🧬 biology

Unifying Disjoint Phenotypic Contexts: A Multimodal Soft Contrastive Approach to Identify DILI Activity Cliffs

The paper introduces Bio $\mathcal{X}$ Mol, a multimodal framework that unifies disjoint phenotypic datasets with molecular structures using a soft contrastive objective to overcome the limitations of traditional methods and significantly improve the identification of drug-induced liver injury (DILI) activity cliffs.

Original authors: Muhammad Arslan Masood, Tianyu Cui, Markus Heinonen, Samuel Kaski

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad Arslan Masood, Tianyu Cui, Markus Heinonen, Samuel Kaski

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

In the high-stakes world of drug discovery, the path from a chemical idea to a life-saving medicine is paved with a single, terrifying obstacle: toxicity. A drug might look perfect in a test tube, binding to its target with precision, only to fail catastrophically when it reaches a human liver. This is the realm of drug-induced liver injury, a leading cause of why promising medicines are pulled from the market or never reach patients at all. The challenge is not just that drugs can be toxic, but that they can be deceptively similar. Scientists often encounter "activity cliffs," a phenomenon where two molecules look almost identical in their chemical structure, yet one is a safe treatment while its near-twin causes severe liver damage. Traditional methods of analyzing these chemicals rely on mapping their structural shapes, but when the shapes are nearly the same, these maps fail to distinguish the safe from the dangerous. To solve this, researchers have begun looking beyond the chemical structure itself, turning instead to the biological fingerprints drugs leave behind: the changes they cause in how cells look and how their genes behave.

A team of researchers at Aalto University and the ELLIS Institute Finland has developed a new approach to navigate these dangerous cliffs, introducing a system they call BioXMol. Their work addresses a critical gap in how computers learn to predict drug safety. While previous attempts to combine chemical data with biological observations existed, they were hampered by two major flaws. First, they required every single drug to have been tested in every type of biological experiment, a condition rarely met in the real world where different research groups generate different types of data for different sets of chemicals. Second, and more subtly, they treated all non-matching drugs as equally different from one another. This assumption ignored the biological reality that two different drugs might still trigger similar responses in a cell, and punishing them as if they were completely unrelated distorted the computer's understanding.

The researchers solved these problems by building a framework that can learn from disjoint sets of data. They combined molecular structures with two massive, independent public datasets: one containing detailed images of how cells change shape when exposed to thousands of different compounds, and another containing gene expression profiles showing how thousands of other compounds alter cellular activity. Crucially, these two datasets did not share the same compounds. Instead of waiting for a perfect overlap, the new system anchored both types of biological data to the chemical structures, allowing it to learn from the union of all available information. This meant the computer could learn the relationship between a drug's shape and its biological effects even if no single drug had been tested in both the imaging and gene-expression experiments.

To refine this learning, the team replaced the standard "hard" way of teaching computers to distinguish between items with a "soft" approach. In a typical training session, a computer is told that if two items do not match, they are completely different. The new method, however, uses a more nuanced strategy. It recognizes that while two drugs might not be the same, they might still be biologically similar. By using a momentum-based system that acts as a stable guide, the model weights the differences between drugs based on their actual biological similarity. If two different drugs cause similar changes in a cell, the system treats them as less dissimilar than two drugs that cause completely opposite effects. This preserves the continuous gradients of biological response, rather than forcing a binary choice between "same" and "different."

When the researchers tested their system on a set of known activity cliffs—pairs of structurally similar drugs with opposite liver toxicity outcomes—the results were striking. The new soft-contrastive approach correctly ranked the toxic drug as more dangerous than the safe one in 80.7% of the cases. In contrast, a version of the same system trained with the traditional "hard" method, which assumed all non-matching drugs were equally different, performed no better than random chance, getting the ranking right only 51% of the time. Even standard chemical fingerprints, which rely solely on structural shape without any biological context, managed only 60.7% accuracy. The study demonstrated that the improvement came specifically from the way the model handled the differences between drugs, not just from having more data.

The findings reveal a crucial insight about how we evaluate drug safety models. When tested on a broad, general set of chemicals using standard methods, all the models performed similarly, clustering together with no clear winner. It was only when the test was narrowed to the specific, difficult cases of activity cliffs that the true power of the new approach emerged. The researchers found that the standard evaluation methods were effectively hiding the model's ability to distinguish the most clinically dangerous scenarios. By focusing on these edge cases, they showed that the soft, biologically aware approach creates a representation where safe and toxic analogs are separated by a clear, meaningful distance, whereas traditional methods leave them bunched together. This work suggests that to truly predict drug safety, we must move beyond simple structural matching and embrace a more fluid, biologically grounded understanding of how chemicals interact with living systems.

Technical Summary: Unifying Disjoint Phenotypic Contexts via BioXMol

Problem Statement
Predicting Drug-Induced Liver Injury (DILI) remains a critical bottleneck in drug discovery, particularly due to the prevalence of "activity cliffs." These are pairs of structurally similar compounds that exhibit divergent safety profiles (e.g., one is safe, the other hepatotoxic). Traditional structure-based representations, such as Extended-Connectivity Fingerprints (ECFP) or self-supervised language models (e.g., MoLFormer, ChemBERTa), fail to distinguish these cliffs because they rely on the assumption that chemical structure determines biological activity. When structures are nearly identical, these models cannot capture the subtle functional differences required to predict toxicity.

While integrating functional biological readouts (morphology and transcriptomics) offers a solution, current multimodal approaches face two severe limitations:

Data Disjointness: Large-scale phenotypic datasets like JUMP-CP (morphology) and LINCS L1000 (transcriptomics) profile largely non-overlapping sets of compounds. Standard contrastive learning requires complete triplets (molecule + morphology + transcriptomics) for every sample, rendering these datasets unusable in their union.
Hard Negative Assumption: Standard contrastive objectives treat all non-matching pairs as equally dissimilar "hard negatives." This is biologically implausible, as chemically distinct molecules may induce overlapping phenotypic responses. Penalizing these functionally related pairs equally distorts the learned embedding space, preventing the model from learning continuous gradients of biological severity.

Methodology: BioXMol
The authors propose BioXMol, a multimodal representation learning framework designed to overcome these limitations through a three-stage training pipeline:

Disjoint Data Integration: Instead of requiring triplets, BioXMol treats each biological modality as an independent view aligned to a shared molecular anchor. This allows the model to be pretrained on the union of available paired data (molecule–morphology and molecule–transcriptomics) rather than their intersection.
Soft Contrastive Objective with Momentum Teachers: To address the hard negative problem, BioXMol replaces the standard one-hot negative targets with a "soft" contrastive objective.
- Architecture: The framework utilizes modality-specific encoders (for morphology and transcriptomics) and a shared molecular encoder. Crucially, biological encoders are instantiated as both "student" and "teacher" networks. The teacher networks are updated via an Exponential Moving Average (EMA) of the student weights, providing stable, slowly evolving similarity targets.
- Mechanism: In the forward direction (molecule $\to$ biology), the model minimizes the cross-entropy between student predictions and soft targets derived from the teacher's latent space. These targets weight negative repulsion by the underlying biological similarity of the compounds. If two non-matching compounds induce similar phenotypic responses, they are penalized less than those with divergent responses. The backward direction (biology $\to$ molecule) retains a hard identity objective to ensure unique mapping.
Downstream Evaluation: The pretrained molecular encoder is frozen and evaluated via linear probing using an ordinal logistic regression classifier (Frank–Hall decomposition) to predict ordered DILI severity labels (No-DILI, Less-DILI, Most-DILI).

Key Contributions
The paper outlines four primary contributions:

A Multimodal Pretraining Framework: BioXMol successfully aligns molecular structures with morphological (JUMP-CP) and transcriptomic (LINCS L1000) data from disjoint datasets, eliminating the requirement for fully paired triplets.
Soft Contrastive Learning: The introduction of a momentum-based teacher network that weights negative pairs by biological similarity, replacing the rigid assumption that all non-matching compounds are equally dissimilar.
Activity Cliff Evaluation Protocol: The authors establish that standard scaffold-split cross-validation is insufficient for detecting representation differences in toxicity prediction. They introduce a specific evaluation protocol using 15 held-out activity cliff pairs to reveal clinically meaningful distinctions.
Empirical Evidence of Objective Superiority: The study demonstrates that the performance gain on activity cliffs is driven specifically by the soft contrastive objective, not merely by multimodal pretraining. A "hard" contrastive baseline trained on the exact same data and architecture fails to achieve the same results.

Results
The evaluation reveals a stark divergence between global metrics and activity cliff performance:

Global Performance (Cross-Validation): Under standard 5 $\times$ 5 scaffold-split cross-validation, BioXMol-Soft, BioXMol-Hard, and traditional ECFP fingerprints are statistically indistinguishable (p > 0.05) across ROC-AUC, Spearman's $\rho$ , and linear-weighted $\kappa$ . All converge to a narrow performance band, while pretrained language models perform significantly worse.
Activity Cliff Performance: When evaluated on 12 held-out activity cliff pairs (structurally similar compounds with divergent hepatotoxicity):
- BioXMol-Soft achieves 80.7% pairwise ranking accuracy.
- ECFP achieves 60.7%.
- BioXMol-Hard (the same architecture with hard contrastive loss) collapses to 51.0% (random chance).
- Traditional language models also perform near random chance.
Geometric Analysis: PCA projections show that while ECFP embeddings for safe and toxic cliff analogs are nearly superimposed, BioXMol-Soft stretches these distances by 3–4 fold, encoding toxicity-relevant phenotypic differences as geometric displacements.

Significance and Claims
The paper claims that soft phenotypic alignment is essential for accurately identifying toxicity cliffs. The authors argue that aggregate metrics (like global AUC) obscure clinically critical differences because they are dominated by easy, structurally distinct compounds. The "hard" contrastive assumption, which treats all non-matches as equally dissimilar, destroys the directional gradient necessary to distinguish a safe drug from its toxic analog. By weighting negative repulsion based on biological similarity, BioXMol preserves the continuous structure of the biological response space.

The authors conclude that the choice of evaluation protocol is critical: standard cross-validation fails to reveal the superiority of soft contrastive learning, whereas activity cliff evaluation is a necessary complement to standard benchmarks for drug discovery tasks involving subtle structural variations. The work suggests that respecting biological gradients in the contrastive objective is more impactful than the mere integration of multimodal data.

Technical Summary: Unifying Disjoint Phenotypic Contexts via BioXMol

More like this