PhenoMoler: Phenotype-Guided Molecule Optimization

• PhenoMoler is a phenotype-guided molecular optimization framework that uses drug-induced transcriptional profiles to steer substructure-level chemical editing.
• It combines a chemistry large language model operating in SELFIES space with a CNN-based tokenizer for differential expression to ensure chemically valid, phenotype-conditioned designs.
• The system demonstrates controllable scaffold, linker, and side-chain optimizations, with in silico evaluations showing enhanced drug-likeness and predicted target affinities.

Searching arXiv for the specified paper to ground the article. PhenoMoler is a phenotype-guided molecular optimization framework that combines a chemistry LLM with drug-induced transcriptional expression profiles to generate or edit molecules in a manner conditioned on desired phenotypic response signatures. It operates in SELFIES space, conditions generation on differential gene expression signatures from drug-treated cells, and is designed for substructure-level molecular optimization, specifically scaffold optimization, side-chain optimization, and linker optimization. The framework is presented as a shift from target-centric molecular generation toward phenotype-aware design by explicitly linking transcriptional responses to chemical structure (Song et al., 25 Sep 2025).

1. Conceptual basis and problem formulation

PhenoMoler is motivated by the observation that current molecular generative models primarily optimize drug-target binding affinity and specificity, while often neglecting system-level phenotypic effects elicited by compounds. In this framework, transcriptional profiles are treated as molecule-level readouts of drug-induced phenotypic shifts, and differential expression signatures are used as conditioning signals for molecular design. The central hypothesis is that drug-induced differential expression signatures are informative readouts of molecular mechanism and cellular response, and therefore can guide the generation of chemical structures that reproduce or improve a desired biological response (Song et al., 25 Sep 2025).

The paper formulates this linkage as a mapping from phenotype space to molecular sequence space,

$f:\mathbb{R}^n \rightarrow \mathcal{Y}^T,$

where $\mathbb{R}^n$ denotes the phenotype or expression space and $\mathcal{Y}^T$ denotes the molecular sequence space. Operationally, this corresponds to “expression signature in” and “molecular fragment reconstruction out.” This suggests a phenotype-to-structure design paradigm in which transcriptomic perturbation patterns are not auxiliary annotations but primary control signals for molecular optimization.

A central claim of the work is that, to the authors’ knowledge, it is the first study to perform substructure-level optimization explicitly guided by phenotypic profiles. This positioning is important because the framework does not restrict itself to de novo whole-molecule generation. Instead, it supports localized editing of medicinally salient regions of a molecule while conditioning those edits on a desired transcriptional response (Song et al., 25 Sep 2025).

2. Model architecture

PhenoMoler consists of three main components: a chemistry LLM, a phenotypic profile tokenizer, and a phenotype-guided molecular optimizer. The chemistry LLM is pretrained autoregressively on molecules from ZINC represented in SELFIES rather than SMILES. The use of SELFIES is emphasized because SELFIES guarantees syntactically valid outputs, which helps ensure that generated molecules remain chemically valid. The decoder is a 6-layer Transformer decoder with multi-head self-attention, feedforward layers, layer normalization, and residual connections (Song et al., 25 Sep 2025).

For self-attention, the model uses

$$\text{SelfAttn}(Q_X, K_X, V_X) = \text{softmax}\left(\frac{Q_XK_X^\top}{\sqrt{d_k} + M}\right)V_X,$$

with

$Q_X = XW^Q,\quad K_X = XW^K,\quad V_X = XW^V,$

where $M$ is an upper-triangular causal mask with $-\infty$ entries blocking access to future tokens. The role of this formulation is to preserve autoregressive generation by predicting each next SELFIES token from previous ones only.

The phenotypic profile tokenizer does not ingest the differential expression profile as a flat vector. Instead, it applies a two-layer multi-channel 1D CNN to extract structured biological patterns. The input is a one-dimensional vector of differential expression values. The first convolution uses 128 filters of size $1 \times 5$ with stride 5, producing a feature tensor of shape $128 \times 195$. The second convolution uses 256 filters of size $128 \times 1 \times 3$ with stride 2, yielding a final phenotype embedding of $\mathbb{R}^n$0. The stated intuition is that co-expressed or co-silenced genes often represent regulatory modules responding to a perturbation, so the tokenizer is intended both to compress expression profiles into manageable tokens and to preserve local gene-expression patterns that may correspond to phenotype-relevant modules (Song et al., 25 Sep 2025).

After tokenization, the phenotype embedding is passed through a 6-layer Transformer encoder. The pretrained molecular decoder is then augmented with cross-attention, with molecular tokens acting as queries and phenotype embeddings acting as keys and values:

$\mathbb{R}^n$1

where $\mathbb{R}^n$2 denotes molecular token embeddings and $\mathbb{R}^n$3 denotes phenotype features. In this configuration, molecular fragment generation is biased by the phenotype signal, creating an explicit mechanism for conditioning chemical reconstruction on transcriptional state (Song et al., 25 Sep 2025).

3. Substructure-level optimization and controllability

A defining feature of PhenoMoler is selective masking and reconstruction of specific molecular substructures rather than unconditional generation of entire molecules. The framework supports three reconstruction tasks: scaffold-mask reconstruction, linker-mask reconstruction, and side-chain-mask reconstruction. This design allows fine-grained, controllable molecular optimization (Song et al., 25 Sep 2025).

Substructures are defined using RDKit. The Murcko scaffold is extracted as the scaffold component. Linkers are identified as the shortest subgraphs connecting scaffold and substituents, and side chains are defined as non-scaffold branches from those linkers. The corresponding SELFIES fragments are replaced with special $\mathbb{R}^n$4 tokens, and the model receives both the visible molecular context and the phenotypic differential expression signature in order to reconstruct the masked region (Song et al., 25 Sep 2025).

This masking strategy narrows the search space relative to whole-molecule generation and aligns with medicinal chemistry workflows in which one aims to preserve part of a molecule while improving another part. The paper explicitly associates this selective editing scheme with fine-grained control, a smaller search space, better interpretability, and more efficient optimization than generating entire molecules blindly. A plausible implication is that the method is intended not only as a generative model but also as a structure-editing system for lead optimization under phenotypic constraints.

The training objective is described as a cross-entropy loss over masked tokens, where the model predicts missing chemical tokens conditioned on both the surrounding molecular context and the phenotype embedding. In substantive terms, the model is optimized to reconstruct masked regions in a way that is simultaneously chemically coherent and phenotypically conditioned (Song et al., 25 Sep 2025).

4. Data resources and evaluation protocol

The chemistry LLM is pretrained on 10,032,879 bioactive molecules from ZINC, converted from SMILES to SELFIES. Phenotype conditioning is based on L1000 drug-induced expression profiles selected at 10 μM and 24 hours. After averaging technical replicates, the phenotype dataset comprises 6,549 drugs, 164 cell lines, 86,400 expression profiles, and 978 landmark genes. The phenotype input uses Level 5 differential expression signatures computed by subtracting control mean expression (Song et al., 25 Sep 2025).

For target-oriented evaluation, the model is tested on ten cancer-relevant targets: AKT1, AKT2, AURKB, CTSK, EGFR, HDAC1, MTOR, PIK3CA, SMAD3, and TP53. Reference ligands are obtained from Drug Target Commons, and protein structures are taken from the PDB. For the main optimization tasks, the conditioning signatures are derived from MCF7 breast cancer cell line perturbations related to the target genes (Song et al., 25 Sep 2025).

The reported evaluation metrics cover both generative quality and medicinal chemistry relevance. These include validity, uniqueness, novelty, diversity, QED, TPSA, Tanimoto similarity, Lipinski Rule of Five metrics—MolWt, LogP, HBD, HBA, and rotatable bond count—and docking scores from LeDock to estimate target binding affinity. This metric suite indicates that evaluation is not restricted to syntactic validity or token-level reconstruction accuracy but extends to structural novelty, physicochemical property profiles, and in silico target engagement (Song et al., 25 Sep 2025).

The study also includes a patient-oriented case study using three TCGA cases with transcriptomic data and treatment histories. Tumor RNA-seq is matched with GTEx normal tissue controls, differential expression is computed, and signatures are inverted to simulate therapeutic reversal. Molecules are then generated for scaffold optimization and side-chain optimization (Song et al., 25 Sep 2025).

5. Reported empirical results

For scaffold optimization across the ten evaluated targets, PhenoMoler reports validity generally above 90%, uniqueness at 100%, and novelty often above 80%. Table-level values highlighted in the summary include AKT1 with validity 95.86% and novelty 96.0%, EGFR with validity 95.53% and novelty 87.0%, and TP53 with validity 96.19% and novelty 74.52%; all targets are reported to have uniqueness of 100% (Song et al., 25 Sep 2025). These results support the claim that the system generates syntactically valid, structurally novel, and non-duplicative samples.

For scaffold optimization specifically, the paper states that for more than 90% of reference compounds, at least one generated molecule had Tanimoto similarity of 1.0 to the original scaffold, indicating complete scaffold reconstruction. It also reports that more than 80% of generated molecules had novel scaffolds and that most outputs were chemically valid. This suggests that the model can preserve known core motifs while varying surrounding structures (Song et al., 25 Sep 2025).

Drug-likeness is evaluated using QED. The reported finding is that the QED distributions of generated molecules closely matched those of known ligands, indicating preserved drug-likeness. The paper further notes that TP53-generated molecules had significantly higher QED than known ligands in some comparisons. Relative to FDA-approved drugs, the abstract states that generated compounds exhibited comparable or enhanced QED, optimized physicochemical properties, and superior binding affinity to key cancer targets (Song et al., 25 Sep 2025).

For target affinity estimation, LeDock docking scores are used. The paper reports that several generated molecules achieved better predicted binding affinity than reference ligands, with EGFR, mTOR, and TP53 highlighted as examples. Because these are docking-based results, they constitute in silico evidence rather than experimental binding measurements, but they are presented as support for the proposition that phenotype-conditioned optimization can produce molecules with competitive target interactions (Song et al., 25 Sep 2025).

Side-chain optimization is reported to improve physicochemical properties relative to reference ligands. Specifically, generated molecules showed lower MolWt, often moving into the 250–400 Da range; LogP shifted toward the 1–3 range; HBD and HBA remained in recommended ranges; and HBA moved toward 0–2. The paper states that these differences were statistically significant under a Wilcoxon rank-sum test with $\mathbb{R}^n$5. For linker reconstruction, MolWt converged around the ideal $\mathbb{R}^n$6 Da, LogP decreased, and TPSA declined. PCA is reported to show that generated molecules overlapped with reference drug space while covering a broader region. This suggests preservation of a drug-like neighborhood alongside exploration of additional chemical space (Song et al., 25 Sep 2025).

In the patient-oriented case study, generated molecules for each of the three TCGA cases were reported to be more similar to effective drugs than to ineffective ones. The paper presents this as evidence for potential utility in patient-specific phenotype-aware drug design (Song et al., 25 Sep 2025).

6. Limitations, interpretation, and significance

The paper’s explicit claims are primarily supported by in silico evaluation, and several limitations are implicit in that setup. The phenotype-to-structure mapping remains indirect: transcriptomic signatures are informative, but they do not uniquely determine molecular structure. Docking scores are computational proxies rather than experimental measurements of binding or efficacy. The evaluation is mostly in silico, so biological validation remains necessary. The approach also depends on the availability and quality of expression profiles and reference perturbation data, and the discussion does not deeply address generalization beyond the selected target set or the specific cell-line settings used in evaluation (Song et al., 25 Sep 2025).

These caveats are relevant to a common misconception that phenotype conditioning alone can identify a unique or experimentally validated therapeutic structure. The reported results do not establish such uniqueness, nor do they replace wet-lab validation. Instead, the framework demonstrates that phenotype-guided conditioning can be used to generate chemically valid and diverse candidates with favorable property distributions and improved docking scores under the reported benchmarks.

Within those bounds, PhenoMoler’s significance lies in its integration of a chemistry LLM with transcriptional phenotype conditioning and its support for controllable substructure-wise optimization. The paper identifies five main contributions: a multimodal generative framework combining a chemistry LLM with transcriptional phenotype conditioning; substructure-wise controllable generation for scaffold, linker, and side-chain editing; phenotype-guided molecular optimization yielding valid, novel, diverse, and drug-like molecules; evidence of improved target affinity and physicochemical properties, especially for cancer-relevant proteins; and a patient-phenotype case study suggestive of personalized therapeutic molecule design (Song et al., 25 Sep 2025).

Taken together, these elements define PhenoMoler as a phenotype-conditioned, structure-controllable molecular generator that connects gene expression response to chemical substructure design. This suggests a molecular optimization paradigm in which transcriptomic perturbation signatures serve as direct conditioning signals for localized chemical editing, thereby situating phenotype-aware generation as a complement to conventional target-centric drug design.