Jul 2026· Journal of Chemical Information and Modeling· Vol 66, pp. 9045-9056· 0 citations· 36 references
Medicine
TL;DR
DF-S4, a conditional molecular generation framework based on Structured State Space Models (S4), which addresses this limitation through a disentangled latent representation and hierarchical feature-wise linear modulation (FiLM) through a disentangled latent representation and hierarchical feature-wise linear modulation.
Abstract
Balancing target-specific biological affinity with drug-likeness remains a central challenge in de novo molecular design, where existing generative models often exhibit limited controllability or reduced structural diversity. Here, we present DF-S4, a conditional molecular generation framework based on Structured State Space Models (S4), which addresses this limitation through a disentangled latent representation and hierarchical feature-wise linear modulation (FiLM). By decoupling structural and property variables and injecting conditional signals across multiple representation levels, DF-S4 enables fine-grained and stable multi-objective control beyond conventional input-level conditioning. Evaluated via a rigorous progressive multi-denominator auditing framework across three kinase targets (EGFR, BRAF, and FGFR1), DF-S4 exhibits robust target-steering performance, yielding favorable intradomain active ratios (74.2-83.9%) while maintaining high novelty (>95%) and competitive internal diversity (∼0.85). Furthermore, DF-S4 shifts the multi-objective Pareto Frontier toward regions of simultaneous high apparent affinity and drug-likeness, mitigating the distributional-trapping trade-offs commonly observed in prior approaches. Molecular docking analyses confirm the physical plausibility of generated candidates, yielding stronger binding affinities than reference inhibitors. Finally, comprehensive ablation studies and latent space dependence analyses mathematically validate that both explicit latent disentanglement and hierarchical FiLM modulation are critical for robust feature isolation, tighter property alignment, and generative stability.
Molecular generation is a core task in drug discovery. Although existing deep generative models can produce valid molecular structures, they lack precise control over molecular properties such as lipophilicity (logP) and drug-likeness (QED). This paper proposes Property-Regularized Graph Variational Autoencoder (PR-GVAE), a conditional graph VAE that introduces Feature-wise Linear Modulation (FiLM) conditioning in the encoder and a dual-branch fusion architecture in the decoder for precise property control. Experiments on a filtered subset of ZINC-20 (approximately 52,000 molecules, at most 20 heavy atoms) across five models spanning graph-based, sequence-based, and fragment-assembly paradigms demonstrate that PRGVAE achieves a Condition Satisfaction Rate (CSR) for both properties of 78.3%, outperforming the unconditional Vanilla GraphVAE by 2.0× (39.4%) and the sequence-based SMILES C-VAE by 40 percentage points (38.3%). Ablation experiments reveal that decoder conditioning is the core driver of property control (removing it reduces CSR-Both by 26.6%), while encoder FiLM conditioning and dual-branch fusion exhibit additive contributions (simultaneous removal reduces CSR-Both by 30.8%). Visualization confirms that PR-GVAE generates structurally diverse drug-like molecules, providing an effective solution for on-demand molecular design.
Lu Yu, Tianyu Han, Shuyue Men et al.· 2026 IEEE 27th China Confere...· 0 citations
A high-accuracy surrogate predictor based on message-passing neural networks is decoupled from heterogeneous generators, and integrated into a closed-loop strategy of generate, score, select, and regenerate to progressively enrich high-performing candidates without modifying the underlying model architectures.
Zhaosheng Zhang, Yanbo Liu, Jiadong Liu et al.· Physical Chemistry, Chemical...· 0 citations
The design of molecules with desired properties is a key challenge in drug discovery and materials science. Traditional methods rely on trial-and-error, while recent deep-learning approaches accelerate molecular generation. However, existing models struggle with generating molecules based on specific textual descriptions. We introduce Mol-CADiff, a diffusion-based framework that uses causal attention mechanisms for text-conditional molecular generation. Our approach explicitly models the causal relationship between textual prompts and molecular structures, overcoming limitations in existing methods. We enhance dependency modeling both within and across modalities, enabling precise control over the generation process. While primarily designed for text-guided tasks, this architecture inherently supports unconditional generation, providing the added capability to autonomously sample the broader chemical space without explicit constraints. Here we show that Mol-CADiff outperforms alternative methods in generating diverse, chemically valid molecules, with better alignment to specified properties, enabling more intuitive language-driven molecular design. By bridging these modalities, our framework provides a versatile method for drug discovery. Computational approaches to molecular design often explore only limited regions of the vast chemical space. This study presents a causality-aware diffusion model that generates valid and diverse molecules with or without text prompts, improving controllability in molecular design.
While recent advances in data synthesis aim to curate high-quality datasets, most generation pipelines still rely on heuristic prompt-based control. This black-box paradigm provides limited insight into how individual samples interact with a model's underlying learning dynamics. To bridge this gap, we propose a circuit-grounded framework that connects training-dynamics-based data valuation with mechanistic interpretability (MI). Specifically, we conceptualize data quality along three complementary utility axes, learnability, challenge, and alignment. First, we uncover specialized model-internal circuits that causally govern these utility signals. Then, moving beyond heuristic prompting toward mechanistic control, we leverage these circuits as controllable interfaces, actively steering generation to produce utility-targeted data. Building on this capability, we introduce SAMS (Stage-Aware Mechanistic Scheduling), which schedules circuit-steered data according to the model's evolving optimization needs. Experiments on multiple-choice QA tasks demonstrate that our approach yields precisely controlled data with greater diversity than prompt-based baselines, consistently improving downstream performance and calibration. Ultimately, this work establishes a principled white-box paradigm for interpretable data generation, pioneering the use of MI not just as an analytical tool, but as a practical, controllable interface.
Diffusion Language Models (DLMs) offer a compelling alternative to autoregressive (AR) generation by enabling bidirectional context and iterative refinement. However, their reliability under natural input noise and adversarial attacks remains under-explored. To address this, we systematically evaluate DLM robustness and calibration against AR baselines, using two parameter-matched pairs (LLaDA-8B vs. LLaMA-3-8B and Dream-7B vs. Qwen2.5-7B) across 32 natural perturbation conditions, adversarial gradient probes, and mechanistic hidden-state analyses. This paired design effectively isolates architecture-intrinsic properties from weight-dependent behaviors. We find a nuanced robustness profile: while highly stochastic DLM loss landscapes naturally resist gradient-based adversarial suffixes, they provide no guaranteed defense against natural noise, proving that everyday robustness is weight-dependent rather than inherently architectural. Furthermore, DLMs exhibit systematic overconfidence, presenting a practical deployment hazard. Most crucially, mechanistic probing reveals that all models perfectly encode input corruption, isolating behavioral fragility entirely to a decoder routing failure. Consistent with this diagnosis, we show that surface-level prompt patching fails to improve over noisy baselines. Ultimately, DLM robustness cannot be patched on; it must be fundamentally integrated into the iterative decoding loop.
Saurabh Yadav, B. N. Patro, V. Agneeswaran· 0 citations
Scientific datasets, such as materials and molecular datasets, are often large, complex, and open-ended, posing a core challenge for data efficiency and model training. While data attribution (DA) offers a principled way to score and select samples for efficient learning, we identify a fundamental misalignment between existing DA assumptions and the phase-structured training process. Most DA methods assign each sample a single and static score under a fixed objective, which is difficult to align with scientific training dynamics. Through preliminary results, we identify a clear phase shift: early training primarily benefits from memorizing common and recurring patterns from abundant in-distribution (ID) samples, while later training increasingly relies on generalization by emphasizing complex, long-tailed out-of-distribution (OOD) samples. This phase-dependent behavior implies that a sample's value is not fixed, but can change with the model's state. To address this, we propose DuetDA, a decomposed and dynamic DA framework with model-state gating. DuetDA assigns each sample two complementary values, i.e., a memorization value and a generalization value, and uses a lightweight gate conditioned on the current model state to adaptively integrate them across training. We train DuetDA via meta-learning and apply it at inference time for data attribution, avoiding expensive second-order computations while directly optimizing for downstream performance gains. Theoretically, we show that single-objective static attribution is inherently limited under phase-structured objectives, and that dynamic integration is necessary to recover phase-dependent sample attribution. Empirically, on scientific benchmark MatBench, we show that (1) for accelerated training, DuetDA achieves more than 50% compute reduction while even improving the baselines on OOD set, and (2) compared to data pruning baselines, DuetDA gains about an overall 11% performance compared to the best general data pruning methods on OOD set, showing the effectiveness for open-ended data. We open-source our code at https://github.com/cjpcool/DuetDA.
Jianpeng Chen, Wangzhi Zhan, Haohui Wang et al.· Proceedings of the 32nd ACM...· 0 citations