Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

This study demonstrates that the quality of both molecular and image data, particularly regarding sparsity, noise, and resolution, is a critical determinant of deep learning performance in predicting spatial transcriptomics from histology images, suggesting that improving data quality offers an essential, orthogonal strategy to model architecture optimization.

The Gist

Imagine you are trying to teach a computer to be a super-detective. Its job is to look at a black-and-white photograph of a city (a tissue slide) and guess exactly what kind of shops, restaurants, and offices are inside every single building (predicting gene expression).

In the real world, to know what's inside a building, you usually have to go inside and take a census. In biology, this "census" is called Spatial Transcriptomics. It tells scientists exactly which genes are active in specific spots of a tissue sample. But here's the catch: doing this census is incredibly expensive and slow, like hiring a team of surveyors to visit every house in a city.

On the other hand, taking a photograph of the city (a Histology Image) is cheap, fast, and routine.

The big idea in this paper is: Can we train a computer to look at the cheap photo and accurately guess the expensive census data?

The authors of this paper wanted to know: Does the quality of the "census data" we use to train the computer matter more than the computer's own intelligence?

Here is the breakdown of their findings, using simple analogies:

1. The Two Types of "Census Data"

The researchers compared two different ways of gathering the "census" data to train their AI:

• The "Blurry, Noisy" Method (Visium): Imagine a surveyor who stands far away and tries to guess what's in a building by looking through a foggy window. They might miss some details (sparsity) or hear things that aren't there (noise). This is cheaper but less accurate.
• The "High-Def, Clear" Method (Xenium): Imagine a surveyor who walks right up to the door, looks inside with a magnifying glass, and counts every single person perfectly. This is expensive but very high quality.

The Discovery: When they trained their AI detective using the "High-Def" data, it became a much better detective than when they trained it on the "Blurry" data. The AI learned about 38% better just by having better training examples.

2. The "Garbage In, Garbage Out" Experiment

To prove that the data quality was the secret sauce and not just the AI's architecture, they ran some clever experiments:

• The "Fog Machine" Test (Molecular Sparsity & Noise): They took the perfect "High-Def" data and artificially added fog and noise to it, making it look like the "Blurry" data.
  • Result: The AI's performance dropped immediately. It was like taking a genius student and forcing them to study with a textbook that had half the pages torn out and random words scribbled over the rest. They couldn't perform well.
• The "Magic Eraser" Test (Imputation): They tried to fix the "Blurry" data using a computer program that guesses the missing parts (like a "Magic Eraser" that fills in the torn pages).
  • Result: The AI got slightly better on the test it knew, but when they gave it a new city to solve, it failed miserably. The "Magic Eraser" had filled in the blanks with guesses that were wrong, teaching the AI bad habits. It couldn't generalize.

3. The "Blurry Camera" Test (Image Quality)

The researchers also looked at the photos themselves.

• They took the high-resolution photos and blurred them to simulate a low-quality camera.
• Result: The AI got worse. But more importantly, they looked at where the AI was looking (using a tool called Grad-CAM).
  • With a clear photo, the AI focused on specific details like cell nuclei (the "rooms" inside the buildings).
  • With a blurred photo, the AI got confused and started looking at random background noise. It lost its ability to understand the structure of the tissue.

4. The "Different Cities" Test

They repeated these experiments on a different type of tissue (colon cancer) with different technologies. The result was the same: Better data quality always led to better predictions.

The Big Takeaway

For years, scientists have been trying to build "smarter" AI models (better detectives) to solve this problem. They've been tweaking the code, changing the algorithms, and building more complex neural networks.

This paper says: "Stop trying to build a smarter detective. Instead, give the current detective a better textbook."

The Analogy:
If you want a student to pass a math test, you can either:

1. Hire a genius tutor who uses a broken textbook full of typos and missing pages (High-tech model, low-quality data).
2. Hire a standard tutor who uses a perfect, clear, and accurate textbook (Standard model, high-quality data).

The authors found that Option 2 wins every time.

Why This Matters

• Cost vs. Quality: High-quality data is expensive. This study suggests that if you want accurate predictions, you can't just cut corners on the data quality and hope a fancy AI will fix it.
• The Future: To make these tools useful in hospitals (where doctors need to trust the AI), we need to prioritize getting the cleanest, highest-quality data possible, rather than just chasing the latest, most complex AI model.

In short: You can't make a great prediction from a blurry picture and a noisy map, no matter how smart your computer is.

Technical Summary

1. Problem Statement

Spatial transcriptomics (ST) technologies allow for high-throughput gene expression quantification with spatial context but are expensive and technically complex. To bypass these costs, deep learning models have been developed to predict spatial gene expression directly from inexpensive, routine Hematoxylin and Eosin (H&E)-stained histology images.

While significant research has focused on optimizing model architectures (e.g., CNNs, Transformers, Graph Neural Networks) to improve prediction accuracy, the impact of training data quality remains largely unexplored. Different ST technologies (e.g., sequencing-based like Visium vs. imaging-based like Xenium) produce molecular data with varying levels of sparsity, noise, and resolution. Similarly, histology images vary in resolution, staining protocols, and scanner optics. The authors investigate whether these data quality variations, rather than model architecture, are the primary drivers of predictive performance differences.

2. Methodology

The study employs a rigorous, controlled experimental framework using in silico ablation and rescue experiments to isolate specific data quality factors.

• Datasets:
  • Primary Dataset: Paired serial tissue sections of breast cancer tissue profiled by 10x Visium (sequencing-based) and 10x Xenium (imaging-based).
  • Validation Dataset: Colon adenocarcinoma (COAD) serial sections profiled by Xenium 5K, VisiumHD, and CosMx 6K.
• Preprocessing & Alignment:
  • Used STalign to co-register Visium and Xenium histology images into a common coordinate system.
  • Rasterized gene expression counts to a uniform spatial resolution (matching Visium spot size) to create "patches" for training.
  • Restricted analysis to genes common across platforms (306 genes for breast cancer; 2,506 for COAD).
• Model Architecture:
  • Baseline: A pretrained ResNet50 feature extractor followed by a four-layer Multilayer Perceptron (MLP) to predict patch-level gene expression.
  • Robustness Checks: Tested alternative architectures including the UNI histology foundation model and RedeHist (a single-cell specific model).
• Experimental Design:
  • Cross-Technology Comparison: Trained models on Visium vs. Xenium data to establish baseline performance gaps.
  • Molecular Ablation: Systematically degraded high-quality Xenium data by introducing sparsity (zeroing counts to match Visium levels) and noise (adding Poisson-distributed random fluctuations).
  • Molecular Rescue: Applied imputation algorithms (KNN, MAGIC, SCVI) to low-quality Visium data to see if artificial augmentation could recover performance.
  • Image Ablation: Applied Gaussian blur to high-resolution Xenium images to simulate lower resolution and assessed the impact on prediction and interpretability (via Grad-CAM).
  • Cross-Image Swapping: Trained models on one technology's molecular data using the other technology's histology image to decouple image vs. molecular effects.

3. Key Contributions

• Data-Centric Insight: Demonstrates that training data quality is a more significant determinant of performance than model architecture in ST prediction tasks.
• Quantification of Sparsity and Noise: Provides empirical evidence that increased molecular sparsity and noise (characteristic of sequencing-based ST) directly degrade prediction accuracy, while high-resolution imaging-based data (Xenium) yields superior results.
• Limitations of Imputation: Shows that while imputation methods can artificially boost performance on held-out test sets, they fail to generalize to independent biological replicates, indicating they introduce bias rather than true signal recovery.
• Interpretability Impact: Reveals that reduced image resolution not only lowers accuracy but also degrades model interpretability, causing attention maps (Grad-CAM) to lose alignment with biologically relevant cellular structures.
• Generalizability Framework: Validates these findings across different tissue types (breast vs. colon) and multiple ST platforms (Visium, Xenium, CosMx, VisiumHD).

4. Key Results

• Performance Gap: Models trained on Xenium data achieved significantly higher prediction accuracy than those trained on Visium data.
  • Mean Pearson Correlation Coefficient (PCC): 0.715 (Xenium) vs. 0.519 (Visium), representing a ~38% improvement.
  • This trend held true across ResNet50, UNI, and RedeHist architectures.
• Molecular Drivers:
  • When using the same histology image, Xenium molecular data consistently outperformed Visium data, proving the molecular data quality is the primary driver.
  • Ablation: Introducing Visium-level sparsity and noise into Xenium data caused a steady decline in PCC, recapitulating the lower performance of Visium-trained models.
  • Rescue Failure: Imputation of Visium data improved test set PCC but resulted in poor generalization on an independent Xenium replicate, confirming that imputation cannot substitute for high-quality ground truth.
• Image Drivers:
  • Image resolution impacts performance, but the effect is conditional on molecular data quality. High-resolution images improved performance primarily when paired with high-quality molecular data (Xenium).
  • Blurring: Applying Gaussian blur to Xenium images degraded PCC and caused Grad-CAM heatmaps to become diffuse, losing focus on nuclei and cellular boundaries.
• Cross-Platform Validation (COAD):
  • Xenium 5K outperformed VisiumHD.
  • CosMx 6K showed a distinct, bimodal performance distribution with higher error, attributed to specific image quality issues (pale staining) and molecular noise patterns.

5. Significance and Implications

• Paradigm Shift: The study argues that the field should shift focus from purely architectural innovation to data-centric AI. Improving the quality of training data (via better ST technologies or rigorous curation) offers an "orthogonal strategy" to boosting model performance.
• Technology Selection: Researchers selecting ST technologies for training predictive models should prioritize platforms with lower sparsity, lower noise, and higher resolution (e.g., imaging-based platforms) over high-throughput but noisy sequencing-based platforms if the goal is accurate gene prediction.
• Clinical Relevance: Since deep learning models are increasingly used for clinical inference, the study highlights that poor data quality (e.g., low-resolution scans or noisy molecular data) can lead to unreliable predictions and misinterpretation of biological features.
• Future Directions: The authors suggest that future work must address off-target probe binding artifacts and develop robust correction strategies. They also emphasize the need for holistic approaches that integrate better data generation with advanced modeling to ensure safe and reliable clinical applications.

In conclusion, this paper establishes that data quality is the bottleneck in spatial gene expression prediction from histology images, and that artificial augmentation (imputation) cannot fully compensate for the inherent limitations of lower-quality ST technologies.