Comparison with SOTA methods (Features)

The LPI-deepGBDT [1] was designed for lncRNA-protein relation prediction. The input is the RNA/protein sequence, while our input contains no protein sequence. However, we checked their GitHub repo and found that their methods provide a feature extractor. Thus, we leveraged their feature extractors for sequence and applied their features in our framework/ and GAE-LGA [2] for more evaluation.

GAE-LGA [2] , a recently published paper (27 October 2022). In their manuscript, multi-omics features for lncRNAs and PCGs were collected. Then, they fed those features into a framework that first calculated similarities between those nodes and then fed into a graph-autoencoder for potential lncRNA-PCG relations prediction.

For the comparison, as LPI-DeepGBDT provided a feature extractor function, thus we leveraged their feature extractors in our frameworks. For GAE-LGA, we replaced their features（Since GAE-LGA were using the PCG whose number is less than ours, we first had to filter the overlap between their method and ours.） with ours and LPI’s generated features and conducted their graph-autoencoder-based framework for performance evaluation.

Methods	Year	J	Codes	Prediction type	Comparison	How (Compare)
LPI‑deepGBDT [1]	2021	BMC bioinformatics	https://github.plhhnu/LPI-deepGBDT	Protein	Yes	Their Feature/Our Framework
GAE-LGA [2]	2022.10	BIB	https://github.com/ meihonggao/GAE-LGA	PCG-Protein coding gene	Yes	Our Feature/LPI [1] feature/ Their Framework

Part I Using different features in our framework

Here we leveraged the features from LPI generating features from the same group of sequences. We applied those features to the downstream framework and calculated AUC/AUPR/F1/MCC for two feature generators. Details can be found in Fig 1. As can be seen in the picture, lncRNA-top-generated features can outperform LPI feature generators in all metrics with p-value <=0.05, indicating the features are better than the SOTA methods generators.

Part II Using Features (LPI’s, ours, and multi-omics features) in the framework of GAE-LGA

  Results of GAE-LGA with different features as input

Method	D1_AUC	D1_AUPR	D1_F1-score	D1_MCC	D2_AUC	D2_AUPR	D2_F1-score	D2_MCC	D3_AUC	D3_AUPR	D3_F1-score	D3_MCC
GAE-LGA_ (Adjust_original features)	0.9522	0.5479	0.7691	0.6071	0.923	0.5984	0.8265	0.6907	0.8149	0.4169	0.7761	0.6007
GAE-LGA_(LPI_Features)	0.9482	0.5407	0.7692	0.6072	0.918	0.5833	0.7901	0.633	0.8404	0.5312	0.8172	0.6524
GAE-LGA_(Our_Features)	0.9489	0.5453	0.7726	0.6124	0.9242	0.6176	0.834	0.7019	0.8568	0.5683	0.8359	0.6809
Increment (%)	-0.346	-0.476	0.4551	0.8730	0.1300	3.2085	0.9074	1.6215	5.142	36.315	7.7051	13.351

We conducted the experiments again in the adjusted network (that removed all non-overlapped lncRNA/gene and corresponding rows and columns) to acquire a benchmark value of GAE-LGA, named as GAE-LGA (Adjust_original features). Then we replaced the original mult-omics features with  LPI  [2] feaures and our feaures (lncRNA-Top features). The results shown that our features can outperformed most of the metrics (10 out of 12 metrics). Notably, as more lncRNA is involved in the dataset( from 117 to 155 to 193), the performance increment increases correspondingly.

Dataset Details After filtered

Dataset	lnc	overlap_lnc	PCG (Protein coding gene)	overlap_gene
Dataset1	208	117	256	211
Dataset2	238	155	716	617
Dataset3	263	193	498	425

Ref:

[1] Zhou L, Wang Z, Tian X, et al. LPI-deepGBDT: a multiple-layer deep framework based on gradient boosting decision trees for lncRNA–protein interaction identification[J]. BMC bioinformatics, 2021, 22(1): 1-24.

[2] Gao M, Liu S, Qi Y, et al. GAE-LGA: integration of multi-omics data with graph autoencoders to identify lncRNA–PCG associations[J]. Briefings in Bioinformatics, 2022, 23(6): bbac452.