Just a reminder that people in that situation can catch up via the News in Brief.
-
New Forum Software Installed (Still a few issues but reopening the forum) More details.
Preprint Dissecting the genetic complexity of myalgic encephalomyelitis/chronic fatigue syndrome via deep learning-powered genome analysis, 2025, Zhang+
- Thread starter SNT Gatchaman
- Start date
We should keep a wide purview. While many of these genes are flagged as neural and relating to synaptic function, they can have important non-canonical roles. The potential link with autism development is fascinating but autism spectrum disorder has other features beyond the neurodevelopmental, eg gastrointestinal dysfunction. So while their effect on synapse formation and maintenance would be important in the primary neurodevelopmental abnormalities we observe, the very common comorbid problems with gut function might be more due to epithelial tight junctions and barrier integrity than with the gut's neural connections.
We should also consider that genes identified in OI might relate more directly to vascular (esp. endothelial cell) function rather than neuronal synapses. Eg NLGN1 and NLGN2 are expressed in vascular endothelial cells. There's also potentially an endocrine aspect, as NLGN2 is also expressed in pancreatic beta cells (insulin secretion), so maybe there's also a link between the suggested GIP secretion (splanchnic vasodilator) and post-prandial increased POTS/OI symptoms.
Neuroligin 1 Induces Blood Vessel Maturation by Cooperating with the α6 Integrin (2014, Journal of Biological Chemistry)
Modulation of Angiopoietin 2 release from endothelial cells and angiogenesis by the synaptic protein Neuroligin 2 (2018, Biochemical and Biophysical Research Communications)
Altered Pancreatic Islet Function and Morphology in Mice Lacking the Beta-Cell Surface Protein Neuroligin-2 (2013, PLOS ONE)
Worsening Postural Tachycardia Syndrome Is Associated With Increased Glucose-Dependent Insulinotropic Polypeptide Secretion (2022, Hypertension)
After looking at the paper again, I realized I should have done my GSEA analysis using p-values, not attention scores. I don't fully understand their method for interpreting the model, but p-values are the metric they used for choosing the top 115 that they say are the most important. It didn't really change the results much since p-value rankings and attention score rankings are pretty similar. (But an exception, for example, is that HLA-C is ranked 22 using p-values and 2077 using attention scores). Again, very similar results on the Genebass CFS data, but the run with p-values can be seen on the last link here.
But I thought it might also be useful to just see the top ranked clusters using the Zhang HEAL2 list of genes. Basically another way to do it from their method of taking their top ranked 115 genes and seeing what pathways these specific genes are enriched in, without considering ranking. Instead I uploaded all of the genes with their -log10 p values to STRING so that it looks at all of them and weighs them by the metric that is apparently useful for interpreting the most important genes.
Here is the link to the enrichment analysis with many different gene set collections. (The page is pretty slow to load and interact with.) I merged items by similarity greater than 0.5, and filtered by FDR < .001 and enrichment score > 1.0. ("Top of input" for direction is what's interesting since it relates to to genes with low p-values.) Filters can be changed near the bottom of the page.
Here are the top 10 STRING local clusters from its protein-protein interaction network, which I think should be most comparable to the enrichment analysis they did that found synaptic function and proteasomes.
Local Network Cluster (STRING)
There are some other interesting gene sets in other collections, though. For example, in DISEASES, there are the following four that match my filters. This includes COVID and two diseases that may be related to sex differences,. If I filter with a less conservative FDR, then autistic disorder is the highest enriched disease with FDR=.0078.
Disease-gene Associations (DISEASES)
For the "Tissue Expression (TISSUES)" collection, the two gene sets are "autonomic nervous system" and "brain ventricle". For "Human Phenotype (Monarch)", the top ranked phenotype is "myeloid leukemia".
Tissue Expression (TISSUES)
Human Phenotype (Monarch)
But I thought it might also be useful to just see the top ranked clusters using the Zhang HEAL2 list of genes. Basically another way to do it from their method of taking their top ranked 115 genes and seeing what pathways these specific genes are enriched in, without considering ranking. Instead I uploaded all of the genes with their -log10 p values to STRING so that it looks at all of them and weighs them by the metric that is apparently useful for interpreting the most important genes.
Here is the link to the enrichment analysis with many different gene set collections. (The page is pretty slow to load and interact with.) I merged items by similarity greater than 0.5, and filtered by FDR < .001 and enrichment score > 1.0. ("Top of input" for direction is what's interesting since it relates to to genes with low p-values.) Filters can be changed near the bottom of the page.
Here are the top 10 STRING local clusters from its protein-protein interaction network, which I think should be most comparable to the enrichment analysis they did that found synaptic function and proteasomes.
Local Network Cluster (STRING)
There are some other interesting gene sets in other collections, though. For example, in DISEASES, there are the following four that match my filters. This includes COVID and two diseases that may be related to sex differences,. If I filter with a less conservative FDR, then autistic disorder is the highest enriched disease with FDR=.0078.
Disease-gene Associations (DISEASES)
For the "Tissue Expression (TISSUES)" collection, the two gene sets are "autonomic nervous system" and "brain ventricle". For "Human Phenotype (Monarch)", the top ranked phenotype is "myeloid leukemia".
Tissue Expression (TISSUES)
Human Phenotype (Monarch)
tralfamadorian97
Established Member
Just finished a quick read of this paper. Overall, this seems promising. I look forward to seeing if this paper’s findings match those of analyses performed by other groups on different data. And the discussion so far in this thread has been insightful. I wish I had the bioinformatics background to contribute.
I do have a bit of background in deep learning research. Some thoughts from that perspective:
I also just realized that the authors are using a non-standard form of attention. In equation 4, the Sigmoid would be replaced by Softmax in standard attention. I am curious as to what motivated them to make this change.
The authors also say "The source code of HEAL2 has been uploaded in the Supplementary Materials.", but I don't see any source code link on medrxiv. Presumably they plan to add this later.
I think it could be clarifying to explore the code a bit. I'll try to do this once the code is available.
I do have a bit of background in deep learning research. Some thoughts from that perspective:
- In the deep learning literature, when you present a complex model, you are typically asked by reviewers to perform an ablation study. This involves systematically removing different components of your system and reporting how this removal affects the system’s performance. Ablation studies reveal which components of the system are most important. I would be interested in seeing an ablation study for HEAL2. I appreciate that the authors compare HEAL1 to HEAL2, but I would like to see finer granularity. This could involve ablating the various input-preprocessing steps, as well as numerous components of the neural architecture. I expect this would shed useful light on how the model is making its predictions.
- Much of the practical value of this paper comes from the list of “ME/CFS risk genes” that it generates. These risk genes are a consequence of attention scores. The use of attention scores for model explanation is reasonable and fairly standard, though not without some controversy (See here and here). It would be interesting to see whether other model explanation techniques (for example, integrated gradients) produce concordant lists of risk genes. If so, this would increase confidence in the results.
I also just realized that the authors are using a non-standard form of attention. In equation 4, the Sigmoid would be replaced by Softmax in standard attention. I am curious as to what motivated them to make this change.
The authors also say "The source code of HEAL2 has been uploaded in the Supplementary Materials.", but I don't see any source code link on medrxiv. Presumably they plan to add this later.
I think it could be clarifying to explore the code a bit. I'll try to do this once the code is available.
Last edited:
Utsikt
Senior Member (Voting Rights)
Can you give an ELI5 of attention scores?
tralfamadorian97
Established Member
I can certainly try. Let me know if this helps:
- The input to HEAL2 is a bunch of statistics describing a person’s genome.
- HEAL2 operates on this input gene by gene.
- That is: it performs one set of neural computations on gene A related input to produce a gene A output, and another set of neural computations on gene B related input to produce a gene B output, etc.
- (Aside: This is slightly complicated by the graph structure: if gene A is known to interact with gene B according to the STRING database, the gene A neural computation is allowed to influence the gene B neural computation and vice versa)
- That is: it performs one set of neural computations on gene A related input to produce a gene A output, and another set of neural computations on gene B related input to produce a gene B output, etc.
- To predict a person’s ME/CFS status, the gene-specific outputs need to be aggregated.
- Attention is a mechanism of aggregation.
- Attention scores are numbers used as aggregation weights. So all else being equal, if the attention score for gene A is larger than the attention score for gene B, we should expect the gene A output to have more influence on the final prediction than the gene B output.
- Attention scores are dynamic. This means that the attention scores of person X for genes A and B will differ from the attention scores of person Y for genes A and B
- The authors identify ME/CFS risk genes by finding genes for which the attention scores of patients tend to be larger than the attention scores of controls.
Last edited: