🧬 biology

Generative Molecular Design for Senescence-Associated Secretory Phenotype (SASP) Reversal: A REINVENT4 Pipeline Trained on Real LINCS L1000 Transcriptomic Data

This study presents a REINVENT4 pipeline trained on real LINCS L1000 transcriptomic data that successfully generates novel, drug-like molecules predicted to reverse the senescence-associated secretory phenotype (SASP), while also identifying and resolving a critical scaffold-collapse failure mode through diversity filtering and reward tuning.

Original authors: Selahattin Alperen Uysal

Published 2026-09-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Selahattin Alperen Uysal

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

As we grow older, our cells do not simply fade away; many of them enter a state of permanent pause. They stop dividing but refuse to die, lingering in the body like silent, stubborn tenants. While these dormant cells are harmless in small numbers, they accumulate over time and begin to secrete a toxic cocktail of inflammatory signals. This secretion, known as the senescence-associated secretory phenotype, or SASP, acts like a slow-burning fire, damaging the healthy tissue around it and driving the aging process as well as age-related diseases. Scientists have long sought a way to reverse this damage, hoping to find a molecule that could silence these inflammatory signals and restore the body's balance. The challenge has been finding the right chemical key to unlock this specific problem without causing harm elsewhere.

A researcher at Ege University has taken a bold step toward solving this puzzle by teaching a computer to design new molecules from scratch. Instead of testing thousands of existing chemicals in a lab, they used a powerful artificial intelligence system to imagine and build entirely new structures. The goal was to create compounds that could specifically reverse the toxic signature of these aging cells. To train this digital architect, the researcher fed it a massive library of real-world data: the effects of over 21,000 different compounds on human cells, recorded in a public database known as LINCS. They focused the computer's attention on a specific list of 106 genes that act as a fingerprint for cellular aging. The computer's job was to learn the relationship between a molecule's shape and its ability to turn down the volume on these aging genes.

The process began with a significant hurdle. The researcher first tried to use an existing, pre-trained computer model that was supposed to predict how chemicals affect genes. However, when they tested this model against known drugs that are already famous for fighting aging, it failed completely. The model incorrectly predicted that these helpful drugs would make aging worse. This failure was a crucial discovery in itself; it showed that relying on a generic, pre-made tool could lead scientists down the wrong path. Instead, the researcher built their own custom prediction tool, or "oracle," trained directly on the real experimental data. This new tool learned to recognize which chemical features were associated with reversing the aging signature, creating a reliable guide for the design process.

With this custom guide in place, the researcher launched a reinforcement learning pipeline, a method where the computer generates millions of candidate molecules and learns from its own successes and failures. The system was rewarded for creating molecules that looked like real drugs, were chemically novel, and could be synthesized in a lab, while also scoring high on their ability to reverse the aging signature. In the first attempt, the computer fell into a trap. It discovered that a very simple, common chemical structure was easy to generate and received a decent score, so it began producing thousands of nearly identical copies of this single structure. This is known as "mode collapse," where the system stops exploring and gets stuck on one easy answer. The researcher realized this error and adjusted the system's rules, tightening the requirements for diversity and lowering the pressure to find a perfect score immediately.

In the second, corrected run, the computer explored a much wider landscape of chemical possibilities. It generated over 92,000 unique, valid molecules, with the vast majority having distinct chemical backbones rather than repeating the same pattern. The results were promising: the average score of these new, computer-designed molecules was significantly higher than the average of the entire library of 21,000 real compounds used for training. This suggests the system did not just memorize existing drugs but actually discovered new chemical patterns that the data suggested would be effective. Among the top candidates, a specific structural motif—a ring-like shape with attached chemical groups—appeared frequently, yet this specific shape was almost entirely absent from the original training library. This indicates the system found a genuine new direction for drug design rather than simply copying what it had seen before.

The study concludes that it is possible to use public data and open-source software to design new molecules that target the specific biology of aging. The researcher has provided a complete, reproducible pipeline that others can use to tackle similar problems. However, they are careful to note that these are virtual candidates; no physical synthesis or lab testing has been performed yet. The molecules exist only as computer models, and their ability to actually reverse aging in living organisms remains to be proven. The work serves as a proof of concept, demonstrating that by combining real biological data with generative design, scientists can navigate the vast chemical space to find novel solutions for complex diseases, provided they remain vigilant against the pitfalls of automated design.

Technical Summary: Generative Molecular Design for SASP Reversal Using REINVENT4 and LINCS L1000 Data

Problem Statement
Cellular senescence contributes to aging and age-related diseases largely through the Senescence-Associated Secretory Phenotype (SASP), a pro-inflammatory secretory program. The "SenMayo" signature, a 125-gene transcriptomic profile, serves as a validated marker for identifying senescent cells and SASP activity. While generative molecular design guided by transcriptomic reversal oracles has been validated for other diseases (e.g., hepatocellular carcinoma), it has not been applied to senescence. The challenge lies in generating novel, drug-like molecules that can reverse the SenMayo signature without relying on proprietary models or new experimental data, while avoiding common pitfalls in reinforcement learning (RL) such as mode collapse.

Methodology
The study employs a multi-stage pipeline integrating real transcriptomic data with open-source generative AI tools:

Data Curation and Scoring:
- Source Data: The pipeline utilizes the LINCS L1000 Level-5 transcriptomic dataset (21,299 compounds, 312,438 signatures).
- Target Signature: 106 of the 125 SenMayo genes were mapped to the LINCS dataset (106 measured or Best-Inferred Genes).
- Reversal Score Calculation: A compound-specific reversal score was computed as the mean z-score of the 106 SenMayo genes. This was standardized against a null distribution generated by 2,000 random 106-gene subsets from the background pool. Positive values indicate down-regulation (reversal) of the SASP signature.
- Validation: The scoring system was validated against known senolytics/senomorphics (e.g., dasatinib, navitoclax) and mechanism-of-action annotations. Notably, a pre-trained transcriptomic predictor (GPS) failed a positive-control check, systematically predicting up-regulation for known reversers, necessitating the use of a custom, real-data-derived scoring approach.
Structure-to-Property Bridge Model:
- Since the generative agent (REINVENT4) produces molecules not present in the LINCS database, a surrogate model was trained to predict reversal scores based solely on molecular structure.
- Features: Morgan count fingerprints (radius 2, 2048 bits) concatenated with 10 physicochemical descriptors.
- Algorithm: Random Forest regression was selected over Ridge regression based on scaffold-split validation (Pearson $r = 0.263$ ), ensuring the model generalizes to unseen chemotypes rather than memorizing training data.
Generative Pipeline (REINVENT4):
- Reward Function: A multi-objective geometric mean reward was constructed combining:
  - The SenMayo reversal oracle (weight 0.5).
  - Quantitative Estimate of Drug-likeness (QED) (weight 0.2).
  - Novelty (inverse Tanimoto similarity to LINCS training set) (weight 0.15).
  - Synthetic Accessibility (Ertl score) (weight 0.15).
- Training Strategy: Two reinforcement learning runs were conducted using the Direct Augmented Prior (DAP) strategy.
  - Run 1: Served as a pilot using default diversity filters. It suffered from "mode collapse," converging on a single trivial scaffold (unsubstituted benzene) despite the diversity filter.
  - Run 2: Addressed the collapse by tightening the diversity filter (reducing bucket size from 25 to 5), lowering reward sharpness ( $\sigma$ from 128 to 100), and restarting from the unbiased prior.

Key Results

Oracle Performance: The Random Forest bridge model achieved a scaffold-split Pearson correlation of $r = 0.263$ against the real LINCS data, a statistically significant signal ( $p < 0.001$ ) despite the inherent difficulty of structure-transcriptome mapping.
Mode Collapse Resolution: Run 1 resulted in 57.2% of converged molecules sharing a single benzene scaffold. Run 2 successfully eliminated this, producing a converged set where 88.4% of molecules possessed distinct Murcko scaffolds, with no single scaffold exceeding a 0.03% share.
Performance Improvement: The molecules generated in Run 2 showed a mean predicted SASP-reversal score of +0.748, representing an improvement of 0.61 standard deviations over the average score of the entire LINCS training library.
Novelty and Structure: The generated set occupied a distinct region of chemical space (UMAP) with low nearest-neighbor similarity (mean Tanimoto = 0.179) to the LINCS library.
De Novo Discovery: A specific substructure, a thiazole-dicarboxamide motif, was identified in the top candidates. This motif was absent from the LINCS training library and Run 1, appearing in only 1.7% of the full Run 2 set but rising to 60% of the top-10 candidates, suggesting the model learned a genuine structure-activity relationship rather than memorizing existing data.
Multi-Objective Balance: Decomposition of the top 10 candidates confirmed that high reversal scores were achieved without sacrificing drug-likeness (QED 0.57–0.71), novelty, or synthetic accessibility.

Significance and Claims
The author claims this work represents the first application of generative molecular design guided by a transcriptomic signature-reversal oracle specifically to the SenMayo senescence signature. The study emphasizes two primary contributions:

Methodological Proof of Concept: It demonstrates that a fully open-source pipeline, using only public data (LINCS L1000) and tools (REINVENT4), can generate novel, drug-like candidates predicted to reverse a complex disease-associated transcriptional phenotype.
Practical RL Insight: The paper documents a concrete failure mode in RL-based molecular generation—silent mode collapse onto trivial scaffolds that persists even when reward curves appear to improve. It provides a practical resolution (tightening diversity filters and adjusting reward sharpness) for researchers applying similar pipelines.

Limitations
The author explicitly notes that the structure-transcriptome relationship was learned from immortalized cancer cell lines (LINCS), and its transferability to primary or senescent cells remains unverified. The bridge model's accuracy ( $r \approx 0.26$ ) sets a modest ceiling on prediction certainty. Furthermore, no wet-lab validation or chemical synthesis has been performed; all reported compounds are virtual candidates.

Technical Summary: Generative Molecular Design for SASP Reversal Using REINVENT4 and LINCS L1000 Data

More like this