstrained long-form reasoning. The paper’s interpretability is therefore best characterized as attribute-centric and verifier-grounded, rather than mechanistic in the sense of token-level or atom-level attribution.

6. Relation to neighboring molecular reasoning systems, limitations, and scope

AttriLens-Mol occupies a distinct position relative to neighboring molecular learning systems. Compared with Mol-PECO, which is a graph-based QSOR model for predicting 118 odor descriptors from molecular structure, AttriLens-Mol does not attempt to learn a dense molecular representation from a Coulomb matrix or Laplacian eigenfunction positional encodings. Mol-PECO integrates a fully connected weighted graph derived from a Coulomb matrix, Laplacian-eigenfunction positional encodings, and graph convolution to predict a multi-label odor vector, reaching AUROC 0.813 and AUPRC 0.181 for Mol-PECO-asym. Its interpretability is primarily embedding-level, including t-SNE odor-space structure and nearest-neighbor retrieval, rather than explicit attribute-centric explanation (Zhang et al., 2023). This suggests that AttriLens-Mol and Mol-PECO address different explanatory strata: the former makes reasoning explicit in textual attribute space, whereas the latter makes structure–perception regularities visible in learned embedding space.

Compared with MT-Mol, AttriLens-Mol shares an emphasis on chemically grounded reasoning but differs in task formulation and system architecture. MT-Mol is a multi-agent framework for molecular optimization rather than property prediction, using five analyst agents, a scientist, a verifier, and a reviewer, together with 154 chemistry-related functions grouped into five tool domains. Its output is an auditable optimization loop with tool-aligned reasoning, reasoning–structure consistency checks, and reviewer feedback, and the paper reports state-of-the-art performance of the PMO-1K benchmark on 17 out of 23 tasks (Kim et al., 27 May 2025). AttriLens-Mol is simpler: it retains a single policy model and outsources verification to reward computation rather than multi-agent deliberation. A plausible implication is that MT-Mol offers a stronger scaffold for iterative design feedback, while AttriLens-Mol offers a more direct framework for learning attribute-centric prediction policies.

The limitations of AttriLens-Mol are explicit or readily apparent from the method description. The rationality reward depends on fuzzy matching between free-text attributes and standardized RDKit descriptors, and on LLM-generated advantageous ranges supplied by GPT-4o or DeepSeek-R1; if the mappings or ranges are noisy, the verifier can be imperfect. The method is built around RDKit-computable descriptors and a binary promote/inhibit schema, which may not capture properties requiring richer mechanistic or 3D reasoning. The RL setup, although modest relative to frontier-scale RL, still requires rollouts, external descriptor computation, verifier logic, and multiple GPUs. Training correctness is explicitly classification-based, so the improvements on ESOL and FreeSolv are transfer effects rather than direct regression optimization. The paper also limits training to three binary classification tasks and does not test domains such as reaction prediction, generative design, 3D conformer-informed tasks, or protein-ligand multimodal settings (Lin et al., 6 Aug 2025).

Within those limits, AttriLens-Mol’s main contribution is to show that for molecular property prediction with LLMs, a reasoning policy can be made more effective by rewarding structured attribute-based outputs, count control, and chemistry-grounded rationality verification. Its reported empirical pattern—strong gains over direct prompting, CoT prompting, and several supervised fine-tuning baselines with only 4,023 RL training samples and 7B/8B models—supports the paper’s narrower thesis that the most useful reasoning in this setting is not longer reasoning, but better-directed reasoning around relevant molecular attributes (Lin et al., 6 Aug 2025).