PMCID
string
Title
string
Sentences
string
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We can see that the FID score is then 0.0 - because the distributions are precisely the same.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We have applied Gaussian blur to this example - in the case of blur with σ = 1, the FID score increases to 18.3; in the case of the blur with σ = 3, it’s 61.5, and in the last case, when Gaussian blur is performed with σ = 7, we are getting FID equal to 70.6.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We can clearly see that the blurring increases the FID score significantly while reducing the data quality FID score of the real and augmented data - example of chr8 8 600 000–10 600 000.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The first example is the real data - we compare the real data distribution with the “generated” distribution, which is also the real data.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We can see that the FID score is then 0.0 - because the distributions are precisely the same.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We have applied Gaussian blur to this example - in the case of blur with σ = 1, the FID score increases to 18.3; in the case of the blur with σ = 3, it’s 61.5, and in the last case, when Gaussian blur is performed with σ = 7, we are getting FID equal to 70.6.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We can clearly see that the blurring increases the FID score significantly while reducing the data quality In the case of FID, we deal with two distributions - one real and one generated.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
In the case of the situation where the real distribution is equal to the generated one, we get FID equal to 0.0 - which is the highest possible FID score.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
That is also the case in our example - when we take real Hi-C data and compare it to itself, we get FID equal to 0.0.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The following steps are blurring the Hi-C matrix and comparing the real distribution (composed of the real data), with the blurred matrix.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We have shown 3 examples with σ = (1, 3, 7).
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
In the first case, where the blur is not that intense, as we are using σ = 1, the FID score equals 18.3.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
In the case of the higher blur, with σ = 3, we are getting a much more augmented matrix - and the FID score rises to 61.5.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
In the last example, we have used Gaussian blur with σ = 7, and the FID score is 70.6.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We can clearly see that, indeed, with the higher blurring of the matrix, we are getting a higher FID score.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
That is why, in our study, the goal was to decrease the FID score, that is, to deblur the Hi-C matrix generated by convolutional encoder-decoder architectures.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The Hi-C Diffusion model that we propose is composed of multiple components (see Fig. 2).
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The first part, encoder-decoder architecture, is similar to the current state-of-the-art tools .
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The input to the network is a genomic sequence - one-hot encoded, and the final output is the Hi-C matrix.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We first use an encoder composed of residual 1D convolutions that transfer 1D sequence into latent space; furthermore, we use a transformer encoder to allow the model to learn long-range context.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Such latent representation is then cast into the 2D matrix, and the decoder, composed of 2D convolutions with exponentially growing dilation, produces the final matrix.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The last part of the encoder-decoder architecture is a convolution that transfers the latent space’s final representation into a classic Hi-C matrix of size 256 × 256 with one channel.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The second network, which uses transfer learning (by taking pre-trained encoder-decoder architecture), is the diffusion model.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Based on the previous findings about diffusion networks, we have decided that the input to the network will be residual between the real Hi-C map and the final prediction of the encoder-decoder network.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Then, we apply Gaussian noise to the residual Hi-C map, and the denoising U-Net is trained to predict the noise, making it easy to obtain the actual residual Hi-C (by subtracting predicted noise from the noised residual Hi-C).
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The network takes as input the noised residual (or, in case of inference - the random input) and the latent representation of the Hi-C heatmap predicted by encoder-decoder architecture.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The latent representation stores the knowledge about the sequence and its meaning in the context of the whole genomic window.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
It can be used for multiple downstream tasks.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
In our case, it guides diffusion to create a heat map that is as close as possible; however, it can be easily applied to classification problems as well (see Supplementary Materials).
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Using hidden representation is necessary to guide the diffusion, thus creating a conditional diffusion model.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
For more information on the technical details of the architecture, see Methods and Fig. 2.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Fig. 2The architecture of HiC Diffusion.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The example used to visualise the prediction & real data is chr8, position 8 600 000–10 600 000 The architecture of HiC Diffusion.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The example used to visualise the prediction & real data is chr8, position 8 600 000–10 600 000 In our study, we have decided to use the context of 2,097,152 nucleotides of sequence and predict the same Hi-C region.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
To validate the model thoroughly, we have performed 22-fold cross-validation - creating 22 models, each with one chromosome excluded for testing purposes (see Methods).
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Our primary motivation behind this approach was to ensure that the model works no matter which chromosomes are used for training and which for validation/testing.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
In a standard use case, one model is sufficient for downstream analysis.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We have calculated the Pearson correlation coefficients for each of the examples in the testing set (for each model), SCC (Stratum-adjusted correlation coefficient), and the FID score.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
To compare ourselves with the current state-of-the-art tool, C.Origami, we have trained the model ourselves, according to the authors’ recommendations; we have used two approaches - in one, we have excluded the epigenetic signal that they used - to keep the results consistent with our findings (see Methods), and second one, with epigenetic signal that they presented as the final model.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We calculated and compared the Pearson correlation coefficients (as well as SCCs) to our model.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The results are consistent and very similar - as in our work, we were aiming to obtain similar metrics in terms of correlation, in our case, even more challenging, i.e. without epigenomic profiles used as an additional input apart from the DNA sequence.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Then, we calculated the FID scores for all the datasets obtained using HiCDiffusion and C.Origami.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We have obtained an average improvement of FID score by 12 times in case of comparison between sequence-only models - with the highest improvement in chr7 (by 88 times).
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
In case of comparison of our sequence-only model to C.Origami enhanced with epigenetics, the average improvement of FID score was by 11 times, and the highest improvement was also obtained in chr11 (by 56 times).
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The visualisation of those results can be seen in Fig. 3, and detailed per-chromosome statistics can be found in Supplementary Figs. 1–3.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Fig. 3In the upper part of the figure, an example output of the 3 models is presented - full HiCDiffusion, HiCDiffusion - only encoder and decoder, and C.Origami (version with epigenetics and sequence, and version with only sequence), along with the real Hi-C matrix.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
All example heatmaps are chr8, position 21 100 000–23 100 000.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The lower part of the chart presents the Pearson correlation coefficient, based on data from all chromosomes, the stratum-adjusted correlation coefficient (SCC) - calculated in similar way, and an average FID obtained by the models (the average is taken from the per-chromosome metric) In the upper part of the figure, an example output of the 3 models is presented - full HiCDiffusion, HiCDiffusion - only encoder and decoder, and C.Origami (version with epigenetics and sequence, and version with only sequence), along with the real Hi-C matrix.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
All example heatmaps are chr8, position 21 100 000–23 100 000.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The lower part of the chart presents the Pearson correlation coefficient, based on data from all chromosomes, the stratum-adjusted correlation coefficient (SCC) - calculated in similar way, and an average FID obtained by the models (the average is taken from the per-chromosome metric) The model was further tested in downstream analysis tasks.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The first one was TAD calling.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The obtained insulation scores were very similar to those observed in the experimental data.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The average Pearson correlation score for the insulation (predicted/real) was 0.63 for chromosome 8.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The full results from the analysis, with the violin plot of the correlations, as well as two examples, can be seen in Fig. 4.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
In the presented examples, we can see that the algorithm correctly finds TAD boundaries in both cases - however, in the case of the real data, in chr8:130 380 000-132 380 000, we detect two TAD boundaries - which are very close to each other and are seen as one in the predicted data.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
More sophisticated TAD calling algorithms could also be applied to the output of the model, provided the algorithms work per heatmap (as the scope of interest in case of the predicted data is 2Mbps - and we are sliding across the diagonal for the prediction of each next heatmap).
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Fig. 4TAD calling analysis in chromosome 8.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The Pearson correlation coefficient between predicted and real insulation scores averaged 0.63.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The two examples show TAD boundaries (orange squares) in chr8:20 100 000–22 100 000 and chr8:88 880 000–90 880 000 TAD calling analysis in chromosome 8.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The Pearson correlation coefficient between predicted and real insulation scores averaged 0.63.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The two examples show TAD boundaries (orange squares) in chr8:20 100 000–22 100 000 and chr8:88 880 000–90 880 000 Another analysis, as presented in Fig. 5, showed that it is possible to call loops from the predicted heatmaps.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
However, for the HiCDiffusion model, that task is much harder - as we are dealing with the sequence-only model, and epigenetic tracks that can indicate very strongly looping factors (e.g., CTCF) are not present in the input data.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Fig. 5Loops called on example regions - Chr8:29 100 000–31 100 000 and chr8:123 380 000-125 380 000 - on the real and predicted values Loops called on example regions - Chr8:29 100 000–31 100 000 and chr8:123 380 000-125 380 000 - on the real and predicted values To see the behaviour of the model in the case of the mutations, we have simulated the event of transduplication of a region resembling a small TAD.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
To do so, we have used the model to predict the genomic window of chr8:32 1000 000–34 100 000 - firstly, using wild type sequence, and secondly, applying a transduplication of chr8:33 450 000–33 650 000 to chr8:32 750 000–32 950 000.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The input to the model in the case of the wild type was the raw sequence, and in the case of the perturbed experiment - the sequence at chr8:32 750 000–32 950 000 was replaced with the one present at chr8:33 450 000–33 650 000.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The results of such an operation can be seen in Fig. 6 (see Supplementary Materials for comparison with C.Origami model), where we can see that the change not only modified the region directly affected but also isolated the top-left corner more (which is shown as the increase of contacts within the top-left corner TAD, as well as decrease of contacts downstream from it), and created a new TAD in the middle of the heatmap.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Fig. 6Modelled transduplication of chr8:32 100 000–34 100 000 region.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The first heatmap is raw model output - with no changes in the sequence.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The second heatmap shows the output of the model with chr8:32 750 000–32 950 000 replaced by a sequence of chr8:33 450 000–33 650 000 Modelled transduplication of chr8:32 100 000–34 100 000 region.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The first heatmap is raw model output - with no changes in the sequence.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The second heatmap shows the output of the model with chr8:32 750 000–32 950 000 replaced by a sequence of chr8:33 450 000–33 650 000 The final analysis that we have undertaken was the question - how well would our method perform on different organisms?
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
To answer this question, we have taken a B cell derived cell line from a mouse , and tested our model on all the chromosomes.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We have found that even if it’s a different organism, the strength of the model persisted - we obtained a genome-wide Pearson correlation score of 0.847 and stratum-adjusted correlation coefficient of 0.761.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The cumulative results of the analysis, as well as an example heatmap, can be seen in Fig. 7.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
For detailed per-chromosome results, see Supplementary Fig. 4.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Fig. 7Results of applying the human model to the mouse data.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Pearson correlation coefficient and SCC were calculated for all heatmaps from all chromosomes.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The example shows chr1:29 500 000–31 500 000 from mouse predicted by the human model Results of applying the human model to the mouse data.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Pearson correlation coefficient and SCC were calculated for all heatmaps from all chromosomes.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The example shows chr1:29 500 000–31 500 000 from mouse predicted by the human model The first step of processing the data is creating the genomic windows analysed in the study.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The sliding window that is used for the processing of the chromosomes is set to 500kbp.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We load the reference genome (GRCh38) and use pyranges to exclude telomeres and centromeres from the analysis.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Then, the sequence is onehot encoded.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The Hi-C matrices used in this research are taken from C.Origami paper - we used GM12878 cell line to allow us to compare our findings with that current state-of-the-art tool.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We take precisely 2,097,152 base pairs of sequence and predict the Hi-C matrix of the same region (resized to 256 × 256 region) - the resolutions were chosen to easily and straightforwardly use convolutions in the network architecture.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
To show the predictive power of the method, we divided the dataset into training (used for training), validation (used for choosing the best models - both encoder/decoder and diffusion), and testing (separate, used only for final testing) datasets.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
This division ensures that the deep learning model is generalising well and that we are unbiased toward examples occurring in the training data.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
To test the model entirely, we decided to create 22 models - in each, the training, validation, and testing datasets are composed of different chromosomes.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Such an approach allows us to be sure that no matter which chromosomes we use for training/validation/testing, the model is still trained properly and maintains its generalisation power.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
For a standard use that does not require such thorough testing, one model is entirely sufficient.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
For each case, we take chromosome i as the testing chromosome, chromosome i + 1 as the validation, and the remaining chromosomes compose the training dataset.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
In the case of testing the last chromosome, chr22, the validation chromosome is chr21.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
We excluded sex chromosomes from the analysis.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The architecture of the model (see Fig. 2) uses the concept of transfer learning.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Firstly, we use encoder-decoder architecture very similar to the ones previously published - e.g. in C.Origami or Akita .
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The encoder first converts the 1D genomic sequence into a sequence of 256, with 256 channels.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
That is done using 13 residual blocks, out of which each is composed of convolution (converting input channels into output channels, with the kernel of size 3 and padding of size 1), batch normalisation, ReLU function, another convolution (this time preserving the number of channels, with kernel of size 3, and padding of size 1), batch normalisation, and maxpooling.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
Additionally, the initial data provided to the residual block is downscaled using convolution (converting input channels directly into output channels, with the kernel of size 3, and padding of size 1).
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
That downscaled data is added to the result from the previously explained sequence of transformations.
PMC11481779
HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences
The final step of the residual block is applying the ReLU function to the output.