Back to Basics: Improving Molecular Understanding in LLMs via SMILES-Graph Translation

Technical Summary: Back to Basics: Improving Molecular Understanding in LLMs via SMILES–Graph Translation

Problem Statement
Despite the strong performance of Large Language Models (LLMs) on molecular understanding and generation tasks, current approaches suffer from a lack of reliable structural grounding. The paper identifies a critical disconnect: while models achieve downstream success, they fail at basic structure recognition tasks, such as counting heavy atoms or bonds from canonical SMILES strings. This contradicts the fundamental chemical principle that "structure determines function." Preliminary evaluations reveal that both generalist LLMs (e.g., GPT, Qwen) and specialist molecular LLMs (e.g., MolT5, ChemLLM) perform poorly on these foundational tasks, suggesting they fail to capture the explicit equivalence between sequential SMILES text and the underlying 2D molecular graph. The authors argue that current paradigms, which often jump to high-level reasoning after simple text pre-training, lack logical interpretability and fail to establish the necessary structural cognition required for reliable reasoning.

Methodology: The MolBasic Framework
To address this, the authors propose MolBasic, a "structure-first" framework designed to ground molecular understanding in fundamental structure comprehension. The methodology consists of three core components:

1. Multi-level Structure Comprehension Benchmark (MSC):
   The authors construct a benchmark comprising eight tasks organized across three levels to evaluate and train 2D graph perception:

   • Node Level: Heavy atom counting and element-specific counting.
   • Edge Level: Total bond counting and specific bond type counting.
   • Graph Level: Molecular formula conversion, substructure recognition, and bidirectional conversion between canonical SMILES and textual graph adjacency lists.
     The core task is the bidirectional SMILES–Graph conversion, which forces the model to learn the equivalence between sequential and topological representations.

2. Staircase Learning Framework:
   The training follows a progressive, linear path ($v_{struct} \rightarrow v_{prop} \rightarrow v_{opt}$) that mirrors the cognitive process of human chemists. This DAG-structured approach ensures that higher-level reasoning capabilities are built upon reliable structural understanding:

   • Stage 1 (Structure Comprehension): Establishing fundamental graph perception.
   • Stage 2 (Structural Property Prediction): Deriving properties (e.g., LogP, MW) from structural features.
   • Stage 3 (Objective Molecular Optimization): Modifying molecules to achieve desired properties.
     This curriculum learning approach prevents the interference of complex tasks on foundational learning.

3. Chain-of-Thought (CoT) Supervision:
   To enhance accuracy and interpretability, the framework employs a two-stage training strategy for each task using LoRA fine-tuning:

   • Stage 1 (Answer Learning): The model learns to predict the correct answer ($a$) to establish domain knowledge anchors.
   • Stage 2 (Reasoning Learning): The model learns the complete response containing a Chain-of-Thought ($z$) distilled from strong reasoning models (specifically GPT-5).
     The training objective jointly models the reasoning process and the final result: $P_\theta(a, z | S, q) = P_\theta(z | S, q) \cdot P_\theta(a | S, z, q)$. The CoT includes deterministic intermediate results for regression tasks (e.g., explicit atom parsing) and an "analysis–strategy–verification" pipeline for optimization tasks.

Key Results
Experiments were conducted on the PubChemSTM dataset and various downstream benchmarks, comparing MolBasic against generalist LLMs, specialist molecular LLMs, and traditional deep learning models.

• Structure Comprehension: MolBasic achieved substantial improvements across all MSC tasks. Most notably, it reached 94.65% accuracy on SMILES-to-Graph conversion and 85.66% on Graph-to-SMILES, whereas all baseline models (including specialized LLMs) performed near zero on these bidirectional conversion tasks.
• Property Prediction: On structural property prediction (MW, LogP, TPSA, etc.), MolBasic achieved the best performance among LLM-based methods, with Pearson correlations exceeding 0.97 for most properties and Mean Absolute Errors (MAE) comparable to or better than task-specific models like Chemprop.
• Molecular Optimization: In the task of increasing LogP, MolBasic achieved the highest LogP improvement (2.27) among LLM-based models while maintaining high validity (98.7%) and diversity (0.86).
• Ablation Studies:
  • Removing the structure comprehension stage ("w/o Struct") led to a monotonic degradation in downstream performance, confirming that structural perception is a necessary foundation.
  • Randomizing the training order ("Mixed SFT") resulted in lower performance than the staircase approach, indicating that the progressive ordering is critical to prevent task interference.
  • Removing CoT supervision ("w/o CoT") caused consistent performance drops, highlighting the necessity of explicit intermediate reasoning guidance.

Significance and Claims
The paper claims that the lack of basic structure comprehension is a fundamental bottleneck in current molecular LLMs. By "backing to basics" and prioritizing the alignment of sequential text with topological graph structures, MolBasic enables models to perform reliable, interpretable, and generalizable molecular reasoning. The authors assert that their structure-first paradigm resolves the contradiction between downstream success and structural failure, providing a robust foundation for future "AI chemist" applications. The work emphasizes that without accurately perceiving molecular graphs, higher-level reasoning tasks are inherently unreliable.

Limitations
The authors acknowledge that the current staircase design models reasoning as a linear chain. As tasks move further from the initial structure comprehension stage, the direct influence of structural perception may weaken, resulting in less pronounced improvements in optimization compared to property prediction. They suggest that future work should explore more flexible, non-linear reasoning frameworks that allow LLMs to explicitly utilize molecular structure as a foundation for diverse tasks without the interference observed in mixed-task training.