💻 bioinformatics

Zero-Shot Protein-Ligand Binding Site Prediction from Protein Sequence and SMILES

This paper introduces a systematic three-stage benchmarking framework that leverages protein and chemical language models to predict protein-ligand binding sites from sequence and SMILES, demonstrating that while ligand conditioning improves performance on known ligands, a zero-shot approach using pretrained chemical embeddings enables effective prediction on previously unseen ligands.

Original authors: Pourmirzaei, M., Alqarghuli, S., Chen, K., Pourmirzaei, M., Xu, D.

Published 2026-11-07 (v2)

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pourmirzaei, M., Alqarghuli, S., Chen, K., Pourmirzaei, M., Xu, D.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Inside every living cell, proteins act as the molecular machines that keep life running. They fold into complex shapes to carry out tasks like copying genetic instructions or fighting off infections. Often, these machines need to grab onto smaller chemical molecules, known as ligands, to function. When a protein grabs a specific molecule, it does so at a precise location on its surface, much like a key fitting into a specific lock. Knowing exactly where this lock is located on the protein is crucial for scientists. If researchers can pinpoint these spots, they can understand how diseases work and design new medicines to either block or activate these interactions. For decades, finding these spots has relied on expensive and slow laboratory experiments that take pictures of the molecules. Now, scientists are turning to computers to predict these locations faster, but a major hurdle remains: computers struggle when they encounter a new type of chemical molecule they have never seen before.

A team of researchers has developed a new way to teach computers to find these binding spots, even when the chemical molecule is completely unfamiliar. They built a system that learns from the vast library of known protein shapes and the chemical formulas of molecules to make educated guesses about where they will connect. The team tested their approach in three different scenarios to see how well it worked. First, they looked at common chemicals that the computer had seen many times before. Second, they tested it on chemicals that appeared only a few times in the training data. Finally, they challenged the system with a zero-shot test, where the computer had to predict binding spots for thousands of chemicals it had never encountered during its training. This final test is the most difficult, mimicking the real-world situation where a scientist discovers a new drug candidate and needs to know how it interacts with a protein immediately.

The researchers started by training a computer model using a massive database of protein sequences and their known binding partners. They first created a baseline system that learned to predict binding spots for one specific chemical at a time. This worked well for the common chemicals but required a separate model for every single type of molecule, which is inefficient. Next, they improved the system by teaching it to recognize different chemicals all at once, using a special code for each one. This allowed a single model to handle many different molecules, and it performed better than the first version. However, this approach still failed when faced with a completely new chemical because the computer did not have a code for it. To solve this, the team introduced a third stage. Instead of using a pre-made code for the chemical, they fed the computer the chemical's structural formula, written as a string of text called a SMILES string. They used a separate computer program, trained on millions of chemical formulas, to turn this text into a mathematical representation that the main system could understand. This allowed the model to make predictions for chemicals it had never seen before.

When the team tested this new zero-shot capability, the results were promising but highlighted the difficulty of the task. On the chemicals the computer had seen before, the system remained accurate. But for the unseen chemicals, the accuracy dropped, which is expected when dealing with entirely new data. The system correctly identified binding spots on about 31 percent of the relevant parts of the protein for these new molecules, a significant achievement for a first attempt at this specific problem. To ensure the computer was actually using the chemical information and not just guessing based on the protein's shape, the researchers ran control experiments. They fed the system the correct protein but replaced the chemical formula with nonsense text or the wrong chemical. In these cases, the computer's performance dropped significantly across all test groups. This proved that the model was indeed paying attention to the specific chemical it was given, rather than relying on a generic guess.

The study also explored how the size of the computer models affected their performance. They tested different versions of the protein-learning model, ranging from small to very large. Generally, the larger models performed better, suggesting that having more computational power helps the system understand the complex patterns of protein shapes. However, making the chemical-learning model larger did not always lead to better results, indicating that the protein side of the equation is currently the more critical factor for accuracy. The researchers found that while the system is not yet perfect, it provides a reliable framework for predicting how proteins and new chemicals might interact. This work establishes a new standard for testing these predictions, distinguishing between what a computer can do with familiar data versus what it can achieve with truly novel discoveries. By proving that the system can generalize to unseen chemicals, the researchers have opened a path for faster drug discovery, allowing scientists to screen new molecules against proteins without needing to wait for slow, traditional experiments.

Technical Summary: Zero-Shot Protein–Ligand Binding Site Prediction from Protein Sequence and SMILES

Problem Statement
Accurate identification of protein–ligand binding sites is critical for mechanistic biology and drug discovery. However, existing computational tools often lack versatility, restricting effectiveness to specific ligand categories, and predictive performance varies significantly across ligand families depending on data availability. While experimental techniques like X-ray crystallography provide structural data, they are labor-intensive, necessitating computational approaches for high-throughput prediction. A key limitation in current research is the lack of systematic evaluation regarding how predictive performance changes across different ligand data-availability regimes: data-rich (overrepresented), data-scarce (underrepresented), and data-absent (zero-shot). Furthermore, distinguishing performance on data-rich ligands from generalization to unseen ligands requires a ligand-aware evaluation framework.

Methodology
The authors present a systematic, ligand-aware benchmark and evaluation framework utilizing a BioLiP2-derived dataset. The dataset was processed using CD-HIT clustering at 40% sequence identity within ligand-specific receptor datasets to reduce redundancy. The evaluation is structured across three distinct regimes:

Overrepresented: Ligands with $\ge$ 100 samples.
Underrepresented: Ligands with 20–99 samples.
Zero-shot: Ligands with 1–19 samples, held out entirely from training and validation.

To assess performance across these regimes, the authors developed a progressive three-stage sequence-based modeling suite:

Stage 1 (Ligand-Specific Baseline): A baseline architecture where a separate model is trained for each ligand type using a pretrained Protein Language Model (PLM, specifically ESM-2) to generate residue-level embeddings. These embeddings are fed into a linear classifier. No ligand-specific input (SMILES or ID) is provided to the model during inference; the model relies solely on the protein sequence and its specific training data.
Stage 2 (Multi-Ligand with Learned Embeddings): A single model is trained to predict binding sites for multiple predefined ligand types. Protein sequences are encoded by the same PLM. Ligand identity is provided as an index querying a trainable embedding table, which is fused with protein features via a bidirectional transformer decoder using cross-attention. This allows the model to learn from a pooled dataset but cannot generalize to unseen ligands.
Stage 3 (Zero-Shot Generalization): The fixed ligand-embedding table is replaced with a pretrained Chemical Language Model (CLM, specifically MolFormer) operating on SMILES strings. The CLM generates ligand representations which are projected to match the PLM feature dimension and fused via the same decoder. This architecture enables prediction for ligand identities absent from the supervised training and validation data.

The study employs rigorous control experiments, including inputs with constant, mismatched, and scrambled SMILES, to verify that the model utilizes specific ligand information rather than generic protein propensity. Additionally, ablation studies were conducted across five scales of ESM-2 (8M to 3B parameters) and multiple CLM configurations (MolFormer, UniMol-2) to analyze the impact of model scale.

Key Contributions

Ligand-Aware Benchmark: The establishment of a standardized evaluation framework with splits for overrepresented, underrepresented, and zero-shot ligands, incorporating within-ligand sequence clustering to systematically assess residue-level prediction across varying data availability.
Progressive Modeling Suite: The development and evaluation of a three-stage modeling progression that characterizes performance as ligand supervision decreases from single-ligand training to multi-ligand conditioning and finally to SMILES-based zero-shot inference.
Controlled Representation Analysis: Systematic assessment of protein and chemical representations through ablations and ligand-input controls (constant, mismatched, scrambled SMILES), demonstrating the contribution of specific ligand inputs to prediction accuracy.

Results

Stage 2 vs. Stage 1: On the overrepresented test set, Stage 2 achieved a higher ligand-macro F1 (0.5832) compared to Stage 1 (0.4769). However, the authors note this comparison involves simultaneous changes in architecture, conditioning, and training regime. A matched ablation on six ligands showed that replacing the linear predictor with the Stage 2 decoder alone yielded only a modest improvement (+0.0136 F1), suggesting the gains are largely driven by multi-ligand training and conditioning.
Stage 3 Performance: The Stage 3 model achieved a pooled residue-level F1 of 0.3109 and a ligand-macro F1 of 0.2338 on previously unseen (zero-shot) ligands. While this represents a modest decrease in performance on represented ligands compared to Stage 2 (e.g., overrepresented ligand-macro F1 dropped from 0.5832 to 0.5526), it demonstrates the capability to predict binding sites for unseen ligand identities.
Control Experiments: Replacing native SMILES with constant, mismatched, or scrambled inputs significantly reduced performance across all three test regimes (e.g., zero-shot pooled F1 dropped from 0.3193 to 0.1335 with constant input). This provides evidence that the model relies on the supplied ligand association rather than solely on generic protein binding-site propensity.
Scaling Analysis: Increasing the PLM scale (ESM-2) generally improved performance across regimes, though the trend was not strictly monotonic. In contrast, increasing the CLM scale (e.g., UniMol-2 84M to 570M) produced smaller and less consistent gains.
Baselines: The Stage 2 model outperformed several sequence- and structure-based baselines (including Boltz-2x, Prot2Token, and LMetalSite) on specific metal ions and cofactors in the overrepresented set.

Significance and Claims
The paper claims to provide a systematic benchmark for quantifying the effect of ligand data availability on residue-level binding-site prediction and for evaluating generalization to previously unseen ligands. The authors emphasize that their zero-shot capability allows for residue-level prediction using only protein sequences and SMILES, without generating 3D complexes.

The study modestly concludes that while the model demonstrates zero-shot generalization, the current architecture relies on a shallow projector for cross-modal fusion, and the limited benefit from scaling CLMs suggests that protein-representation capacity has a stronger effect on performance under the current setup. The authors clarify that the zero-shot setting is defined at the ligand-identity level (ligand IDs absent from training) and does not imply the absence of receptor sequences or receptor families from the training data. The work aims to support reproducible progress in ligand-aware and zero-shot residue-level binding prediction by releasing code, data, and model weights.

Technical Summary: Zero-Shot Protein–Ligand Binding Site Prediction from Protein Sequence and SMILES

More like this