CSGD: Composable Score-based Graph Diffusion

• The paper introduces CSGD as the first model extending score matching to discrete graphs using concrete scores to enhance multi-conditional molecular generation.
• CSGD employs composable guidance and subset-selectable control to flexibly enforce arbitrary combinations of molecular property constraints.
• Empirical results show significant improvements in validity and controllability metrics over traditional graph diffusion methods.

Searching arXiv for the specified CSGD paper and related graph diffusion context. Search query: "Composable Score-based Graph Diffusion Model for Multi-Conditional Molecular Generation" Searching arXiv now. Composable Score-based Graph Diffusion Model (CSGD) is a diffusion model for multi-conditional molecular generation that operates directly on discrete molecular graphs and targets controllable generation under multiple property constraints (Qiao et al., 11 Sep 2025). In the paper’s formulation, the generated object is a molecular graph $G=(\mathbf X,\mathbf E)$, where node and edge variables are categorical, and the objective is to balance distribution learning / fidelity against controllability under a condition set $\mathcal C=\{c_1,\dots,c_M\}$. CSGD is presented as the first model that extends score matching to discrete graphs via concrete scores, and it couples that formulation with Composable Guidance (CoG) and Probability Calibration (PC) to support flexible conditioning on arbitrary subsets of properties during sampling (Qiao et al., 11 Sep 2025).

1. Problem formulation and multi-conditional objective

CSGD addresses the setting in which a molecular graph is generated subject to multiple requested properties, each of which may be numerical or categorical. The target conditional distribution is written as

$$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$

Within this formulation, the model is required to satisfy two competing objectives. The first is distribution learning / fidelity, meaning that generated molecules should resemble the data distribution $p(G)$. The second is controllability, meaning that generated molecules should satisfy the requested properties through alignment with $p(\mathcal C\mid G)$ (Qiao et al., 11 Sep 2025).

A central practical requirement is flexibility at sampling time. The paper emphasizes that one may want to specify only an arbitrary subset of properties rather than the entire condition set. This requirement is treated as a first-class design constraint rather than an afterthought. In that sense, CSGD is not merely a conditional generator; it is a generator designed for subset-selectable control over multiple properties (Qiao et al., 11 Sep 2025).

The paper frames prior discrete graph diffusion methods as insufficiently adapted to this setting for two reasons. First, methods such as MOOD and GraphDiT use monolithic joint conditioning: even when numerical and categorical properties are encoded separately, they are merged into a single combined condition representation. Second, most discrete graph diffusion methods rely on mean-prediction formulations rather than explicit score matching. The paper argues that this deprives such methods of a score representation suitable for compositional guidance, and it further notes that predictor-guidance approaches such as DiGress require relaxing discrete graphs into continuous variables, which can harm generation fidelity (Qiao et al., 11 Sep 2025).

2. Discrete graph diffusion and concrete-score formulation

CSGD builds a score-based formulation directly on discrete graphs. Node and edge alphabets are categorical,

$$\mathcal X=\{1,\dots,a\}, \qquad \mathcal E=\{1,\dots,b\},$$

and a graph is represented as

$$G=(\mathbf X,\mathbf E), \quad \mathbf X=(x^i)_{i=1}^n,\quad \mathbf E=(e^{ij})_{i,j=1}^n.$$

The diffusion process factorizes over nodes and edges: $$p_t(G_t)=\prod_{i=1}^n p_t^X(x_t^i)\prod_{i,j=1}^n p_t^E(e_t^{ij}).$$ Each token follows a continuous-time Markov process,

$$\frac{d p_t^X}{dt}=Q_t^X p_t^X,\qquad \frac{d p_t^E}{dt}=Q_t^E p_t^E,$$

with time-dependent generators

$Q_t^X=\sigma(t)Q^X,\qquad Q_t^E=\sigma(t)Q^E.$

The closed-form forward transition from the clean state uses matrix exponentials with

$\mathcal C=\{c_1,\dots,c_M\}$0

and the full forward transition factorizes over node and edge tokens (Qiao et al., 11 Sep 2025).

The paper considers two standard transition choices. Absorb introduces a MASK absorbing state. Uniform transitions uniformly among non-current states. These choices control the form of the corruption process on categorical graph variables (Qiao et al., 11 Sep 2025).

The distinctive step is the replacement of continuous gradients by concrete scores. In continuous diffusion, a score is $\mathcal C=\{c_1,\dots,c_M\}$1. In CSGD, the discrete analogue is a local ratio of probabilities between an alternative token state and the current token state. For nodes,

$\mathcal C=\{c_1,\dots,c_M\}$2

An analogous expression is used for edges. Thus, the model does not directly estimate the clean-sample distribution; it estimates local ratios that play the role of discrete scores. The denoiser is a Graph Transformer that processes node and edge tokens and outputs node and edge concrete scores (Qiao et al., 11 Sep 2025).

3. Training objective, reverse process, and sampling dynamics

The training objective is Diffusion Weighted Denoising Score Entropy, adapted from SEDD. The node loss is defined by comparing the predicted concrete score against the corresponding forward-process probability ratio; the edge loss is analogous. The final objective is

$\mathcal C=\{c_1,\dots,c_M\}$3

Optimization uses Adam, gradient clipping, and learning-rate warmup (Qiao et al., 11 Sep 2025).

This training criterion is the operational core of the paper’s claim that CSGD extends score matching to discrete graphs via concrete scores. The target of learning is no longer a denoised graph distribution in the conventional mean-prediction sense, but a collection of token-level score ratios defined on the noisy marginal (Qiao et al., 11 Sep 2025).

The reverse process constructs token-level reverse transition probabilities by combining the learned concrete score with the forward generator for a step $\mathcal C=\{c_1,\dots,c_M\}$4. The full reverse transition again factorizes across nodes and edges: $\mathcal C=\{c_1,\dots,c_M\}$5 Sampling proceeds by iterating from

$\mathcal C=\{c_1,\dots,c_M\}$6

down to $\mathcal C=\{c_1,\dots,c_M\}$7, using $\mathcal C=\{c_1,\dots,c_M\}$8-leaping / Euler-style discrete updates token by token (Qiao et al., 11 Sep 2025).

The paper’s ablation evidence links this formulation to improved discrete molecular generation. When CSGD is used without PC and with CFG instead of CoG, it still improves validity over GraphDiT: on Polymers, validity rises from 82.45 for GraphDiT to 90.15 with CSGD Absorb and 95.81 with CSGD Uniform; on HIV, validity rises from 69.06 to 74.01 and 81.30, respectively. The paper interprets this as support for the claim that concrete-score modeling itself improves discrete molecular generation (Qiao et al., 11 Sep 2025).

4. Composable Guidance and Probability Calibration

The conditional guidance mechanism is where CSGD most sharply departs from fixed joint-conditioning schemes. As a baseline, the paper considers standard classifier-free guidance (CFG): $\mathcal C=\{c_1,\dots,c_M\}$9 This still treats $$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$0 as a single joint condition (Qiao et al., 11 Sep 2025).

The paper’s main guidance mechanism is Composable Guidance (CoG). It begins from the factorization

$$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$1

which yields the composable score

$$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$2

Because each property contributes its own score term, inference can condition on any subset $$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$3: all properties, a single property, a mixed subset, or separately reweighted properties. Training uses condition dropout so that the model learns both conditional scores $$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$4 and the unconditional score $$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$5 (Qiao et al., 11 Sep 2025).

The second control mechanism is Probability Calibration (PC), introduced to address a train–test mismatch caused by strong guidance scales. The paper observes that large guidance can drive guided scores outside the range encountered during training, damaging validity and sometimes controllability. PC is applied after guided transition probabilities are computed. It first performs dynamic thresholding: raw probabilities are clipped at zero, percentile-based thresholds are used to define a stable interval, and values are rescaled into that interval. It then applies temperature scaling: $$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$6 Here $$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$7 yields sharper transitions, whereas $$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$8 yields flatter transitions. The paper treats PC as a plug-and-play postprocessing step for guided discrete diffusion (Qiao et al., 11 Sep 2025).

The sampling pipeline integrates these two mechanisms in sequence: sample $$p(G \mid \mathcal C)=p(G \mid c_1,\dots,c_M)\propto p(G)\,p(c_1,\dots,c_M\mid G).$$9; for each reverse step, compute the unconditional score, compute conditional scores for each selected property, compose them with CoG, convert the result into transition probabilities, optionally apply PC, and sample node and edge states independently tokenwise. In the paper’s summary, CoG decides what conditions to enforce, and PC decides how strongly and stably the reverse probabilities should behave (Qiao et al., 11 Sep 2025).

5. Architecture and implementation

The implementation reuses GraphDiT’s backbone. The denoiser is a Graph Transformer equipped with adaptive layer norm (AdaLN), a separate categorical encoder, and a separate numerical/cluster encoder. The timestep embedding is combined with the condition embedding as

$p(G)$0

The denoiser processes noisy node and edge features, updates hidden states with Graph Transformer layers conditioned on $p(G)$1, and outputs node and edge log-concrete scores. Condition dropout is used so that the model can learn both conditional and unconditional estimates (Qiao et al., 11 Sep 2025).

The reported training configuration is explicit. The model uses 1000 diffusion steps, batch size 1200, learning rate $p(G)$2, gradient clipping norm 1, linear warmup for 1500 iterations, and a log-linear noise schedule. Training is reported on a single NVIDIA A800 80GB GPU (Qiao et al., 11 Sep 2025).

These implementation choices place CSGD in a pragmatic lineage of graph diffusion systems that inherit backbone architectures from earlier discrete graph generators while modifying the probabilistic objective and sampling rule. A plausible implication is that the main novelty is concentrated in the score representation and guidance mechanism, rather than in a wholly new graph encoder. That implication follows from the paper’s explicit reuse of GraphDiT’s backbone and its emphasis on concrete scores, CoG, and PC (Qiao et al., 11 Sep 2025).

6. Datasets, metrics, and empirical results

The empirical study uses four public molecular datasets. The Polymers benchmark includes four numerical properties: Synthesizability (SAS + SCS as a 2D vector), $p(G)$3 permeability, $p(G)$4 permeability, and $p(G)$5 permeability. The drug-design benchmarks are HIV, BBBP, and BACE; each sample there has a numerical synthesizability condition and a categorical task-specific property, respectively HIV replication inhibition, BBBP label, and BACE label (Qiao et al., 11 Sep 2025).

Evaluation uses three metric families: Validity; Distribution learning, measured with Diversity, Similarity, and FCD; and Controllability, measured with MAE for numerical properties and Accuracy for categorical properties. The evaluations use MOSES metrics and an Oracle random forest predictor (Qiao et al., 11 Sep 2025).

The headline empirical claim is that CSGD achieves state-of-the-art performance across the four datasets, with an average 15.3% improvement in controllability over prior methods, while maintaining high validity and fidelity (Qiao et al., 11 Sep 2025). On Polymers, relative to GraphDiT, CSGD Uniform reaches 99.38% validity versus 82.45%, and FCD 6.4351 versus 6.6443. The controllability MAEs also improve: Synth 1.1905 versus 1.2973; $p(G)$6 0.6054 versus 0.7440; $p(G)$7 0.6993 versus 0.8857; and $p(G)$8 0.6232 versus 0.7550. The paper reports an average 15.3% relative improvement in MAE over the strongest baseline on this benchmark (Qiao et al., 11 Sep 2025).

On the drug datasets, CSGD again outperforms or matches prior methods, especially in validity and controllability. For HIV, GraphDiT reports validity 69.06 and inhibition accuracy 0.9437, whereas CSGD Uniform reports validity 83.36 and accuracy 0.9898. For BBBP, GraphDiT reports validity 78.72, accuracy 0.9089, and MAE 0.4492, whereas CSGD Uniform reports validity 95.63, accuracy 0.9831, and MAE 1.0745. The paper explicitly notes that this indicates a fidelity/control trade-off depending on the transition choice: Uniform tends to push validity and accuracy higher but can sometimes worsen MAE on some tasks. For BACE, GraphDiT reports validity 76.00 and accuracy 0.9132, whereas CSGD Uniform reports validity 96.92 and accuracy 0.9250 (Qiao et al., 11 Sep 2025).

The ablation study isolates the contributions of CoG and PC. The paper states that CoG performs comparably to CFG on the same backbone, whereas PC consistently improves validity and controllability across datasets. On Polymers, Uniform, CFG without PC gives validity 95.81, CFG with PC gives 99.51, CoG without PC gives 98.84, and CoG with PC gives 99.36. On HIV, Uniform, CFG without PC gives validity 81.30 and accuracy 0.9898, CFG with PC gives 82.34 and 0.9910, CoG without PC gives 80.89 and 0.9869, and CoG with PC gives 83.36 and 0.9898 (Qiao et al., 11 Sep 2025).

The most direct evidence for compositional control comes from experiments on arbitrary subsets of the Polymers properties. CoG is tested on only Synthesizability, only gas permeability properties, mixed subsets such as Synth + one gas property, and the full property set. Reported examples include $p(G)$9: validity 99.78, Synth MAE 0.9179; $p(\mathcal C\mid G)$0: validity 99.38, gas MAEs all around 0.60–0.69; and $p(\mathcal C\mid G)$1: validity 99.50, with strong control on the selected properties. This is the paper’s clearest demonstration that CoG provides subset-selectable control (Qiao et al., 11 Sep 2025).

7. Position within graph diffusion research and related terminology

A 2024 survey of diffusion-based graph generative methods does not mention “Composable Score-based Graph Diffusion Model (CSGD)” explicitly by name in the provided text, but it gives a framework in which CSGD fits naturally (Chen et al., 2024). The survey organizes graph diffusion into DDPMs, SGMs, and Score SDEs, and explicitly states that score-based generative models (SGMs) intend to restore the original data distribution through estimating the Stein score, which is the gradient of data $p(\mathcal C\mid G)$2. In the survey’s taxonomy, CSGD belongs most naturally to the SGM / Score SDE family, rather than the DDPM family, because its central operation is score estimation and score composition on noised graph states (Chen et al., 2024).

The survey also emphasizes a structural difficulty that is directly relevant to CSGD: “due to the intrinsic discrete nature of graphs, there are a number of significant challenges when implementing diffusion models to graph generation.” It further notes that there is “no universal approach for graph diffusion and edge generation.” Within that landscape, graph-specific score-based or score-SDE methods such as EDP-GNN, GraphGDP, GSDM, NVDiff, and SLD form the closest conceptual neighborhood for CSGD, although the survey does not discuss CSGD’s specific mechanisms of compositional guidance or probability calibration (Chen et al., 2024).

A separate but similarly named method is CCSD, “Combinatorial Complex Score-based Diffusion Modelling through Stochastic Differential Equations,” which targets combinatorial complexes rather than multi-conditional molecular graph generation (Carrel, 2024). CCSD generalizes graph generation to higher-order topological structures and unifies score-based paradigms under an SDE formalism. The two methods therefore share score-based language but address different modeling objects and problem settings: CSGD centers on discrete molecular graphs, concrete scores, and flexible multi-property conditioning, whereas CCSD centers on combinatorial complexes and higher-order relations (Carrel, 2024).

The resulting conceptual significance of CSGD is specific. The paper states that it changes the conditioning paradigm for discrete graph diffusion from fixed joint conditioning to composable per-property guidance and that, by moving from mean prediction to concrete score matching, it enables score-space manipulations that were previously unavailable for discrete molecular graph diffusion (Qiao et al., 11 Sep 2025). This suggests that CSGD’s main contribution is not only improved controllability metrics, but also a reformulation of conditional discrete graph diffusion in a score-compositional language that supports property-wise inclusion, exclusion, and reweighting at inference time.