SynLaD: Synthesis-aware Latent Diffusion

• SynLaD is a latent diffusion framework that integrates 3D pharmacophore conditioning with synthesis planning in a shared latent space.
• It employs a two-stage architecture combining a variational autoencoder and a conditional diffusion model to jointly optimize molecular geometry and feasible synthesis routes.
• Empirical results show improved analogue generation, higher synthesizability, and enhanced hit rates compared to traditional sequential approaches.

SynLaD, expanded as “Synthesis-aware Latent Diffusion,” is a latent diffusion framework for small-molecule generation that conditions on 3D pharmacophore profiles while jointly modeling synthetic accessibility. It is designed to unify ligand-based drug design objectives—described as “what to make”—with explicit synthesis planning—described as “how to make it”—within a single learned latent space that decodes to both 3D molecular structures and synthesis pathways (Cretu et al., 1 Jul 2026). The framework addresses a bottleneck identified in prior approaches: current models typically optimize one of these objectives at the expense of the other, which hinders the discovery of molecules that are simultaneously high-scoring, shape-aligned, diverse, and synthesizable.

1. Problem formulation and design objective

SynLaD is motivated by the observation that small-molecule design and synthesis planning are often treated as separate computational tasks. In the formulation given for SynLaD, this separation is operationalized as a mismatch between optimizing pharmacophore- or geometry-based ligand objectives and optimizing route feasibility. The framework therefore combines reaction-constrained generation with pharmacophore-conditioned 3D design by learning a shared latent representation that supports both modalities (Cretu et al., 1 Jul 2026).

The central object is a latent variable $Z$ derived from molecular atom types and 3D coordinates. This latent is used as the common interface between structure generation and synthesis-route generation. A plausible implication is that SynLaD treats synthesizability not as a post hoc filter but as a coequal generative target. This differs from workflows in which 3D generation is performed first and retrosynthetic evaluation is applied only after candidate enumeration.

The model is explicitly framed around analogue generation tasks for bioactive ligands. Across these tasks, the intended output is not only a molecule with appropriate 3D pharmacophore alignment, but also a feasible synthetic route expressed in a serialized reaction-based notation. This dual-output design defines the method’s scope.

2. Two-stage architecture and shared latent space

SynLaD uses a two-stage pipeline. Stage 1 is a variational autoencoder with a shared encoder and two decoder heads. Stage 2 is a conditional latent diffusion model, specifically a Diffusion Transformer (DiT), that samples in the learned latent space and feeds its outputs back into the two decoders (Cretu et al., 1 Jul 2026).

The encoder is written as

$f(A,X) \to Z,$

with inputs atom types $A \in \mathbb{Z}^n$ and 3D coordinates $X \in \mathbb{R}^{3 \times n}$. Its architecture is a standard non-equivariant transformer with per-atom embeddings and random-rotation augmentation. The use of random-rotation augmentation is a factual architectural detail; this suggests an attempt to stabilize learning over 3D conformational orientation without imposing an explicitly equivariant inductive bias.

The first decoder head is geometric:

$g_{3D}(Z) \to (\hat{A}, \hat{X}).$

It is a bidirectional transformer that reconstructs atom types via cross-entropy and coordinates via mean-squared error. The second decoder head is an autoregressive synthesis model:

$g_{\text{syn}}(Z) \to S=(s_1,\ldots,s_L),$

implemented as a causal transformer that emits a serialized reaction-based synthesis plan token by token, conditioned via cross-attention on $Z$.

This arrangement makes the latent variable a multi-modal code. It is shared across geometry and synthesis, rather than partitioned into independent branches learned from separate objectives. The reported ablations later reinforce that this shared training is not incidental: joint training of geometry and synthesis heads improves synthesizability of 3D-decoder outputs.

3. Multi-modal VAE objective and latent diffusion dynamics

The Stage 1 objective is a joint reconstruction loss with KL regularization:

$$L = \lambda L_{3D} + (1-\lambda) L_{\text{synth}} + \beta D_{KL}[N(\mu_Z,\sigma_Z)\|N(0,I)].$$

The geometric term is

$$L_{3D} = \frac{1}{N}\sum_i H(a_i,\hat{a}_i) + \frac{1}{3N}\sum_i \|(x_i-\bar{x})-\hat{x}_i\|^2/\sigma^2,$$

and the synthesis term is

$$L_{\text{synth}} = \frac{1}{L}\sum_i H(s_i,\hat{s}_i),$$

where $f(A,X) \to Z,$0 is the cross-entropy on atom or token types. In this formulation, $f(A,X) \to Z,$1 trades off geometric versus synthetic reconstruction, and $f(A,X) \to Z,$2 controls the VAE’s KL penalty (Cretu et al., 1 Jul 2026).

Stage 2 defines a diffusion process in latent space. The forward noising process is a Markov chain on $f(A,X) \to Z,$3:

$f(A,X) \to Z,$4

with a predefined schedule $f(A,X) \to Z,$5. The reverse process is

$f(A,X) \to Z,$6

where $f(A,X) \to Z,$7 is predicted by the DiT denoiser and $f(A,X) \to Z,$8 denotes conditioning.

The DDPM loss is given in simplified $f(A,X) \to Z,$9-prediction form:

$A \in \mathbb{Z}^n$0

with

$A \in \mathbb{Z}^n$1

In SynLaD, however, a flow-matching variant is used; the stated core idea remains a mean-squared error on the network’s vector-field prediction $A \in \mathbb{Z}^n$2 against the ground-truth $A \in \mathbb{Z}^n$3.

The complete training objectives are therefore separated by stage:

$A \in \mathbb{Z}^n$4

and

$A \in \mathbb{Z}^n$5

This formalization indicates that the latent space is first regularized for joint decoding and only then endowed with a conditional generative prior.

4. Pharmacophore conditioning and synthesis-route serialization

A pharmacophore $A \in \mathbb{Z}^n$6 is represented by discrete feature types $A \in \mathbb{Z}^n$7 and 3D positions $A \in \mathbb{Z}^n$8. These $A \in \mathbb{Z}^n$9 features are embedded by a small transformer,

$X \in \mathbb{R}^{3 \times n}$0

and the resulting conditioning tensor $X \in \mathbb{R}^{3 \times n}$1 is introduced into every DiT block via cross-attention (Cretu et al., 1 Jul 2026).

During training, up to 3 randomly dropped pharmacophore features are used to teach robustness; at inference, the full set is supplied. SynLaD also uses classifier-free guidance:

$X \in \mathbb{R}^{3 \times n}$2

so that $X \in \mathbb{R}^{3 \times n}$3 controls conditioning strength. This makes the pharmacophore signal an explicit, adjustable driver of latent generation rather than a fixed embedding appended only once.

The synthesis plan is represented as a directed acyclic graph of building-block and reaction nodes, then serialized bottom-up into tokens. The serialization includes:

• Building block token: B <block> to add a purchasable building block.
• Reaction tokenization: F <reactants> F_r P <product> to apply a reaction.
• Termination token: S to stop.

The autoregressive transformer writes this sequence while embedding each token by summing three components: a learned action embedding, a molecular fingerprint embedding based on a projected 2048-bit Morgan fingerprint, and a positional embedding. Inference uses beam search with width $X \in \mathbb{R}^{3 \times n}$4 or sampling plus top-$X \in \mathbb{R}^{3 \times n}$5.

Product identities at inference are filled in by a separately trained reaction-prediction oracle, specified as a BART-based encoder–decoder with $X \in \mathbb{R}^{3 \times n}$6 top-1 accuracy, with the stated role of ensuring end-to-end synthesizability. A plausible implication is that route validity is not inferred solely from token well-formedness; it is grounded in an external reaction-prediction component.

5. Empirical performance

The reported experiments cover reconstruction, unconditional sampling, pharmacophore-conditioned analogue generation, and out-of-distribution hit diversification (Cretu et al., 1 Jul 2026).

Setting	Result	Notes
Autoencoder reconstruction	98.5% atom-type match	latent dim = 16, $X \in \mathbb{R}^{3 \times n}$7, $X \in \mathbb{R}^{3 \times n}$8
Autoencoder reconstruction	RMSD $X \in \mathbb{R}^{3 \times n}$9 Å	geometric reconstruction
Autoencoder reconstruction	63.4% synthesis-plan match	Tanimoto $g_{3D}(Z) \to (\hat{A}, \hat{X}).$0
Unconditional sampling, 3D outputs	90% valid	100 latents, $g_{3D}(Z) \to (\hat{A}, \hat{X}).$1 steps
Unconditional sampling, 3D outputs	PoseBusters pass $g_{3D}(Z) \to (\hat{A}, \hat{X}).$2	same setting
Unconditional sampling, 3D outputs	AiZynthFinder success $g_{3D}(Z) \to (\hat{A}, \hat{X}).$3	same setting
Unconditional sampling, synthesis outputs	100% valid	same setting
Unconditional sampling, synthesis outputs	AiZynthFinder $g_{3D}(Z) \to (\hat{A}, \hat{X}).$4	same setting
In-distribution analogue generation, synthesis outputs	$g_{3D}(Z) \to (\hat{A}, \hat{X}).$5 hits/query	Tanimoto Combo $g_{3D}(Z) \to (\hat{A}, \hat{X}).$6
In-distribution analogue generation, synthesis outputs	$g_{3D}(Z) \to (\hat{A}, \hat{X}).$7 unique scaffolds	same task
In-distribution analogue generation, synthesis outputs	AiZynth success 0.80	same task
Baseline library screen	0 hits on average	$g_{3D}(Z) \to (\hat{A}, \hat{X}).$8 compounds under same query budget
OOD Lit-PCBA, 3D outputs	$g_{3D}(Z) \to (\hat{A}, \hat{X}).$9 hits avg	500 samples/query, 10 queries
OOD Lit-PCBA, 3D outputs	18.8 unique scaffolds	same task
OOD Lit-PCBA, 3D outputs	max Combo $g_{\text{syn}}(Z) \to S=(s_1,\ldots,s_L),$0	same task
OOD Lit-PCBA, synthesis outputs	$g_{\text{syn}}(Z) \to S=(s_1,\ldots,s_L),$1 hits avg	same task
OOD Lit-PCBA, synthesis outputs	AiZynth success 0.75	same task

These results are paired with comparative statements. In out-of-distribution Lit-PCBA hit diversification, ShEPhERD and SynFormer both hit fewer synthesizable analogues and achieve lower pharmacophore overlap. A reinforcement-learning baseline, REINVENT, can match or exceed hit counts on some queries but requires $g_{\text{syn}}(Z) \to S=(s_1,\ldots,s_L),$2 h of per-query retraining versus $g_{\text{syn}}(Z) \to S=(s_1,\ldots,s_L),$3 min amortized sampling for SynLaD. The comparison isolates a specific advantage: SynLaD’s sampling cost is amortized across queries rather than dominated by per-target retraining.

The analogue-generation results are particularly notable because the baseline library screen of approximately 67k compounds finds 0 hits on average under the same query budget, whereas SynLaD synthesis outputs find on average approximately 29 hits per query with approximately 9.5 unique scaffolds. This suggests that the model’s search space is not simply reproducing a fixed enumerated compound collection.

6. Ablations, interpretation, and methodological significance

Three ablation findings are stated. First, joint training of geometry and synthesis heads improves synthesizability of 3D-decoder outputs. Second, removing the 3D VAE/diffusion branch collapses diversity, yielding few hits even though the synthesis decoder remains pharmacophore-conditioned. Third, replacing the 3D VAE with a SMILES VAE preserves diversity but sacrifices 3D pharmacophore overlap and hit rates (Cretu et al., 1 Jul 2026).

Taken together, these ablations delimit the role of explicit geometry in SynLaD. The model is not merely an overview-conditioned sequence generator augmented with a pharmacophore descriptor. Rather, the explicit 3D VAE and 3D diffusion components appear necessary for maintaining pharmacophore overlap and hit quality, while the synthesis head contributes makeability. This suggests that the shared latent space is carrying information that neither an overview-only pathway nor a SMILES-based latent alone preserves adequately.

A common misconception in this problem setting is that synthesizability can be appended as a downstream filter without materially affecting upstream generative performance. SynLaD is presented against precisely that separation: current models typically optimize one objective at the expense of the other. The reported results and ablations support the narrower claim that, within this framework, co-training geometry and synthesis produces more useful analogue-generation behavior than optimizing either axis in isolation.

In summary, SynLaD unites three-dimensional pharmacophore conditioning with explicit reaction-route generation by learning a shared latent space that can be diffused under pharmacophore guidance and decoded to both a valid 3D molecule and a feasible synthesis plan (Cretu et al., 1 Jul 2026). Its methodological significance lies in treating molecular shape alignment, scaffold diversity, and route feasibility as coupled outputs of a single amortized generative system rather than sequentially optimized modules.