🤖 machine learning

PhAME: Phenotype-Aware Molecular Editing via Latent Diffusion

PhAME is a latent diffusion framework that enables precise, controllable molecular editing by optimizing small-molecule structures for desired phenotypic signatures while maintaining structural proximity to a seed molecule through a novel compositional classifier-free guidance scheme.

Original authors: Łukasz Janisiów, Sebastian Musiał, Bartosz Zieliński, Dawid Rymarczyk, Tomasz Danel

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Łukasz Janisiów, Sebastian Musiał, Bartosz Zieliński, Dawid Rymarczyk, Tomasz Danel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master chef trying to create a new dish. You have a "hit" recipe—a soup that tastes good but isn't quite perfect. You want to tweak it so it has a specific, complex flavor profile (like "spicy and smoky") that matches a famous critic's description, but you don't want to change the recipe so much that it becomes a completely different dish (like turning soup into a cake).

This is the exact problem drug scientists face. They have a "seed" molecule that works a little bit, and they need to edit it to fix its biological effects (the "flavor") without breaking its chemical structure (the "recipe").

Enter PhAME (Phenotype-Aware Molecular Editing). Think of PhAME as a super-smart, AI-powered "recipe editor" that helps you tweak a molecular recipe perfectly.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Tightrope Walk"

Usually, when scientists try to improve a molecule using AI, they face a frustrating trade-off.

If they push the AI too hard to match the desired biological effect (the "phenotype"), the molecule changes so much it loses its original structure and might stop working.
If they tell the AI to keep the structure exactly the same, the molecule doesn't change enough to get the desired biological effect.

It's like trying to walk a tightrope: if you lean too far left (change the structure), you fall. If you lean too far right (keep it the same), you never move forward.

2. The Solution: Two Separate Knobs

PhAME solves this by giving the user two independent control knobs instead of one.

Knob A (The "Flavor" Knob): This controls how strongly the AI tries to match the desired biological effect (like a gene expression profile or cell shape).
Knob B (The "Structure" Knob): This controls how strictly the AI must keep the molecule looking like the original "seed" molecule.

In previous methods, these two goals were tied together with a single knob. If you turned it up to get the flavor, you lost the structure. PhAME lets you turn Knob A up to get the perfect biological effect while keeping Knob B set to keep the structure safe. You have full control over the balance.

3. The "Secret Sauce": The Latent Space

How does PhAME actually do the editing?
Imagine the molecule isn't a physical object, but a point on a giant, invisible map (called a "latent space").

The Seed: The original molecule is a dot on this map.
The Target: The desired biological effect is a different area on the map.
The Process: PhAME takes the seed dot, adds a little bit of "noise" (like shaking the map), and then uses its two knobs to guide the dot toward the target area.
- If you turn the "Structure" knob up, the dot stays close to where it started.
- If you turn the "Flavor" knob up, the dot moves closer to the target area.

Because it works on this map rather than trying to rebuild the molecule from scratch, it ensures the final result is still a valid, usable molecule.

4. What the Paper Claims It Achieved

The authors tested PhAME in three main ways, and it won in all of them:

Tuning Properties: When asked to change a specific chemical property (like how well it dissolves in fat) while keeping the molecule similar to the original, PhAME did a better job than any other existing method. It found the perfect balance between changing the property and keeping the shape.
Finding New Binders: When asked to generate molecules that stick tightly to specific disease-causing proteins (like PARP1 or BRAF), PhAME found more "hits" (successful candidates) than its competitors. It even found molecules that were better than the best ones in its own training library.
Reading the "Cellular Mood": This is the most complex test. The AI was given a "mood profile" of a cell (based on how the cell looks under a microscope or how its genes are behaving) and asked to create a molecule that would cause that specific mood.
- The Result: PhAME was the only method that could reliably create molecules that matched the target "mood" (Mode of Action) without just copying old molecules. It could even take a starting molecule and "edit" it to turn into a known drug (like Aspirin) just by looking at the target cell's biological profile.

Summary

PhAME is a new tool that lets scientists edit drug molecules with surgical precision. Instead of guessing how to balance "changing the effect" vs. "keeping the shape," it gives them two separate controls. This allows them to create new, effective drug candidates that are chemically valid, novel, and perfectly tuned to the biological signals they are targeting.

Technical Summary: PhAME – Phenotype-Aware Molecular Editing via Latent Diffusion

Problem Statement

Small-molecule drug discovery increasingly relies on optimizing candidate molecules against high-dimensional biological signatures (phenotypes), such as cell morphology (Cell Painting) and transcriptomic profiles (L1000), rather than single-target binding assays. The dominant workflow involves Hit-to-Lead (H2L) and Lead Optimization (LO), where an active "seed" molecule must be iteratively refined to improve potency or alter mechanism of action while maintaining structural proximity to the original hit.

Existing generative methods fail to address this specific workflow effectively:

Ligand-centric generators produce valid molecules but lack mechanisms to steer toward complex phenotypic targets or anchor to a specific seed.
Phenotype-conditioned generators (e.g., GxVAE, SmilesGEN) typically generate molecules de novo from a phenotypic target, ignoring the need to start from a known hit.
Standard Classifier-Free Guidance (CFG) in diffusion models uses a single guidance scale, forcing an unavoidable trade-off between matching the target phenotype and preserving the structural similarity of the seed molecule.

Methodology: PhAME Framework

The authors propose PhAME (Phenotype-Aware Molecular Editing), a latent diffusion framework that reformulates molecular optimization as an editing task within the latent space of a pretrained graph-based Variational Autoencoder (VAE).

1. Latent Space and Architecture

Base Model: PhAME operates on the latent space of DGVAE [34], a pretrained graph-based VAE that encodes molecular graphs into continuous latent vectors ( $z_0$ ) and decodes them back to valid SMILES strings. The encoder and decoder are frozen; only the diffusion model is trained.
Phenotype-Aware Embedding: To anchor the generation to a seed molecule while respecting phenotypic relevance, PhAME utilizes InfoAlign [24] embeddings. Unlike raw chemical fingerprints, InfoAlign embeddings align molecular structure with cellular response profiles (Cell Painting and L1000) via an information bottleneck. This embedding ( $z_{align}$ ) serves as a structural anchor that reflects phenotypically relevant similarity.

2. Training Strategy

Data Construction: Training pairs are constructed by selecting structurally similar molecules that fall on opposite sides of a property threshold (e.g., binder vs. non-binder, or high vs. low logP).
Objective: The denoising network $\epsilon_\theta$ $ϵ_{θ}$ is trained to predict noise added to the latent representation of a target molecule ( $z_0$ $z_{0}$ ) conditioned on:
1. The target phenotypic signature ( $c$ ).
2. The InfoAlign embedding of the seed molecule ( $z_{align}$ ).
Loss Function: The training objective combines the standard diffusion loss ( $L_{diff}$ ) with an alignment loss ( $L_{align}$ ) that penalizes the denoised sample for drifting too far from the seed's InfoAlign embedding, ensuring structural coherence.

3. Compositional Classifier-Free Guidance (c-CFG)

The core innovation of PhAME is the replacement of standard CFG with a compositional guidance scheme. Instead of a single guidance scale $w$ , PhAME decomposes the guidance signal into two independent terms with separate scales:
$\tilde{\epsilon}_\theta(\hat{z}_t, t, c, z_{align}) = \epsilon_\theta(\hat{z}_t, t, \emptyset, \emptyset) + w_c \Delta_c + w_a \Delta_a$
Where:

$\Delta_c$ is the gradient direction toward the target phenotype ( $c$ ).
$\Delta_a$ is the gradient direction toward the seed structure ( $z_{align}$ ).
$w_c$ controls the strength of phenotype matching.
$w_a$ controls the strength of structural anchoring.

This decoupling allows practitioners to explicitly control the trade-off between optimizing for a new biological property and maintaining structural similarity to the seed, removing the forced compromise inherent in single-scale CFG.

4. Inference Modes

Molecular Editing: A seed molecule is encoded, partially noised (controlling the exploration radius), and then denoised using the dual-guidance scheme to generate analogs.
De Novo Generation: The model can sample from pure noise ( $\hat{z}_T \sim \mathcal{N}(0, I)$ ) with $w_a=0$ , functioning as a standard phenotype-conditioned generator.

Key Contributions

Latent-Diffusion Molecular Editing: Introduction of a framework that enables systematic Hit-to-Lead optimization guided by complex biological signatures within a graph-VAE latent space.
Compositional Guidance Mechanism: A novel generalization of CFG that utilizes two independent scales ( $w_c$ and $w_a$ ) to independently control phenotype matching and seed proximity, effectively resolving the trade-off between these objectives.
State-of-the-Art Performance: Demonstration of superior results across diverse modalities, including docking score optimization, transcriptomic-conditioned generation, and Cell Painting-driven Mode of Action (MoA) generation, while maintaining high chemical validity and novelty.

Experimental Results

The authors evaluated PhAME across four distinct benchmarks:

1. Controlled Property Editing (logP)

Task: Optimize logP values while preserving structural similarity to seed molecules on the MOSES benchmark.
Result: PhAME achieved the highest NTS score (a composite metric of Novelty, Target success, and Similarity), outperforming baselines like Mol-CycleGAN, DiGress, and PURE. It demonstrated that independent tuning of $w_c$ and $w_a$ allows the model to match the specific strengths of specialized methods (e.g., high target success like DiGress or high similarity like PURE) simultaneously.

2. Docking Score Optimization

Task: Generate novel high-affinity binders for five protein targets (PARP1, FA7, 5HT1B, BRAF, JAK2) using a three-stage curriculum learning approach on ZINC250k.
Result: PhAME set new state-of-the-art results, achieving the best docking scores on all five targets and the highest novel hit ratio on four of the five targets. Notably, PhAME's top 5% of generated molecules were more potent than the top 5% of the entire ZINC250k training set.

3. Transcriptomic-Guided Generation

Task: Generate molecules conditioned on gene expression profiles (L1000) for specific targets (e.g., AKT1, HDAC1).
Result: PhAME outperformed gene-conditioned baselines (TRIOMPHE, GxVAE, SmilesGEN) in drug-likeness (QED) and synthetic accessibility (SA). Crucially, it generated compounds with significantly higher structural similarity to known binders (up to +90% improvement for HDAC1) compared to baselines, demonstrating effective recovery of known pharmacophores using only expression data.

4. Cell Phenotype-Guided Generation (MoA)

Task: Generate de novo molecules conditioned on Cell Painting morphology centroids corresponding to specific Modes of Action (MoA).
Result: PhAME significantly outperformed baselines (DGVAE, CPMolGAN, GFlowNet) in Top-1 Cluster Accuracy and MoA Retrieval Rate across multiple representation spaces (ECFP, CLOOME, InfoAlign). It was the only method to consistently clear the chance baseline across all spaces, indicating genuine learning of structure-activity relationships rather than overfitting to a specific embedding.

Significance and Claims

The paper claims that PhAME addresses a critical bottleneck in modern drug discovery: the inability of current generative models to simultaneously edit a known hit and steer it toward complex, system-level biological signals. By recasting optimization as latent space editing and introducing compositional guidance, PhAME provides a controllable mechanism to balance the conflicting objectives of structural conservation and phenotypic optimization.

The authors assert that this framework enables a more efficient Hit-to-Lead process, allowing researchers to explore chemical space around known active compounds while targeting specific biological outcomes (e.g., gene knockdowns or morphological changes) without relying on single-target binding assays. The results suggest that PhAME can recover known therapeutics from seed molecules using only phenotypic guidance, highlighting its potential for drug repurposing and lead optimization.