Dmitry Penzar
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Dmitry Penzar has a total of 150 co-authors. Here are the top co-authors by number of shared publications:
- Ivan Vladimirovich Kulakovskiy
Institute of Protein Research, Russian Academy of Sciences, Pushchino, Russia
7 shared publications - Arsenii Zinkevich
Faculty of Bioengineering and Bioinformatics, Lomonosov Moscow State University, Moscow, Russia
5 shared publications - Vsevolod Jurievich Makeev
Vavilov Institute of General Genetics, Russian Academy of Sciences, Moscow, Russia
4 shared publications - Abdul Muntakim Rafi
School of Biomedical Engineering, University of British Columbia, Vancouver, British Columbia, Canada
3 shared publications Andrey Lando
Yandex N.V. , Moscow 119021, Russia
3 shared publications- Anton Buzdin
Institute of Personalized Oncology, I.M. Sechenov First Moscow State Medical University, Moscow 119991, Russia
3 shared publications
Latest Publications
- Open Access
Classic machine learning on top of multiple position weight matrices improves genomic prediction of transcription factor binding sites
Motivation: DNA motifs recognised by transcription factors are typically represented as position weight matrices (PWMs), assuming independent contributions of individual nucleotides to protein binding specificity. Many alternative models accounting for correlations of positional contributions have been introduced in the past decades. However, performance gains have generally not outweighed the advantages of simplicity, interpretability, and practical applicability of PWMs with the well-established codebase. Existing software tools and motif databases provide multiple non-identical PWMs for the same transcription factor or even for the same dataset. It remains a practical question whether these PWMs can be effectively combined into a single improved model. Results: Here we describe ArChIPelago ( https://github.com/autosome-ru/ArChIPelago ), a computational framework that combines multiple PWMs into a joint model using classic machine learning techniques, from linear regression to ensembles of decision trees. We show that such a combination improves prediction of transcription factor binding sites in genomic sequences. With a diverse collection of 704 ChIP-Seq datasets spanning 36 orthologous human and mouse transcription factors of diverse structural families, we show that ArChIPelago consistently outperforms the best available individual mono- and dinucleotide PWMs as well as sparse local inhomogeneous mixture models. Furthermore, using both human and mouse data, we demonstrate that PWM ensembles are capable of making reliable cross-species predictions.
Open AccessClassic machine learning on top of multiple position weight matrices improves genomic prediction of transcription factor binding sites
Motivation: DNA motifs recognised by transcription factors are typically represented as position weight matrices (PWMs), assuming independent contributions of individual nucleotides to protein binding specificity. Many alternative models accounting for correlations of positional contributions have been introduced in the past decades. However, performance gains have generally not outweighed the advantages of simplicity, interpretability, and practical applicability of PWMs with the well-established codebase. Existing software tools and motif databases provide multiple non-identical PWMs for the same transcription factor or even for the same dataset. It remains a practical question whether these PWMs can be effectively combined into a single improved model. Results: Here we describe ArChIPelago ( https://github.com/autosome-ru/ArChIPelago ), a computational framework that combines multiple PWMs into a joint model using classic machine learning techniques, from linear regression to ensembles of decision trees. We show that such a combination improves prediction of transcription factor binding sites in genomic sequences. With a diverse collection of 704 ChIP-Seq datasets spanning 36 orthologous human and mouse transcription factors of diverse structural families, we show that ArChIPelago consistently outperforms the best available individual mono- and dinucleotide PWMs as well as sparse local inhomogeneous mixture models. Furthermore, using both human and mouse data, we demonstrate that PWM ensembles are capable of making reliable cross-species predictions.
- Open Access
- 7 November 2025
- Published by Springer Nature in Communications Biology
- Vol. 8 (1) , 1-19
Cross-platform motif discovery and benchmarking to explore binding specificities of poorly studied human transcription factors
A sequence motif representing the DNA-binding specificity of a transcription factor (TF) is commonly modelled with a positional weight matrix (PWM). Focusing on understudied human TFs, we processed results of 4,237 experiments for 394 TFs, assayed using five different experimental platforms. By human curation, we approved a subset of experiments that yielded consistent motifs across platforms and replicates, and evaluated quantitatively the cross-platform performance of PWMs obtained with ten motif discovery tools. Notably, nucleotide composition and information content are not correlated with motif performance and do not help in detecting underperformers, while motifs with low information content, in many cases, describe well the binding specificity assessed across different experimental platforms. By combining multiple PMWs into a random forest, we demonstrate the potential of accounting for multiple modes of TF binding. Finally, we present the Codebook Motif Explorer ( https://mex.autosome.org ), cataloguing motifs, benchmarking results, and the underlying experimental data.
- 7 November 2025
- Published by Springer Nature in Communications Biology
- Vol. 8 (1) , 1-19
Open AccessCross-platform motif discovery and benchmarking to explore binding specificities of poorly studied human transcription factors
A sequence motif representing the DNA-binding specificity of a transcription factor (TF) is commonly modelled with a positional weight matrix (PWM). Focusing on understudied human TFs, we processed results of 4,237 experiments for 394 TFs, assayed using five different experimental platforms. By human curation, we approved a subset of experiments that yielded consistent motifs across platforms and replicates, and evaluated quantitatively the cross-platform performance of PWMs obtained with ten motif discovery tools. Notably, nucleotide composition and information content are not correlated with motif performance and do not help in detecting underperformers, while motifs with low information content, in many cases, describe well the binding specificity assessed across different experimental platforms. By combining multiple PMWs into a random forest, we demonstrate the potential of accounting for multiple modes of TF binding. Finally, we present the Codebook Motif Explorer ( https://mex.autosome.org ), cataloguing motifs, benchmarking results, and the underlying experimental data.
- Top Cited PapersOpen Access
- 11 January 2025
- Published by Oxford University Press (OUP) in Nucleic Acids Research
- Vol. 53 (2)
GENA-LM: a family of open-source foundational DNA language models for long sequences
Recent advancements in genomics, propelled by artificial intelligence, have unlocked unprecedented capabilities in interpreting genomic sequences, mitigating the need for exhaustive experimental analysis of complex, intertwined molecular processes inherent in DNA function. A significant challenge, however, resides in accurately decoding genomic sequences, which inherently involves comprehending rich contextual information dispersed across thousands of nucleotides. To address this need, we introduce GENA language model (GENA-LM), a suite of transformer-based foundational DNA language models capable of handling input lengths up to 36 000 base pairs. Notably, integrating the newly developed recurrent memory mechanism allows these models to process even larger DNA segments. We provide pre-trained versions of GENA-LM, including multispecies and taxon-specific models, demonstrating their capability for fine-tuning and addressing a spectrum of complex biological tasks with modest computational demands. While language models have already achieved significant breakthroughs in protein biology, GENA-LM showcases a similarly promising potential for reshaping the landscape of genomics and multi-omics data analysis. All models are publicly available on GitHub (https://github.com/AIRI-Institute/GENA_LM) and on HuggingFace (https://huggingface.co/AIRI-Institute). In addition, we provide a web service (https://dnalm.airi.net/) allowing user-friendly DNA annotation with GENA-LM models.
- 11 January 2025
- Published by Oxford University Press (OUP) in Nucleic Acids Research
- Vol. 53 (2)
Top Cited PapersOpen AccessGENA-LM: a family of open-source foundational DNA language models for long sequences
Recent advancements in genomics, propelled by artificial intelligence, have unlocked unprecedented capabilities in interpreting genomic sequences, mitigating the need for exhaustive experimental analysis of complex, intertwined molecular processes inherent in DNA function. A significant challenge, however, resides in accurately decoding genomic sequences, which inherently involves comprehending rich contextual information dispersed across thousands of nucleotides. To address this need, we introduce GENA language model (GENA-LM), a suite of transformer-based foundational DNA language models capable of handling input lengths up to 36 000 base pairs. Notably, integrating the newly developed recurrent memory mechanism allows these models to process even larger DNA segments. We provide pre-trained versions of GENA-LM, including multispecies and taxon-specific models, demonstrating their capability for fine-tuning and addressing a spectrum of complex biological tasks with modest computational demands. While language models have already achieved significant breakthroughs in protein biology, GENA-LM showcases a similarly promising potential for reshaping the landscape of genomics and multi-omics data analysis. All models are publicly available on GitHub (https://github.com/AIRI-Institute/GENA_LM) and on HuggingFace (https://huggingface.co/AIRI-Institute). In addition, we provide a web service (https://dnalm.airi.net/) allowing user-friendly DNA annotation with GENA-LM models.
- Open Access
Cross-platform DNA motif discovery and benchmarking to explore binding specificities of poorly studied human transcription factors
A DNA sequence pattern, or “motif”, is an essential representation of DNA-binding specificity of a transcription factor (TF). Any particular motif model has potential flaws due to shortcomings of the underlying experimental data and computational motif discovery algorithm. As a part of the Codebook/GRECO-BIT initiative, here we evaluated at large scale the cross-platform recognition performance of positional weight matrices (PWMs), which remain popular motif models in many practical applications. We applied ten different DNA motif discovery tools to generate PWMs from the “Codebook” data comprised of 4,237 experiments from five different platforms profiling the DNA-binding specificity of 394 human proteins, focusing on understudied transcription factors of different structural families. For many of the proteins, there was no prior knowledge of a genuine motif. By benchmarking-supported human curation, we constructed an approved subset of experiments comprising about 30% of all experiments and 50% of tested TFs which displayed consistent motifs across platforms and replicates. We present the Codebook Motif Explorer ( https://mex.autosome.org ), a detailed online catalog of DNA motifs, including the top-ranked PWMs, and the underlying source and benchmarking data. We demonstrate that in the case of high-quality experimental data, most of the popular motif discovery tools detect valid motifs and generate PWMs, which perform well both on genomic and synthetic data. Yet, for each of the algorithms, there were problematic combinations of proteins and platforms, and the basic motif properties such as nucleotide composition and information content offered little help in detecting such pitfalls. By combining multiple PMWs in decision trees, we demonstrate how our setup can be readily adapted to train and test binding specificity models more complex than PWMs. Overall, our study provides a rich motif catalog as a solid baseline for advanced models and highlights the power of the multi-platform multi-tool approach for reliable mapping of DNA binding specificities. Abstract Figure: Graphical Abstract
Open AccessCross-platform DNA motif discovery and benchmarking to explore binding specificities of poorly studied human transcription factors
A DNA sequence pattern, or “motif”, is an essential representation of DNA-binding specificity of a transcription factor (TF). Any particular motif model has potential flaws due to shortcomings of the underlying experimental data and computational motif discovery algorithm. As a part of the Codebook/GRECO-BIT initiative, here we evaluated at large scale the cross-platform recognition performance of positional weight matrices (PWMs), which remain popular motif models in many practical applications. We applied ten different DNA motif discovery tools to generate PWMs from the “Codebook” data comprised of 4,237 experiments from five different platforms profiling the DNA-binding specificity of 394 human proteins, focusing on understudied transcription factors of different structural families. For many of the proteins, there was no prior knowledge of a genuine motif. By benchmarking-supported human curation, we constructed an approved subset of experiments comprising about 30% of all experiments and 50% of tested TFs which displayed consistent motifs across platforms and replicates. We present the Codebook Motif Explorer ( https://mex.autosome.org ), a detailed online catalog of DNA motifs, including the top-ranked PWMs, and the underlying source and benchmarking data. We demonstrate that in the case of high-quality experimental data, most of the popular motif discovery tools detect valid motifs and generate PWMs, which perform well both on genomic and synthetic data. Yet, for each of the algorithms, there were problematic combinations of proteins and platforms, and the basic motif properties such as nucleotide composition and information content offered little help in detecting such pitfalls. By combining multiple PMWs in decision trees, we demonstrate how our setup can be readily adapted to train and test binding specificity models more complex than PWMs. Overall, our study provides a rich motif catalog as a solid baseline for advanced models and highlights the power of the multi-platform multi-tool approach for reliable mapping of DNA binding specificities. Abstract Figure: Graphical Abstract
- Open Access
Ribonanza: deep learning of RNA structure through dual crowdsourcing
Prediction of RNA structure from sequence remains an unsolved problem, and progress has been slowed by a paucity of experimental data. Here, we present Ribonanza, a dataset of chemical mapping measurements on two million diverse RNA sequences collected through Eterna and other crowdsourced initiatives. Ribonanza measurements enabled solicitation, training, and prospective evaluation of diverse deep neural networks through a Kaggle challenge, followed by distillation into a single, self-contained model called RibonanzaNet. When fine tuned on auxiliary datasets, RibonanzaNet achieves state-of-the-art performance in modeling experimental sequence dropout, RNA hydrolytic degradation, and RNA secondary structure, with implications for modeling RNA tertiary structure.
Open AccessRibonanza: deep learning of RNA structure through dual crowdsourcing
Prediction of RNA structure from sequence remains an unsolved problem, and progress has been slowed by a paucity of experimental data. Here, we present Ribonanza, a dataset of chemical mapping measurements on two million diverse RNA sequences collected through Eterna and other crowdsourced initiatives. Ribonanza measurements enabled solicitation, training, and prospective evaluation of diverse deep neural networks through a Kaggle challenge, followed by distillation into a single, self-contained model called RibonanzaNet. When fine tuned on auxiliary datasets, RibonanzaNet achieves state-of-the-art performance in modeling experimental sequence dropout, RNA hydrolytic degradation, and RNA secondary structure, with implications for modeling RNA tertiary structure.
- Open Access
- 25 July 2023
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 39 (8)
LegNet: a best-in-class deep learning model for short DNA regulatory regions
Motivation: The increasing volume of data from high-throughput experiments including parallel reporter assays facilitates the development of complex deep-learning approaches for modeling DNA regulatory grammar. Results: Here, we introduce LegNet, an EfficientNetV2-inspired convolutional network for modeling short gene regulatory regions. By approaching the sequence-to-expression regression problem as a soft classification task, LegNet secured first place for the autosome.org team in the DREAM 2022 challenge of predicting gene expression from gigantic parallel reporter assays. Using published data, here, we demonstrate that LegNet outperforms existing models and accurately predicts gene expression per se as well as the effects of single-nucleotide variants. Furthermore, we show how LegNet can be used in a diffusion network manner for the rational design of promoter sequences yielding the desired expression level. Availability and implementation: https://github.com/autosome-ru/LegNet. The GitHub repository includes Jupyter Notebook tutorials and Python scripts under the MIT license to reproduce the results presented in the study.
- 25 July 2023
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 39 (8)
Open AccessLegNet: a best-in-class deep learning model for short DNA regulatory regions
Motivation: The increasing volume of data from high-throughput experiments including parallel reporter assays facilitates the development of complex deep-learning approaches for modeling DNA regulatory grammar. Results: Here, we introduce LegNet, an EfficientNetV2-inspired convolutional network for modeling short gene regulatory regions. By approaching the sequence-to-expression regression problem as a soft classification task, LegNet secured first place for the autosome.org team in the DREAM 2022 challenge of predicting gene expression from gigantic parallel reporter assays. Using published data, here, we demonstrate that LegNet outperforms existing models and accurately predicts gene expression per se as well as the effects of single-nucleotide variants. Furthermore, we show how LegNet can be used in a diffusion network manner for the rational design of promoter sequences yielding the desired expression level. Availability and implementation: https://github.com/autosome-ru/LegNet. The GitHub repository includes Jupyter Notebook tutorials and Python scripts under the MIT license to reproduce the results presented in the study.
- Open Access
GENA-LM: A Family of Open-Source Foundational DNA Language Models for Long Sequences
Recent advancements in genomics, propelled by artificial intelligence, have unlocked unprecedented capabilities in interpreting genomic sequences, mitigating the need for exhaustive experimental analysis of complex, intertwined molecular processes inherent in DNA function. A significant challenge, however, resides in accurately decoding genomic sequences, which inherently involves comprehending rich contextual information dispersed across thousands of nucleotides. To address this need, we introduce GENA-LM, a suite of transformer-based foundational DNA language models capable of handling input lengths up to 36,000 base pairs. Notably, integration of the newly-developed Recurrent Memory mechanism allows these models to process even larger DNA segments. We provide pre-trained versions of GENA-LM, demonstrating their capability for fine-tuning and addressing a spectrum of complex biological tasks with modest computational demands. While language models have already achieved significant breakthroughs in protein biology, GENA-LM showcases a similarly promising potential for reshaping the landscape of genomics and multi-omics data analysis. All models are publicly available on GitHub https://github.com/AIRI-Institute/GENA LM and HuggingFace https://huggingface.co/AIRI-Institute.
Open AccessGENA-LM: A Family of Open-Source Foundational DNA Language Models for Long Sequences
Recent advancements in genomics, propelled by artificial intelligence, have unlocked unprecedented capabilities in interpreting genomic sequences, mitigating the need for exhaustive experimental analysis of complex, intertwined molecular processes inherent in DNA function. A significant challenge, however, resides in accurately decoding genomic sequences, which inherently involves comprehending rich contextual information dispersed across thousands of nucleotides. To address this need, we introduce GENA-LM, a suite of transformer-based foundational DNA language models capable of handling input lengths up to 36,000 base pairs. Notably, integration of the newly-developed Recurrent Memory mechanism allows these models to process even larger DNA segments. We provide pre-trained versions of GENA-LM, demonstrating their capability for fine-tuning and addressing a spectrum of complex biological tasks with modest computational demands. While language models have already achieved significant breakthroughs in protein biology, GENA-LM showcases a similarly promising potential for reshaping the landscape of genomics and multi-omics data analysis. All models are publicly available on GitHub https://github.com/AIRI-Institute/GENA LM and HuggingFace https://huggingface.co/AIRI-Institute.
- Open Access
Evaluation and optimization of sequence-based gene regulatory deep learning models
Neural networks have emerged as immensely powerful tools in predicting functional genomic regions, notably evidenced by recent successes in deciphering gene regulatory logic. However, a systematic evaluation of how model architectures and training strategies impact genomics model performance is lacking. To address this gap, we held a DREAM Challenge where competitors trained models on a dataset of millions of random promoter DNA sequences and corresponding expression levels, experimentally determined in yeast, to best capture the relationship between regulatory DNA and gene expression. For a robust evaluation of the models, we designed a comprehensive suite of benchmarks encompassing various sequence types. While some benchmarks produced similar results across the top-performing models, others differed substantially. All top-performing models used neural networks, but diverged in architectures and novel training strategies, tailored to genomics sequence data. To dissect how architectural and training choices impact performance, we developed the Prix Fixe framework to divide any given model into logically equivalent building blocks. We tested all possible combinations for the top three models and observed performance improvements for each. The DREAM Challenge models not only achieved state-of-the-art results on our comprehensive yeast dataset but also consistently surpassed existing benchmarks on Drosophila and human genomic datasets. Overall, we demonstrate that high-quality gold-standard genomics datasets can drive significant progress in model development.
Open AccessEvaluation and optimization of sequence-based gene regulatory deep learning models
Neural networks have emerged as immensely powerful tools in predicting functional genomic regions, notably evidenced by recent successes in deciphering gene regulatory logic. However, a systematic evaluation of how model architectures and training strategies impact genomics model performance is lacking. To address this gap, we held a DREAM Challenge where competitors trained models on a dataset of millions of random promoter DNA sequences and corresponding expression levels, experimentally determined in yeast, to best capture the relationship between regulatory DNA and gene expression. For a robust evaluation of the models, we designed a comprehensive suite of benchmarks encompassing various sequence types. While some benchmarks produced similar results across the top-performing models, others differed substantially. All top-performing models used neural networks, but diverged in architectures and novel training strategies, tailored to genomics sequence data. To dissect how architectural and training choices impact performance, we developed the Prix Fixe framework to divide any given model into logically equivalent building blocks. We tested all possible combinations for the top three models and observed performance improvements for each. The DREAM Challenge models not only achieved state-of-the-art results on our comprehensive yeast dataset but also consistently surpassed existing benchmarks on Drosophila and human genomic datasets. Overall, we demonstrate that high-quality gold-standard genomics datasets can drive significant progress in model development.
- Open Access
LegNet: a best-in-class deep learning model for short DNA regulatory regions
Motivation: The increasing volume of data from high-throughput experiments including parallel reporter assays facilitates the development of complex deep learning approaches for DNA regulatory grammar. Results: Here we introduce LegNet, an EfficientNetV2-inspired convolutional network for modeling short gene regulatory regions. By approaching the sequence-to-expression regression problem as a soft classification task, LegNet secured first place for the autosome.org team in the DREAM 2022 challenge of predicting gene expression from gigantic parallel reporter assays. Using published data, here we demonstrate that LegNet outperforms existing models and accurately predicts gene expression per se as well as the effects of single-nucleotide variants. Furthermore, we show how LegNet can be used in a diffusion network manner for the rational design of promoter sequences yielding the desired expression level. Availability and Implementation: https://github.com/autosome-ru/LegNet . The GitHub repository includes the Python code under the MIT license to reproduce the results presented in the study and a Jupyter Notebook tutorial. Supplementary Information: Online-only supplementary data are available at Bioinformatics online. Contact: dmitrypenzar1996@gmail.com , ivan.kulakovskiy@gmail.com
Open AccessLegNet: a best-in-class deep learning model for short DNA regulatory regions
Motivation: The increasing volume of data from high-throughput experiments including parallel reporter assays facilitates the development of complex deep learning approaches for DNA regulatory grammar. Results: Here we introduce LegNet, an EfficientNetV2-inspired convolutional network for modeling short gene regulatory regions. By approaching the sequence-to-expression regression problem as a soft classification task, LegNet secured first place for the autosome.org team in the DREAM 2022 challenge of predicting gene expression from gigantic parallel reporter assays. Using published data, here we demonstrate that LegNet outperforms existing models and accurately predicts gene expression per se as well as the effects of single-nucleotide variants. Furthermore, we show how LegNet can be used in a diffusion network manner for the rational design of promoter sequences yielding the desired expression level. Availability and Implementation: https://github.com/autosome-ru/LegNet . The GitHub repository includes the Python code under the MIT license to reproduce the results presented in the study and a Jupyter Notebook tutorial. Supplementary Information: Online-only supplementary data are available at Bioinformatics online. Contact: dmitrypenzar1996@gmail.com , ivan.kulakovskiy@gmail.com
- Open Access
- 31 October 2019
- Published by Frontiers Media SA in Frontiers in Genetics
- Vol. 10, 1078
What Do Neighbors Tell About You: The Local Context of Cis-Regulatory Modules Complicates Prediction of Regulatory Variants
Many problems of modern genetics and functional genomics require the assessment of functional effects of sequence variants, including gene expression changes. Machine learning is considered to be a promising approach for solving this task, but its practical applications remain a challenge due to the insufficient volume and diversity of training data. A promising source of valuable data is a saturation mutagenesis massively parallel reporter assay, which quantitatively measures changes in transcription activity caused by sequence variants. Here we explore the computational predictions of the effects of individual single-nucleotide variants on gene transcription measured in the massively parallel reporter assays, based on the data from the recent "Regulation Saturation" Critical Assessment of Genome Interpretation challenge. We show that the estimated prediction quality strongly depends on the structure of the training and validation data. Particularly, training on the sequence segments located next to the validation data results in the "information leakage" caused by the local context. This information leakage allows reproducing the prediction quality of the best CAGI challenge submissions with a fairly simple machine learning approach, and even obtaining notably better-than-random predictions using irrelevant genomic regions. Validation scenarios preventing such information leakage dramatically reduce the measured prediction quality. The performance at independent regulatory regions, entirely excluded from the training set, appears to be much lower than it is needed for practical applications, and even the performance estimation will become reliable only in the future with richer data from multiple reporters. The source code and data are available at https://bitbucket.org/autosomeru_cagi2018/cagi2018_regsat and https://genomeinterpretation.org/content/expression-variants
- 31 October 2019
- Published by Frontiers Media SA in Frontiers in Genetics
- Vol. 10, 1078
Open AccessWhat Do Neighbors Tell About You: The Local Context of Cis-Regulatory Modules Complicates Prediction of Regulatory Variants
Many problems of modern genetics and functional genomics require the assessment of functional effects of sequence variants, including gene expression changes. Machine learning is considered to be a promising approach for solving this task, but its practical applications remain a challenge due to the insufficient volume and diversity of training data. A promising source of valuable data is a saturation mutagenesis massively parallel reporter assay, which quantitatively measures changes in transcription activity caused by sequence variants. Here we explore the computational predictions of the effects of individual single-nucleotide variants on gene transcription measured in the massively parallel reporter assays, based on the data from the recent "Regulation Saturation" Critical Assessment of Genome Interpretation challenge. We show that the estimated prediction quality strongly depends on the structure of the training and validation data. Particularly, training on the sequence segments located next to the validation data results in the "information leakage" caused by the local context. This information leakage allows reproducing the prediction quality of the best CAGI challenge submissions with a fairly simple machine learning approach, and even obtaining notably better-than-random predictions using irrelevant genomic regions. Validation scenarios preventing such information leakage dramatically reduce the measured prediction quality. The performance at independent regulatory regions, entirely excluded from the training set, appears to be much lower than it is needed for practical applications, and even the performance estimation will become reliable only in the future with richer data from multiple reporters. The source code and data are available at https://bitbucket.org/autosomeru_cagi2018/cagi2018_regsat and https://genomeinterpretation.org/content/expression-variants