💻 bioinformatics

Chemi-Proteome Language Attention Network Empowers Fragment-Based Ligand Interactome and Binding Sites Discovery with Evidence

The paper introduces C-PLANK, a deep learning framework trained on cellular chemoproteomics data that outperforms existing models in predicting fragment-protein interactions and successfully identified a novel SIRT3 agonist by integrating physicochemical embeddings with a systems-level biological plausibility metric.

Original authors: Liao, B., He, J., zhao, M., Cui, X., Cui, Y., Dong, C., Sun, H., Zhang, L., Zhang, J.

Published 2026-08-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Liao, B., He, J., zhao, M., Cui, X., Cui, Y., Dong, C., Sun, H., Zhang, L., Zhang, J.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Finding a new medicine often feels like searching for a key that fits a specific lock, but the search happens in a crowded room filled with thousands of other keys and locks. Scientists have long relied on computer programs to predict which chemical keys might fit into protein locks, hoping to speed up the discovery of treatments for diseases. These programs are usually trained on data from simple laboratory tests where chemicals interact with isolated proteins in a test tube. While this data is useful, it misses a crucial detail: in the human body, proteins do not exist in isolation. They are part of a complex, living network where they interact with many other molecules, and their behavior changes depending on the cellular environment. Because current computer models are disconnected from this living context, they often struggle to predict how a drug will actually behave inside a cell, leading to failures later in the development process.

To bridge this gap, a team of researchers has developed a new artificial intelligence framework called C-PLANK. Unlike previous models that learn from static, isolated experiments, this system is trained directly on data generated from living cells. The researchers used a technique called chemoproteomics, which involves introducing tiny chemical fragments into cells to see which proteins they stick to. These fragments are small, simple pieces of molecules that can bind to proteins, but because the binding is often weak and fleeting, capturing these interactions requires a special approach. The team used probes that could permanently tag the proteins they touched, allowing scientists to map out exactly where and how these small molecules interacted with the vast network of proteins inside the cell. By feeding this rich, cellular-level data into their AI, the researchers created a model that understands not just whether a chemical and a protein might interact, but how that interaction fits into the broader landscape of the cell.

The core of this new system is its ability to look at both the big picture and the fine details simultaneously. It analyzes the overall environment of the cell while also zooming in to see exactly which parts of a protein and which atoms of a chemical are touching. This dual perspective allows the model to generate a kind of fingerprint for each interaction, showing not only if a connection is likely but also how confident the system is in that prediction. The researchers tested this framework against a large collection of data from eight different studies, involving hundreds of chemicals and thousands of proteins. They found that C-PLANK consistently outperformed existing state-of-the-art models, especially when asked to predict interactions for proteins it had never seen before. This ability to generalize to new targets is critical for drug discovery, where the goal is often to find treatments for proteins that are not yet well understood.

Beyond simply predicting interactions, the system provides a level of transparency that is rare in artificial intelligence. It can point out exactly which parts of a chemical structure and which specific amino acids on a protein are responsible for the binding. The researchers verified this by comparing the model's predictions with physical structures of proteins and known binding sites, finding a strong match between the AI's "fingerprint" and the actual biological reality. This suggests that the model is learning genuine biological rules rather than just memorizing patterns in the data. To prove its practical value, the team used C-PLANK to screen a large library of chemical fragments to find a new way to activate a specific protein called SIRT3, which plays a role in metabolism and aging. The model identified a promising candidate that had been overlooked by other methods.

Following this prediction, the researchers synthesized the chemical and tested it in the lab. The compound successfully bound to the target protein inside living cells and, importantly, increased the protein's activity, acting as an agonist. This initial hit was then refined through a process of chemical optimization, leading to a more potent version of the drug candidate. The final molecule showed strong activity at low concentrations and was confirmed to bind directly to the target protein using biophysical methods. This success story demonstrates that by training artificial intelligence on data that reflects the true complexity of living cells, scientists can move beyond simple predictions and begin to model the dynamic state of biological systems. The work establishes a foundation for future tools that could act as digital twins of human biology, helping researchers navigate the complex journey from a chemical idea to a working medicine with greater confidence and precision.

Technical Summary: Chemi-Proteome Language Attention Network (C-PLANK)

Problem Statement

Current deep learning models for drug discovery are predominantly trained on in vitro affinity datasets (e.g., BindingDB, Davis, BioSNAP). While effective for predicting pairwise binding events under controlled biochemical conditions, these models fail to capture the complexity of native cellular environments. They lack the ability to reflect dynamic cellular contexts, including protein complexes, post-translational regulation, and signaling activity. Furthermore, many existing datasets lack experimentally verified negative interactions, often relying on computationally generated negatives that may introduce bias. Consequently, modern AI systems remain disconnected from the biological systems they aim to model, struggling to predict how compounds interact with the proteome within a living cellular context.

Fragment-based ligand discovery (FBLD) offers a pathway to system-aware modeling by using chemically simple fragments to probe the cellular proteome. However, fragment-protein interactions are typically weak and transient, making them difficult to detect with traditional assays. While fully functionalized fragment (FFF) chemoproteomics enables the covalent capture of these transient interactions in living cells, the resulting data is resource-intensive and costly to generate. There is a critical need for computational models capable of learning directly from cellular chemoproteomic perturbations to accelerate discovery and reduce experimental burdens.

Methodology: C-PLANK Architecture

The authors introduce C-PLANK (Chemi-Proteome Language Attention NetworK), a deep learning framework designed to predict and interpret fragment-protein interactions directly from cellular chemoproteomic data. Unlike conventional Drug-Target Interaction (DTI) models, C-PLANK learns from proteome-wide interaction landscapes generated by FFF probes in living cells.

The framework comprises three primary components:

Dual-Channel Encoders:
- Protein Encoder: Utilizes the ESM-2 protein language model to capture evolutionary context, augmented with curated physicochemical descriptors (molecular weight, polarity, hydropathicity, pKa, topology). These are processed through stacked 1D Convolutional Neural Networks (1D-CNNs) to extract both global contextual representations (via max-pooling) and position-resolved feature matrices for residue-level analysis.
- Ligand Encoder: Encodes ligand SMILES strings at the atom level using a learnable embedding matrix for core atoms (C, N, O, F, P, S, Cl, Br, I) and specific physicochemical features (LogP, charges, TPSA, hybridization). These are also processed by 1D-CNNs to extract intra-molecular patterns.
Bilinear Attention Network (BAN) Decoder:
- Instead of standard cross-attention, C-PLANK employs a Bilinear Attention Network to model pairwise relationships between protein residues and ligand atoms.
- The BAN generates an interaction attention map highlighting residue-atom hotspots and a pooled local interaction vector. This allows the model to capture fine-grained, localized interaction evidence (crucial for weak, transient fragment interactions) while integrating it with global system-level context.
Evidential Layer and Cellular Interaction State Index (CISI):
- Evidential Deep Learning (EDL): The model outputs an evidence parameter ( $\alpha$ ) to jointly learn interaction propensity, uncertainty, and biological credibility.
- CISI: A novel systems-level metric defined by the enrichment of a specific fragment-protein interaction relative to the global promiscuity of the ligand and protein across the entire chemoproteomic dataset. CISI contextualizes predictions against the observed cellular interaction landscape, stratifying interactions into high-, medium-, and low-confidence groups based on prior knowledge.
- Loss Function: The objective function combines standard classification loss with a "knowledge loss" term that constrains uncertainty estimates based on CISI groups, ensuring that interactions strongly supported by the cellular landscape are assigned lower uncertainty.

Key Contributions

Cellular Context Modeling: C-PLANK is the first framework to train directly on cellular chemoproteomic datasets (FFF probes) rather than in vitro affinity data, enabling the modeling of interactions within the native cellular interactome.
Interpretable Interaction Fingerprints: By leveraging the BAN architecture, the model generates atom- and residue-level attention scores that identify molecular determinants of interaction states, providing mechanistic insights without requiring explicit 3D structural inputs.
Uncertainty-Aware Inference: The integration of EDL and CISI allows the model to quantify prediction confidence based on both model outputs and the global cellular interaction landscape, distinguishing reliable predictions from artifacts.
Generalization to Unseen Targets: The framework demonstrates robust performance in "cold-protein" settings, successfully inferring interaction states for proteins not present in the training set.

Results

The authors evaluated C-PLANK on a curated dataset of 431 ligand interactomes derived from eight independent chemoproteomic studies, comprising 90,494 interactions between 407 ligands and 2,629 proteins.

Benchmark Performance: Under both random and stringent "cold-protein" splits, C-PLANK consistently outperformed seven state-of-the-art baselines (including HyperAttentionDTI, TransformerCPI, MolTrans, DrugBAN, and GraphBAN). In the cold-protein setting, C-PLANK achieved an AUROC of 0.8106, outperforming the next best model by approximately 1–2%.
Ablation Studies: Removing either global contextual representations or local BAN-derived interaction features significantly degraded performance, confirming the necessity of modeling both system-level context and fine-grained residue-atom interactions. Replacing BAN with Transformer-style cross-attention also reduced performance, highlighting the superiority of bilinear attention for sparse fragment interactions.
Interpretability Validation:
- Chemical Logic: Attention scores correctly prioritized variable fragment moieties and photoreactive diazirine groups over enrichment handles (alkyne), mirroring FFF probe chemistry.
- Structural Alignment: High-attention residues showed substantial overlap with experimentally determined binding sites (Dizco data) and predicted druggable pockets (Fpocket), with the majority located within 10 Å of labeled sites.
- Case Study (RAB14): For the FFF004-RAB14 interaction, C-PLANK correctly identified key residues (K24, I44, G21, F36) consistent with co-crystal structures and docking poses.
Prospective Discovery (SIRT3 Agonist):
- C-PLANK was used to screen a large fragment library for SIRT3 modulators. It identified SH00, a previously unrecognized binder.
- Experimental validation confirmed SH00 binds SIRT3 and acts as an agonist (1.3-fold activation).
- Through iterative structural optimization guided by the model's insights, the team developed SH10, a potent SIRT3 agonist with an EC50 of 9.23 μM and a Kd of 10.9 μM. SH10 demonstrated enhanced deacetylation activity in cellular assays, validating the model's utility in guiding lead optimization.

Significance and Claims

The paper posits that C-PLANK represents a shift from isolated pairwise interaction prediction toward cellular interaction-state modeling. By learning directly from cellular chemoproteomics, the framework captures the biological plausibility of interactions within the context of the whole proteome.

The authors claim that C-PLANK establishes a computational foundation for future digital-twin frameworks in drug discovery. Unlike current models that predict static binding affinities, C-PLANK reasons about the dynamic biological systems in which interactions occur. The successful identification and optimization of a SIRT3 agonist (SH10) serves as proof-of-concept that learning from cellular interaction landscapes can translate into the discovery of biologically active modulators, bridging the gap between computational prediction and functional cellular response.

The authors acknowledge limitations, noting that the current training data is limited to 407 ligands and relies on 1D sequences due to the scarcity of high-resolution structures. They suggest future work could integrate 3D structural data (e.g., from AlphaFold) and extend the model to handle protein complexes, but maintain that the current framework already provides a robust, interpretable, and generalizable approach for fragment-based discovery in cellular contexts.