1. Abstract and Research Context

The pertTF model is a native, transformer-based artificial intelligence architecture, designed from the ground up as a multimodal tool for single-cell transcriptomics analysis, or scRNA-seq. The overarching goal of this computational environment is the precise prediction of multidimensional cellular responses to genetic inactivation.

The "3Ps" Paradigm in scRNA-seq Transcriptomic Analysis

The foundation of the model's predictive architecture is an innovative approach defined as the "3Ps" (Prediction Framework). Thanks to this, the system goes beyond classical, one-dimensional mRNA level estimation, offering parallel modeling across several planes.

The application of the 3Ps framework allows for the simultaneous prediction of:

Changes in the gene expression profile (gene expression) in response to a given perturbation.
Shifts in the spatial location (spatial location) of cellular structures.
The direct impact of introduced mutations on the cell's differentiation trajectory (differentiation trajectory).

The "3Ps" prediction framework signifies a complete departure from simple transcriptome assessment; pertTF concurrently and multidimensionally predicts changes in expression, spatial location, and differentiation trajectories.

2. Architecture of the pertTF Solution

Transformers in the Latent Space and the GEPC Module

Unlike standard NLP models, which by default utilize mean squared error (MSE) in their loss functions, the pertTF engineers had to confront the specificity of biological data. Raw transcriptomic data is characterized by massive dispersion and a high level of noise, resulting primarily from technical read drop-outs, referred to in bioinformatics jargon as the "drop-outs" phenomenon.

To solve the problem of zero counts and background noise, a dedicated GEPC (Gene Expression Prediction guided by Cell embedding) module was implemented in the architecture. It constitutes a key element connecting the vector representation of cells with the final prediction of the molecular profile.

Rigorous Loss Function Optimization: NB-NLL instead of Classical MSE

The GEPC module diametrically changes the mathematical approach to network optimization. It completely replaces classical MSE errors in favor of a rigorous optimization of the negative binomial negative log-likelihood – in short, NB-NLL.

The implementation of the NB-NLL loss function naturally and extremely precisely decodes the statistical variance of molecular reads, outclassing standard MSE metrics and bringing a dramatic improvement for genes with high dispersion.

Integration of GNN (GEARS) for Extrapolating "Unseen" Genes

A critical challenge for in silico models is predicting the effects of de novo gene knockouts, i.e., those that did not appear in the training data corpus. To enable the model to generalize in this way, the pertTF architecture integrates a specialized module utilizing the GEARS algorithm, based on Graph Neural Networks (GNN).

The inference process for new genetic targets runs on two tracks:

The GEARS algorithm projects "unseen" genes into vectors, utilizing graph representations of relationships from structured Gene Ontology (GO) databases.
The graph representations generated in this way are then combined with cell vectors inside fully connected neural network layers (FCNN).
The fusion of this data in FCNN layers ultimately allows extrapolating the global behavior and state of a new cell after a virtual perturbation, achieving Zero-Shot inference capability.

3. Hard Data, Performance Metrics (Hard Metrics)

Training Set and Absolute Dominance in 8 Metrics

The pertTF architecture underwent rigorous testing on unified transcriptomic datasets, encompassing thousands of unique genetic perturbations (including silencing with CRISPR/Cas9 systems). In a direct clash with existing SOTA (State-of-the-Art) models, the model proved its superiority in comparative evaluation.

In rigorous tests, the algorithm dominated the competition in 8 out of 8 key metrics assessing gene expression prediction. The application of the aforementioned NB-NLL loss function brought measurable, quantified benefits:

A 400% increase in prediction accuracy for genes with low baseline expression, which in classical architectures were lost in background noise.
A global area under the curve (AUC) of 0.79 in cell state classification tasks following complex perturbations.
A 47% reduction in latent space reconstruction error compared to classical variational autoencoders (VAE).

The lochNESS Cellular Identity Metric in Latent Space

To reliably assess whether the model actually decodes biology and does not merely fit statistics, researchers implemented an advanced metric for evaluating distance in multidimensional space – lochNESS. It is used to quantify the preservation of cellular identity in the generated latent space.

The lochNESS metric ruthlessly verifies the model's ability to maintain phenotypic consistency; pertTF achieves results 32% higher here than the competing scGen model, proving a deep understanding of "hidden cellular contexts."

"Zero-Shot" Efficacy and Virtual In Silico Screens

The true test for AI in drug design is the ability to generalize to OOD (Out-of-Distribution) data. Thanks to integration with the GEARS module, pertTF demonstrates an unprecedented capability for Zero-Shot inference. This means it can flawlessly predict the effects of knocking out a gene that the network was never shown during the training phase. The effectiveness of this extrapolation means that virtual in silico screens can replace weeks of costly in vitro laboratory research.

4. Bottlenecks and Limitations

The Sparsity Phenomenon and Single GPU Scaling

Despite spectacular results, the architecture is not free from engineering bottlenecks. The biggest problem remains the computational handling of sparse matrices (the sparsity phenomenon), typical for scRNA-seq data, where even 90% of the matrix can be zeros. The massive memory requirements of attention mechanisms in transformers make training the full model using the GEPC module extremely difficult to scale efficiently on a single GPU cluster without aggressive parameter quantization.

Vulnerability to Graph Database Errors (GNN Dependency)

The model's ability for the aforementioned Zero-Shot generalization is entirely dependent on the quality of input data from the GNN module. If the relationships in the Gene Ontology (GO) databases, from which GEARS draws knowledge about connections between genes, are incomplete or erroneous, the model inevitably propagates these errors into the vector space. This vulnerability to external knowledge base error poses a significant risk in research on rare, poorly mapped signaling pathways.

Nonlinearity of Complex Multigene Perturbations (Epistasis)

The ultimate and most difficult barrier to overcome in deploying these models for virtual clinical trials remains the phenomenon of epistasis. This is a complex biological interaction in which the effect of one gene depends on the presence of others.

Full modeling of complex epistatic mechanisms still represents the boundary of AI capabilities; the simultaneous, virtual knockout of five genes often generates a nonlinear effect in pertTF that diametrically differs from the simple, vector sum of individual knockouts.

5. Data Compilation and Reference Sources

The pertTF project is the result of advanced collaboration between leading research institutes, including the University of Maryland School of Medicine (Prof. Wei Li's team), Columbia University Irving Medical Center, and Children's National Hospital. This tool, in accordance with modern standards, has been made available in the spirit of Open Science.

Code Repositories and Open Data

GitHub Repository: The full source code of the model, trained network weights, and evaluation scripts are available at the official address: https://github.com/davidliwei/pertTF.
3Ps Architecture: A detailed discussion of the predictive framework published by the Institute of Metabolism and Integrative Biology (Fudan University): https://imien.fudan.edu.cn/info/1278/1947.htm.

Bibliography and Source Publications

bioRxiv Preprint: The main research manuscript titled "pertTF: context-aware AI modeling for genome-scale and cross-system perturbation prediction" (published: March 12, 2026).
DOI Identifier: 10.64898/2026.03.12.711379 (Direct link to the full PDF file: https://www.biorxiv.org/content/10.64898/2026.03.12.711379v1.full.pdf).