SynCraft: Reasoning-Driven Molecular Optimization

• SynCraft is a molecular optimization framework that reformulates synthesizability challenges as discrete graph-editing problems.
• It integrates LLM-driven chain-of-thought reasoning with RDKit-based deterministic edits to ensure valid chemical modifications.
• Evaluations on synthesis cliff datasets demonstrate improved synthetic feasibility and structural preservation compared to template-based methods.

SynCraft is a reasoning-driven molecular optimization framework designed to address challenges in molecular synthesizability encountered by generative models in computational chemistry. Unlike prior approaches based on post-hoc filtering or projection into pre-defined templates, SynCraft formulates synthesizability optimization as a discrete graph-editing problem, leveraging LLMs for edit planning and deterministic chemoinformatics toolkits for execution. Central to its design is the navigation of “synthesis cliffs,” regions where minimal atom- or bond-level modifications produce qualitatively enhanced synthetic feasibility while maintaining structural and pharmacophoric integrity (Li et al., 23 Dec 2025).

1. Framework Design and Workflow

SynCraft incorporates a two-component workflow. First, a LLM (Gemini-2.5-Pro) is prompted using Chain-of-Thought reasoning to identify synthetic liabilities within a molecule and propose an explicit, executable edit sequence. The edit commands are JSON-encoded and specify discrete atom-, bond-, and stereochemistry-level operations. Second, a deterministic chemoinformatics toolkit (RDKit) applies these edit operations to the molecular graph, ensuring chemical validity.

Traditional generative approaches issue the top-level prompt “generate a synthesizable analog of this SMILES.” In contrast, SynCraft issues the prompt “given identified synthetic liabilities (and optional biochemical constraints), propose a minimal sequence of atom- and bond-level edit commands.” This paradigm decouples molecular reasoning (LLM) from deterministic graph transformation (RDKit).

A synthesis cliff is operationally defined as a pair $(M_\text{src}, M_\text{tgt})$ where $M_\text{tgt}$ is both highly similar (in ECFP4-Tanimoto space) and synthesizable, whereas $M_\text{src}$ is unsynthesizable. Figure 1 of the reference provides qualitative illustrations of synthesis cliff navigation via edit sequences.

2. Formal Problem Specification

Molecules are modeled as labeled graphs $G = (V, E)$, with atoms $V=\{v_1, \ldots, v_n\}$ carrying element labels $\ell(v_i)\in\{\text{C},\text{N},\ldots\}$ and bonds $E \subset V \times V$ annotated with orders $o(e) \in \{1,2,3,\text{aromatic}\}$. An edit sequence $\mathbf{a} = (a_1, a_2, \ldots, a_T)$, $a_t \in \mathcal{A}$, is drawn from a defined action space $\mathcal{A}$.

The synthesizability optimization objective is formalized as: $$\mathbf{a}^* = \arg\max_{\mathbf{a}}\, \mathrm{Sim}(G_\text{src}, G_\text{src} \oplus \mathbf{a}) \quad \text{s.t.}\quad \mathrm{Retro}(G_\text{src} \oplus \mathbf{a}) = 1,$$ where $\oplus \mathbf{a}$ denotes sequential application of $\mathbf{a}$ to $G_\text{src}$, $\mathrm{Sim}(\cdot, \cdot)$ is the ECFP4-Tanimoto similarity, and $\mathrm{Retro}(\cdot)$ is a binary indicator denoting whether SimpRetro can find a valid retrosynthetic route within 30 minutes. The approach is extensible to continuous synthetic accessibility scores $S(G)$.

3. Edit Space and LLM-Driven Editing

3.1 Edit Operations

The edit action set $\mathcal{A}$ includes:

• $\text{DEL\_ATOM}(i)$: Remove atom $i$ and its bonds.
• $\text{MUTATE\_ATOM}(i, X \rightarrow Y)$: Change atom $i$’s element.
• $\text{ADD\_ATOM}(j, Y)$: Introduce new atom $j \geq 500$ with element $Y$.
• $\text{ADD\_BOND}(i,j, \text{order})$: Add a bond.
• $\text{DEL\_BOND}(i,j)$: Remove specified bond.
• $\text{CHANGE\_BOND}(i,j, \text{new\_order})$: Adjust bond parameters/aromaticity.
• $\text{SET\_CHIRAL}(i, R/S)$ and $\text{SET\_BOND\_STEREO}(i,j, E/Z)$: Assign stereochemistry.

These operations are parameterized by atom-map identifiers and, where applicable, bond orders and stereochemical descriptors.

3.2 Prompt Engineering and Inference Pipeline

SynCraft uses retrieval-augmented few-shot prompting. For a given input, it retrieves the top-$k$ ($k=5$) most relevant synthesis cliff pairs $(M_\text{src}, M_\text{tgt})$ and formats each example as:

• Source SMILES string
• Chain-of-Thought reasoning statement describing liabilities
• JSON array with the edit sequence

The LLM is instructed to output first an analysis, then the sequence of edits. The core inference pipeline is captured in the following pseudocode:

function Optimize(G_src): D_knn = RetrieveTopK(G_src, k=5) prompt = BuildPrompt(G_src, D_knn) LLM_out = LLM_Generate(prompt) a = ParseJSON(LLM_out) G_new = ApplyEdits(G_src, a) return G_new

Prompt instructions enforce concise natural-language reasoning followed by edit command output, minimizing syntactic deviations.

4. Optimization, Constraints, and Loss Formulation

Interaction-aware prompting extends SynCraft to incorporate biological constraints derived from structure-based data. A PLIP-derived interaction profile, after docking the ligand into a target protein, is mapped to atom-map indices and described per atom, e.g.,

Atom [i] (Element E, connected to [N]): interaction_data

This is prepended under “[CRITICAL BIOLOGICAL CONSTRAINTS],” explicitly instructing the LLM to preserve or only bioisosterically modify these atoms.

No neural loss is trained; instead, explicit prompting guides the LLM to optimize for

• Structural fidelity (ECFP4-Tanimoto similarity)
• Synthetically feasible output $(\mathrm{Retro}(G_\text{new})=1)$
• Optional bio-constraints

Ablation results indicate edit-sequence generation avoids “SMILES hallucinations” common with direct sequence generation, yielding higher output validity and similarity at each threshold.

5. Datasets, Benchmarks, and Comparative Evaluation

SynCraft was evaluated using:

• The Synthesis Cliff dataset, containing 3,332 $(M_\text{src}, M_\text{tgt})$ pairs from GenBench3D outputs (five generative models), matched to eMolecules neighbors $(\text{Tanimoto} > 0.5$, $\text{Pharm2D} > 0.5)$.
• Test sets: 1,025 unsynthesizable Pocket2Mol (P2M) compounds and 1,725 ResGen outputs, both labeled by SimpRetro feasibility.

Baselines included ChemProjector, SynFormer, and ReaSyn (projection-based). The primary metric is the fraction of cases for which a synthesizable analog $G_\text{new}$ satisfies $\mathrm{Sim}(G_\text{src}, G_\text{new}) > \tau$ and $\mathrm{Retro}(G_\text{new}) = 1$ for $\tau \in \{0.5, 0.6, 0.7, 0.8\}$.

Dataset	Similarity $\tau$	SynCraft (%)	Best Baseline (%)
Pocket2Mol	$>0.5$	42.7	$\approx$30
Pocket2Mol	$>0.6$	28.4	15.4
ResGen	$>0.5$	44.7	37.5
ResGen	$>0.6$	29.1	22.0

Across thresholds, SynCraft consistently outperforms baselines, especially in the medium similarity regime.

6. Medicinal Chemistry Case Studies

6.1 PLK1 Inhibitor Editing (Retrospective)

For input “lig-886” (bearing a 2,5-dimethylpiperazine core causing two chiral centers), SynCraft’s output chain-of-thought was: “Remove methyl groups to avoid stereochemical mixtures.” The executed DEL_BOND and MUTATE_ATOM edits yielded a simple piperazine, matching the human-synthesized IIP0944 lead’s scaffold.

6.2 RIPK1 Candidate Rescue (Prospective)

Case I: For a biaryl linkage between electron-deficient rings, SynCraft suggested inserting an ether via ADD_ATOM (O) and CHANGE_BOND. This enabled a C–O bond formation pathway in retrosynthesis, bypassing problematic cross-coupling, and preserved H-bonding (Vina docking: $-10.7 \rightarrow -10.2$ kcal/mol).

Case II: Modifying a fused diazepino-purine to a 2,4-disubstituted pyrimidine (scaffold hop), SynCraft’s edits preserved substituent vectors and H-bond acceptor at N1. Docking showed maintained binding efficiency ($-10.2 \rightarrow -9.9$ kcal/mol).

7. Limitations and Future Research

SynCraft reframes synthesizability as a targeted, minimal graph-editing problem. This approach outperforms template-projection methods, which may introduce large topological distortions. Its modular architecture—separating LLM-driven reasoning from deterministic chemoinformatics execution—guarantees valid chemical outputs while enabling explicit pharmacophore preservation via interaction-aware prompting.

Reported limitations include operational costs (approximately \$3–5 per 100 optimizations), limiting suitability for ultra-large library enumeration. Proposed future directions include:

• Integration of a learned, continuous feasibility–fidelity loss
• Automated retrosynthetic feedback loops
• Extension to multi-objective optimization (e.g., ADMET properties)
• Adaptation to broader graph-structured domains

Source code, datasets, and full prompt templates are publicly available at github.com/catalystforyou/SynCraft-Core (Li et al., 23 Dec 2025).