MolReasoner: Interpretable Molecular Reasoning

• MolReasoner is a two-stage framework that transitions molecular language models from rote memorization to explicit, interpretable chemical reasoning.
• The first stage uses synthetic chain-of-thought fine-tuning with GPT-4o to provide reasoning priors, while the second employs reinforcement learning with chemical reward functions.
• Empirical evaluations demonstrate significant gains in captioning and generation metrics, confirming enhanced interpretability and structural fidelity in outputs.

MolReasoner is a specialized two-stage training framework for molecular LLMs that aims to push them beyond “memorizing” molecule–text pairs toward genuine, interpretable chemical reasoning. It is designed for two tasks—molecule captioning and text-based de novo molecule generation—and combines synthetic Chain-of-Thought (CoT) supervision with reinforcement learning under chemically aware reward functions. The framework is presented as a way to transition LLMs from memorization toward chemical reasoning, while improving interpretability, molecular understanding, and generalization (Zhao et al., 4 Aug 2025).

1. Problem formulation and design rationale

MolReasoner is motivated by the claim that molecular reasoning requires more than surface-level pattern matching: it demands an understanding of atoms, bonds, functional groups, and their semantic interplay. In the formulation of the framework, prior “prompt-only” methods rely on general-purpose LLMs such as GPT-4o with hand-crafted instructions or a few in-context examples. These models may be fluent, but they frequently hallucinate chemically invalid structures or miscount atoms because they do not internalize chemistry-specific reasoning patterns. By contrast, straightforward fine-tuning on molecule–text pairs tends to produce models that memorize end-to-end mappings without explicit intermediate logic, which limits generalization to novel scaffolds and offers little interpretability (Zhao et al., 4 Aug 2025).

The framework is organized as a two-stage solution. Mol-SFT “warm-starts” explicit reasoning by distilling synthetic CoT examples from GPT-4o and verifying their chemical correctness. Mol-RL then refines that reasoning ability through reinforcement learning with specialized reward functions designed explicitly to align chemical structures with linguistic descriptions at multiple levels of granularity. This sequencing is presented as a transition from rote recall to robust, interpretable molecular reasoning (Zhao et al., 4 Aug 2025).

A central conceptual distinction in MolReasoner is between valid output and reasoned output. The framework does not treat validity alone as sufficient. Instead, it operationalizes molecular reasoning as the ability to produce intermediate, human-readable steps that connect structural features, functional groups, and linguistic descriptions to the final caption or SELFIES output. This suggests that interpretability is treated as an intrinsic training target rather than a post hoc diagnostic.

2. Mol-SFT: synthetic Chain-of-Thought fine-tuning

The first stage, Mol-SFT, is intended to teach the model how to “think aloud” in a style that human chemists would recognize. Because manually producing large numbers of molecular rationales is prohibitively expensive, the framework uses GPT-4o as a synthetic oracle. Starting from the ChEBI-20 training set, described as approximately $42\text{ K}$ molecule–text pairs, the prompting procedure injects chemistry-rich information into the prompt template, including statistical structural features such as ring count, aromaticity, and molecular weight, together with enumerated key functional groups via the EFGs method. GPT-4o then produces stepwise reasoning chains aligned to molecule captioning and text-based de novo molecule generation (Zhao et al., 4 Aug 2025).

The synthetic CoT outputs are filtered by three checks. The first is semantic consistency, meaning that the distilled natural-language caption matches the ground-truth description. The second is validity of SELFIES-encoded molecules. The third is structural fidelity: for generation, the SELFIES must be parsable to the reference SMILES and must match on fingerprint and fragment metrics. After filtering, the framework retains approximately $18\text{ K}$ captioning examples and $24\text{ K}$ generation examples (Zhao et al., 4 Aug 2025).

The fine-tuning objective is standard autoregressive maximum likelihood over the concatenated instruction $x$ and CoT trajectory $y=(y_1,\dots,y_T)$:

$$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$

The base model is Qwen2.5-7B-Instruct. In the description of the framework, this objective teaches both the preferred CoT format and the domain-specific vocabulary and logical steps that underlie chemical reasoning (Zhao et al., 4 Aug 2025).

Component	Specification	Role
Base LLM	Qwen2.5-7B-Instruct	Molecular reasoning backbone
Source dataset	ChEBI-20 training set ($\sim42\text{ K}$ pairs)	Synthetic CoT generation
Filtered CoT data	$\sim18\text{ K}$ captioning, $\sim24\text{ K}$ generation	Mol-SFT supervision

3. Mol-RL: reinforcement learning with molecular reward functions

After Mol-SFT establishes shallow reasoning priors, MolReasoner switches to an on-policy reinforcement-learning stage, Mol-RL, to deepen chemical understanding through outcome-based feedback. The optimization method is Group Relative Policy Optimization (GRPO). For each input $q$, GRPO samples $18\text{ K}$0 candidate outputs $18\text{ K}$1 under the old policy $18\text{ K}$2, computes a reward $18\text{ K}$3 for each candidate, and normalizes the reward within the group as

$18\text{ K}$4

The GRPO objective clips importance weights and adds a KL penalty to a fixed reference policy $18\text{ K}$5:

$18\text{ K}$6

For molecule captioning, the reward is

$18\text{ K}$7

where $18\text{ K}$8 is the mean of BLEU-2/4, METEOR, and ROUGE-1/2/L. For text-based de novo generation, the reward is

$18\text{ K}$9

with

$24\text{ K}$0

The structural terms are further specified as

$24\text{ K}$1

$24\text{ K}$2

$24\text{ K}$3

$24\text{ K}$4

The policy is optimized by stochastic gradient ascent on $24\text{ K}$5. The RL stage is reported to run for $24\text{ K}$6 epochs with $24\text{ K}$7 rollouts per input and a small learning rate $24\text{ K}$8 (Zhao et al., 4 Aug 2025).

4. Interpretability mechanisms

MolReasoner identifies two explicit mechanisms for transparency. The first is explicit CoT steps. Every output is prefaced with numbered, human-readable reasoning steps, such as “1. Identify core scaffold…” through “n. Final SELFIES answer.” This format is intended to let users audit how the model derived each atom or functional group. The second mechanism is structured attention. During Mol-RL, attention weights are described as tending to concentrate on tokens describing structural features, including ring counts and fragment tags, when the model generates matching SELFIES or text tokens. These attention maps are used to show which part of the prompt influenced each reasoning step (Zhao et al., 4 Aug 2025).

The interpretability claim in MolReasoner is therefore narrower than a guarantee of mechanistic faithfulness. The framework provides explicit CoT text and attention visualizations, both of which are inspectable. This suggests that its notion of interpretability is grounded in traceable generation rather than in a formal proof that the latent computation is chemically faithful in all cases.

A common misconception is to equate chemically valid outputs with interpretable reasoning. MolReasoner is specifically framed against that reduction. Its stated objective is not only to generate valid molecules or fluent captions, but also to expose chemically grounded thought chains that users can inspect.

5. Empirical evaluation and ablation evidence

The experimental setup uses warm-up CoT data from ChEBI-20, the same ChEBI-20 splits for GRPO training, and a held-out ChEBI-20 test set in SELFIES representation. For captioning, the reported metrics are BLEU-2/4, METEOR, and ROUGE-1/2/L. For generation, the metrics are BLEU, exact match, Levenshtein distance, Tanimoto-based fingerprint similarities for Morgan, MACCS, and RDKit, Frag-J, Frag-R, FG-Match, and overall validity (Zhao et al., 4 Aug 2025).

In molecule captioning, MolReasoner (7B) reports BLEU-2 $24\text{ K}$9, BLEU-4 $x$0, METEOR $x$1, and ROUGE-1 $x$2. The paper states that these scores far surpass GPT-4o $x$3 and open-source baselines, with the best open-source BLEU-2 around $x$4 and BLEU-4 around $x$5. In text-based de novo generation, MolReasoner reports BLEU $x$6, exact $x$7, Tanimoto-mean $x$8, MACCS $x$9, Morgan $y=(y_1,\dots,y_T)$0, Frag-J $y=(y_1,\dots,y_T)$1, Frag-R $y=(y_1,\dots,y_T)$2, FG-Match $y=(y_1,\dots,y_T)$3, and validity $y=(y_1,\dots,y_T)$4. The comparison given in the paper includes GPT-4o with BLEU $y=(y_1,\dots,y_T)$5 and validity $y=(y_1,\dots,y_T)$6, and Mol-Instruction with BLEU $y=(y_1,\dots,y_T)$7, validity approximately $y=(y_1,\dots,y_T)$8, but poorer similarity scores (Zhao et al., 4 Aug 2025).

Ablation studies are used to separate the effects of warm-up and reward shaping. For captioning, warm-up alone lifts BLEU-2 from $y=(y_1,\dots,y_T)$9 to $$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$0, and adding format accuracy and language-similarity rewards brings BLEU-2 to $$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$1. For generation, warm-up alone raises BLEU-2 to $$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$2 and validity to $$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$3; adding fingerprint, fragment, and FG-match rewards increases BLEU-2 to $$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$4 and Tanimoto similarity from $$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$5 to $$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$6 (Zhao et al., 4 Aug 2025).

These ablations are important because they localize the contribution of each stage. The evidence is presented as showing that Mol-SFT establishes a strong reasoning prior, while Mol-RL sharpens structural and semantic alignment rather than merely improving formatting.

6. Limitations, scope, and relation to adjacent molecular reasoning systems

The framework identifies three limitations and future directions. First, synthetic CoT depends on GPT-4o’s accuracy and may inherit its biases; calibrating chain confidence remains open. Second, the current rewards focus on 2D structural alignment; future work is proposed to incorporate synthetic accessibility, ADMET properties, or 3D conformational feasibility. Third, the two-stage pipeline, especially on-policy GRPO, is computationally demanding, with approximately $$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$7 GPU-hours for RL; more efficient RL algorithms or off-policy methods are suggested for scalability to larger models and molecule libraries (Zhao et al., 4 Aug 2025).

These limitations clarify the scope of MolReasoner. It is a framework for molecular reasoning on captioning and text-based generation under structure-aware rewards, not a complete solution to medicinal chemistry optimization or assay prediction. A plausible implication is that its strongest claims concern interpretable generation and structure–text alignment rather than downstream pharmacology.

The broader research landscape contains related but distinct systems. Bolek injects a radius-2 Morgan fingerprint of length $$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$8 bits into a Qwen3 decoder as a dedicated <molecule> token, and it is trained on molecular alignment tasks and downstream reasoning over $$L_{\mathrm{Mol\text{-}SFT}}(\theta) = -\sum_{t=1}^{T} \log p(y_t \mid x,y_{<t};\theta).$$9 TDC binary classification tasks using synthetic CoTs anchored in concrete molecular features; its auditability is tied to verifiable descriptor citation and agreement with RDKit (Grabowski et al., 4 May 2026). “A Multi-task Large Reasoning Model for Molecular Science” describes a multi-specialist layer, a reasoning controller, curated CoT data, and reinforcement learning infused with molecular knowledge across $\sim42\text{ K}$0 molecular tasks and $\sim42\text{ K}$1 metrics, emphasizing structured and reflective reasoning (Liu et al., 13 Mar 2026). DrugR treats molecular optimization as a reasoning-driven, multi-objective generation problem and combines domain-specific continual pretraining, supervised fine-tuning via reverse data engineering, and self-balanced multi-granular reinforcement learning to improve ADMET properties while preserving scaffold similarity and docking affinity (Liu et al., 9 Feb 2026).

Taken together, these systems indicate a broader shift from purely data-driven prediction or prompt-only molecular language modeling toward architectures that couple explicit reasoning traces with verifiable molecular signals. Within that trajectory, MolReasoner is specifically characterized by synthetic CoT warm-starting, chemically aware GRPO rewards, and outputs accompanied by transparent, chemically grounded thought chains (Zhao et al., 4 Aug 2025).