Machine learning accelerates molecular property prediction, yet state-of-the-art Large Language Models and Graph Neural Networks operate as black boxes. In drug discovery, where safety is critical, this opacity risks masking false correlations and excluding human expertise. Existing interpretability methods suffer from the effectiveness-trustworthiness trade-off: explanations may fail to reflect a model's true reasoning, degrade performance, or lack domain grounding. Concept Bottleneck Models (CBMs) offer a solution by projecting inputs to human-interpretable concepts before readout, ensuring that explanations are inherently faithful to the decision process. However, adapting CBMs to chemistry faces three challenges: the Relevance Gap (selecting task-relevant concepts from a large descriptor space), the Annotation Gap (obtaining concept supervision for molecular data), and the Capacity Gap (degrading performance due to bottleneck constraints). We introduce GlassMol, a model-agnostic CBM that addresses these gaps through automated concept curation and LLM-guided concept selection. Experiments across thirteen benchmarks demonstrate that \method generally matches or exceeds black-box baselines, suggesting that interpretability does not sacrifice performance and challenging the commonly assumed trade-off. Code is available at https://github.com/walleio/GlassMol.
Figures & tables
Figure 1: The proposed GlassMol framework . The inputs to the encoder are the raw data, either a SMILES string or a molecular graph. The embeddings generated by the LLM or GNN are projected onto a set of task-relevant concepts selected by an LLM. The predicted concept values are then passed to a linear layer to produce the final prediction. The importance of each concept is determined by the magnitude of its contribution to the final output, computed as the product of the concept value and its weight.
Dataset
LLM Architecture
GNN Architecture
Baseline
GlassMol (Ours)
Baseline
GlassMol (Ours)
—ADME Properties—
BBB
0.856 ± 0.010
0.892 ± 0.023
0.840 ± 0.006
0.867 ± 0.014
Bioavailability
0.556 ± 0.054
0.679 ± 0.010
0.615 ± 0.044
0.572 ± 0.045
HIA
0.880 ± 0.111
0.989 ± 0.007
0.768 ± 0.041
0.867 ± 0.057
Pgp
0.801 ± 0.005
0.903 ± 0.003
0.899 ± 0.011
0.861 ± 0.014
Table 1: Experimental results on Therapeutics Data Commons datasets. Besides the absence of a concept loss and concept layer, the baseline architectures are identical to their respective GlassMol architectures. We report the AUROC (mean ± std) over five runs. Bold indicates the best performance within the same architecture. The columns for our method ( GlassMol ) are highlighted.
Figure 2: t-SNE visualization of learned representations on the HIA dataset. Left: The baseline shows entangled representations with no clear decision boundary. Right: GlassMol (ours) produces well-separated clusters for the two classes.
Figure 3: Qualitative comparison of concept attribution against ground-truth structural importance. The blue highlights represent the TopoPool ground-truth importance clusters. Red circles indicate the substructures identified by GlassMol where solid lines GNN architecture and dashed lines LLM architecture.
Figure 4: Ablation studies on critical model components for LLM (top row) and GNN (bottom row) architectures. Left: Comparison of model backbones. For LLMs, our task-specific small model (red) significantly outperforms general large models (blue/beige). For GNNs, our GINE-based architecture remains competitive. Middle: Evaluation of concept selection strategies. GPT-4 selected concepts (red) achieve the best performance, with the open-source Llama-3-70B selector (blue) serving as a strong alternative to proprietary models. Right: Robustness analysis showing that our method maintains high performance even with noisy concept labels (blue).
Figure 5: Hyperparameter sensitivity analysis. (a) Performance vs. number of concepts ( K ). Accuracy improves rapidly and saturates around K=40 , balancing representation power and complexity. (b) Performance vs. loss weight λ . A balanced weight of λ=1 yields optimal results, while extreme values degrade performance.
Optimizing absorption, distribution, metabolism, and excretion (ADME) is an important part of small molecule drug discovery. Many machine learning models have been built to predict ADME properties to facilitate this optimization process, but explaining model predictions is challenging. We propose a new graph neural network architecture with built-in atom attributions. Our model MolLedger learns a global context vector for each molecule and a per-atom head to output atom scores that sum to the predicted property. The atom scores are regularized to align with relevant chemical properties. We prove that MolLedger is a universal approximator and demonstrate empirically that the new architecture obtains explainability with little effect on performance. We show that the interpretations from MolLedger are faithful, concordant with held-out physical properties, and align with the changes between matched molecular pairs. Our case studies comparing interpretations from multiple methods on molecular pairs reveal that MolLedger is much better at producing sensible explanations for predicted property changes.
Molecular property models increasingly support high-stakes drug-discovery decisions, but their outputs are often difficult to audit: classical predictors return scores without rationale, while language models can produce fluent explanations weakly grounded in the input molecule. We introduce Bolek, a compact multimodal language model that grounds natural-language reasoning in molecular structure by injecting a Morgan fingerprint embedding into an instruction-tuned text decoder. Bolek is fine-tuned on molecular alignment tasks, including molecule description, RDKit descriptor prediction, and substructure detection, and on downstream reasoning over 15 TDC binary classification tasks using synthetic chains-of-thought anchored in concrete molecular features. Across these tasks, Bolek outperforms its Qwen3-4B-Instruct base on all endpoints in yes/no mode and on 13 of 15 in chain-of-thought mode, raising mean ROC/PR AUC from 0.55 to 0.76. It also outperforms TxGemma-9B-Chat on 13 of 15 binary classification tasks despite being less than half its size. Bolek's explanations are more grounded than those of the baseline LLMs: it cites numerical descriptors 10-100x more often per chain-of-thought, and the cited values agree strongly with RDKit for key descriptors such as TPSA, MolLogP, and MolWt (Spearman rho = 0.87-0.91). Generalisation extends beyond the training panel: on 15 unseen TDC classification endpoints, Bolek matches TxGemma on five, and it produces non-trivial rank correlations on three held-out regression endpoints despite never seeing downstream regression during training. These results suggest that targeted modality injection and reasoning supervision tied to verifiable molecular features can yield compact, auditable molecular reasoning models.
Frederic Grabowski, Jacek Szczerbiński, Maciej Jaśkowski +4
1Ingenix.ai, Warsaw, Poland · Laboratory of Bioinformatics and Computational Genomics, Warsaw University of Technology, Warsaw, Poland
Large language models (LLMs) are widely applied across chemical tasks, such as molecular property prediction, which underpins drug discovery. Molecular LLMs represent a molecule through several modalities, notably a 1D SMILES sequence or a 2D molecular graph. Both encode molecular information implicitly, so the contribution of individual substructures remains opaque. Retrieval and augmentation methods add context, but from external sources. However, the cues chemists reason over are the internal substructures that drive a property up or down. We propose MR-MoL, a multi-granular rationale-guided molecular LLM that supplies this evidence directly. A fine-tuned GNN scores each substructure through masking, and the most influential ones are serialized as a ranked, direction-tagged rationale that the LLM reads alongside the SMILES sequence and molecular graph. The rationale spans three levels of granularity: Murcko scaffolds with their side chains, BRICS fragments, and functional groups. This is, to our knowledge, the first method to expose GNN-derived attributions to an LLM as evidence for property prediction. On eight MoleculeNet tasks, MR-MoL achieves the best overall results among generalist models and narrows the gap to specialist models tuned for each task. Five diagnostics further confirm that the model reads the rationale rather than merely benefiting from its presence. Its direction, rank, and substructure each shape the prediction, and its attributions reproduce known structure-property relationships.