Incorrect evaluation workflow for essential genes in estimateEssentialGenes

[JHL-452b]

Current behavior:

Recently, I have been working on evaluating essential genes. I've found that there are issues with the current evaluation workflow (also in auto-tasks in github) in estimateEssentialGenes.

ihuman = readYAMLmodel('model/Human-GEM.yml');
taskStruct = parseTaskList('data/metabolicTasks/metabolicTasks_Essential.txt');
[eGenes, INIT_output] = estimateEssentialGenes(ihuman, 'Hart2015_RNAseq.txt', taskStruct);
results = evaluateHart2015Essentiality(eGenes);

I found that the output context-specific models were very strange, with only a small amount of content as you can see below.

cell_type	DLD1	GBM	HCT116	HELA	RPE1
genes	475	475	475	475	475
rxns	250	250	250	250	250
mets	339	339	339	339	339

Further investigation revealed that the reason for this result is due to the fourth parameter useGeneSymbol of the estimateEssentialGenes function defaulting as true, which then converts the genes in the template model into geneSymbol format. However, in reality, the genes in the Hart2015_RNAseq.txt data are in the 'ENSG0000' format, leading to no gene matches and thus no gene expression being detected by default.

So, I manually tried changing the fourth parameter to false, and while the content of the resulting model was much more normal.

cell_type	DLD1	GBM	HCT116	HELA	RPE1
genes	1734	1731	1772	1743	1669
rxns	6870	6265	6888	6902	6097
mets	5680	4986	5649	5665	4845

However, the result of essential gene evaluation turned out to be all zeros because the genes in in Hart2015_TableS2.xlsx (Experimental result) are geneSymbol format. So, I believe that after the model is generated, all genes in the model (include template model) need to be converted into GeneSymbol format before performing the essential gene evaluation.

[JHL-452b]

I've noticed that currently, the gene shortNames in the model and the function replaceGrRules are used to complete the gene ID conversion.

idMapping = [model.genes, model.geneShortNames];
[grRules,genes,rxnGeneMat] = replaceGrRules(model.grRules,idMapping);
model.grRules = grRules;
model.genes = genes;
model.rxnGeneMat = rxnGeneMat;

Currently, one Ensembl ID of a gene in Human-GEM corresponds to one Symbol ID (shortName). However, in reality, one Ensembl ID of a gene might correspond to multiple Symbol IDs. This could lead to a situation where the Symbol ID converted through the aforementioned method does not exist in the Hart2015 experimental dataset. For example:

Ensembl	All_Symbols	Symbol in model(whether in Hart2015)	Symbol (whether in Hart2015)
ENSG00000014257	['ACP3', 'ACP-3', 'ACPP', 'PAP', 'TM-PAP']	ACP3 (No)	ACPP (Yes)
ENSG00000031698	['SARS1', 'SARS', 'SERS']	SARS1 (No)	SARS (Yes)
ENSG00000035687	['ADSS2', 'ADSS']	ADSS2 (No)	ADSS (Yes)
ENSG00000065427	['KARS1', 'DFNB89', 'KARS', 'KARS2']	KARS1 (No)	KARS (Yes)
ENSG00000066379	['POLR1H', 'A12.2', 'HTEX-6', 'hZR14', 'RPA12', 'tctex-6', 'ZNRD1']	POLR1H (No)	ZNRD1 (Yes)
ENSG00000090861	['AARS1', 'AARS', 'AlaRS', 'CMT2N']	AARS1 (No)	AARS (Yes)
ENSG00000104331	['BPNT2', 'FLJ20421', 'gPAPP', 'IMPA3', 'IMPAD1']	BPNT2 (No)	IMPAD1 (Yes)
ENSG00000104522	['GFUS', 'FX', 'P35B', 'SDR4E1', 'TSTA3']	GFUS (No)	TSTA3 (Yes)
ENSG00000106105	['GARS1', 'CMT2D', 'DSMAV', 'GARS', 'GlyRS', 'SMAD1']	GARS1 (No)	GARS (Yes)
ENSG00000110619	['CARS1', 'CARS']	CARS1 (No)	CARS (Yes)
ENSG00000113163	['CERT1', 'CERT', 'COL4A3BP', 'GPBP', 'STARD11']	CERT1 (No)	COL4A3BP (Yes)

These examples could lead to errors in the essentiality assessment of some genes in the model, resulting in inaccurate evaluation of essential genes.

[JHL-452b]

I believe rather than determining which Ensembl ID corresponds to which Symbol ID for each gene, it would be better to convert all Symbol IDs in the Hart2015 experimental dataset directly into Ensembl IDs. This approach not only resolves the issue of incorrect matching mentioned above but also eliminates the need to convert gene IDs within the workflow.

I have organized and prepared the conversion results of the Hart2015 experimental dataset. You can refer to here:
Hart2015_TableS2_add_ID.xlsx

[JHL-452b]

I have already organized the results of the two methods.

First is about changing Ensembl ID to Symbol ID in all tissue-specific models (same as the current workflow)

cellLine	TP	TN	FP	FN	accuracy	sensitivity	specificity	F1	MCC
DLD1	94	2085	132	224	0.8596	0.2956	0.9405	0.3456	0.2744
GBM	84	2053	147	250	0.8433	0.2515	0.9332	0.2973	0.217
HCT116	110	2113	122	245	0.8583	0.3099	0.9454	0.3748	0.3074
HELA	82	2156	154	200	0.8634	0.2908	0.9333	0.3166	0.2426
RPE1	63	2095	166	210	0.8516	0.2308	0.9266	0.251	0.1702
all	40	2308	165	79	0.9059	0.3361	0.9333	0.2469	0.2089

Second is about keeping the Ensembl ID and using the new mapping data.

I have organized and prepared the conversion results of the Hart2015 experimental dataset. You can refer to here:
Hart2015_TableS2_add_ID.xlsx

cellLine	TP	TN	FP	FN	accuracy	sensitivity	specificity	F1	MCC
DLD1	120	2160	140	235	0.8588	0.338	0.9391	0.3902	0.3174
GBM	108	2130	156	261	0.8429	0.2927	0.9318	0.3412	0.2595
HCT116	140	2191	127	258	0.8582	0.3518	0.9452	0.4211	0.3527
HELA	108	2234	164	212	0.8617	0.3375	0.9316	0.3649	0.289
RPE1	87	2167	179	222	0.849	0.2816	0.9237	0.3026	0.2192
all	60	2396	181	81	0.9036	0.4255	0.9298	0.3141	0.2772

Obviously, the second result is superior to the first one. This indicates that the gene Symbol IDs not present in the experimental dataset, as previously described, have a negative impact on the assessment of essential genes by the model. Therefore, I recommend using the Ensembl ID of genes as the standard ID for assessing gene essentiality.

[feiranl]

@johan-gson Johan, Could you check this? Thanks!

[johan-gson]

Hi Jiahao,

Nice work, I'm sure you are right, thank you for a thorough investigation! I have been away from this quite some time, and actually never worked with the Hart dataset. @feiranl, is there any way we can place this new file somewhere where it is accessible and update any code/docs to refer to this file instead?

I also recommend looking at the DepMap data, there are many more cell lines there.