Using CHESS, a large set of transcripts assembled from nearly 10,000 human RNA-seq experiments, we identified all protein-coding gene isoforms that were 1000aa or less in length (see Materials and methods). The 233,973 transcripts at 20,666 gene loci that fit this description encoded 127,398 distinct protein isoforms, and we predicted structures for all of them. Additionally, we included 3302 structure predictions for proteins with length >1000aa from the AlphaFold Protein Structure Database (Varadi et al., 2022) for isoforms with an exact protein sequence match in CHESS 3. This resulted in a total of 237,295 transcripts at 20,817 gene loci encoding 130,700 distinct protein isoforms. As shown in Figure 1, we observe no strong trend in the relationship between predicted local distance difference test (pLDDT) and either protein length or overall transcript expression in the GTEx data. The lack of a clear linear relationship implies protein structure prediction may provide an orthogonal source of useful information for genome annotation efforts.

In total, 22,644 transcripts (9.7% of all examined transcripts) encoded an isoform that scored a higher pLDDT than the isoform encoded by the corresponding MANE transcript. However, many of these higher-scoring transcripts encoded relatively short, often low-expressed, likely non-functional fragments of larger proteins. Therefore, we incorporated filters based on RNA-seq expression data when determining which higher-scoring isoforms appeared superior to MANE isoforms (see ‘Filtering MANE comparisons’). Based on the combination of RNA-seq evidence and protein foldability, we identified 940 unique alternate isoforms at 632 loci (3.4% of all MANE loci) which appeared to have a more stable structure than the annotated primary isoform. Data for these 940 alternate isoforms can be found in Supplementary file 3. Worth noting is that over 96% of the MANE loci evaluated here contained no higher-scoring alternate isoforms that passed our filtering criteria. This is a testament to both the high degree of consistency in MANE and the sensitivity of protein structure prediction for finding instances where alternative isoforms may create functional products. Additionally, for 35% of all human protein-coding gene loci in our analysis, the most commonly observed isoform scored a pLDDT below 70, suggesting that intrinsic disorder may be an important feature of proteins at these loci.

Gene identifiers for all predicted protein isoforms as well as pLDDT scores and evolutionary conservation data from mouse can be found in Supplementary file 1. Predicted scores and GTEx expression data for all isoforms overlapping a MANE locus can be found in Supplementary file 2. All predicted protein structures as well as data for all tables in this paper are publicly available at our website, isoform.io.

To illustrate the improvements in human gene annotation that can be obtained using accurate structure prediction, we describe a small set of proteins, selected from Supplementary file 3, where an alternate isoform appears to be superior to the isoform chosen for inclusion in MANE. For these examples, the alternative isoform is clearly functional based on structure as well as evolutionary conservation and, in some cases, additional expression evidence from RNA-seq data. For some of these examples, the MANE isoform is missing critical structural elements and may not be functional at all.

Alternative splicing allows genes to code for multiple functional protein products (Matlin et al., 2005). Thus, rejecting all but the top-scoring isoform based on predicted structure may eliminate lower-scoring yet functional proteins. The risk of discarding functional transcripts by relying too heavily on the pLDDT score is well-illustrated by vascular endothelial growth factor B (VEGFB), a growth factor implicated in cancer and diabetes-related heart disease (Lal et al., 2018). The human VEGFB gene encodes two well-characterized protein isoforms: VEGFB-167 and VEGFB-186. Alternative splicing that skips part of the sixth exon in VEGFB-167 leads to sequestration of the protein to the cell surface due to a highly basic C-terminal heparin binding domain. Full inclusion of exon six in VEGFB-186 results in a soluble protein freely transported to the blood stream (Li, 2010).

Both isoforms shown in Figure 8 represent highly expressed and similarly functional products, containing a well-conserved cysteine-knot motif (Iyer and Acharya, 2011), yet VEGFB-167 receives a pLDDT score of 81.7 while VEGFB-186 receives a much lower pLDDT score of 69.5. The MANE isoform (CHS.9039.1, RefSeq NM_003377.5, GENCODE ENST00000309422.7) encodes the freely soluble protein VEGFB-186, while the alternate isoform (CHS.9039.2, RefSeq NM_001243733.2, GENCODE ENST00000426086.3) encodes the sequestered protein VEGFB-167. Additionally, VEGFB-186 is present as a full-length cDNA clone (MGC:10373 IMAGE:4053976) in the Mammalian Gene Collection (Temple et al., 2009), a fact which strongly supports its functionality. Due to the large pLDDT score difference between the two functional VEGFB isoforms, a naïve attempt to use protein folding prediction scores as the sole oracle of protein function might inadvertently discard VEGFB-186, a clearly functional transcript. Thus, one must be careful to incorporate multiple sources of information when making decisions about isoform functionality.

While examining the evolutionary conservation of the 3D structure of TXNDC8 in human, we noticed that the predicted structure for the same gene in Mus musculus (house mouse) seemed to contain a poorly folded region similar to CHS.56446.6, a low-scoring human protein. Further inspection of TXNDC8 in the mouse genome revealed that the primary RefSeq transcript contains a misfolding fifth exon, while the alternate RefSeq transcript skips the misfolding exon, similar to the functional human isoform, shown in Figure 9. Interestingly, both mouse transcripts contain a third exon homologous to the sequence missing in the human MANE isoform. A BLASTP search of the misfolding exon resulted in only a single significant hit outside the order Rodentia to an unnamed protein product. In contrast, a BLASTP search of the third exon present in all mouse transcripts, homologous to the missing exon in the human MANE transcript, revealed significant hits to at least 31 homologs outside Rodentia.

Unlike the functional human isoform, however, the RefSeq alternate mouse transcript is annotated with a start site 26 codons upstream of the translation initiation site annotated in the human orthologs. A BLASTP search of the 26 amino acid additional sequence resulted in zero significant hits outside Rodentia. Folding this alternate mouse transcript revealed that these additional N-terminal amino acids fail to form any confident predicted protein structure. As a result, none of the three transcripts annotated in mouse fold into the highly conserved structure of TXNDC8 in human. Predicted structures and exon alignments for isoforms in human and mouse are shown in Figure 9.

We hypothesized that a truly functional isoform of TXNDC8 in mouse should be similar to the human isoform in 3D structure. Based on our observations, this similarity might be realized if the mouse alternate isoform simply started at the downstream start site that matches human, yielding a 105aa protein rather than the 131aa protein that is currently annotated. Folding the coding sequence of the alternate mouse transcript, minus the 26 N-terminal amino acid residues, revealed a predicted structure remarkably similar to both human and cattle TXNDC8. Figure 5 shows the 3D alignment of the human and cow proteins to our predicted (105aa) isoform of mouse TXNDC8. Remarkably, the predicted average RMSD between aligned heavy atoms of the putative human and mouse proteins is just 0.83 Å. For reference, the atomic diameter of one carbon atom is 1.4 Å.

In a further investigation of mouse TXNDC8 transcription, we aligned 8 gigabases of RNA-seq cDNA from a mouse testis sample (SRR18337982) to the GRCm38 reference genome using HISAT2 (Kim et al., 2019) then assembled transcripts using StringTie2 (Kovaka et al., 2019). This resulted in two putative transcripts at the mouse TXNDC8 locus, with neither transcript containing the upstream start site present in the RefSeq annotation. Examination of the read coverage confirmed that both putative TXNDC8 transcripts appear to use the start site of our proposed shorter protein-coding sequence, with 1060 reads supporting the canonical start site and zero reads supporting the upstream start site. As hypothesized, one of these newly assembled transcripts contained the protein-coding sequence necessary to exactly match our predicted structure-conserved isoform.

We believe this represents the first experimentally confirmed novel isoform in any organism discovered due to a hypothesis derived from comparison of computationally predicted protein structures. All in all, the structure-guided identification and subsequent experimental confirmation of a novel functional TXNDC8 isoform in mouse demonstrates the potential of 3D protein structure prediction to enhance functional annotation in any genome.

The examples discussed here are only a small subset of the 130,700 unique protein structures we generated at 20,817 human gene loci. These structures and associated pLDDT scores have already been used to avoid filtering out transcripts encoding functional, clinically relevant isoforms while building the latest version of the CHESS human gene catalog. We provide all of these structures as a searchable and downloadable database, at isoform.io, to create a public resource for improving the annotation of the human genome. The large majority of CHESS isoforms in this collection have direct support from RNA-seq data, as they were assembled from the large GTEx collection, a high-quality set of deep RNA-seq experiments across dozens of human tissues. A very small number of MANE proteins were not assembled from GTEx data, but structures of these too are included in this analysis so that no MANE genes would be omitted.

In the online resource, we provide for each transcript: (1) the nucleotide and amino acid sequences; (2) the predicted structure, as a file that can be viewed in a standard structure viewer such as PyMOL (Schrödinger, 2015) (3) the pLDDT score of that structure; (4) the length of the isoform; (5) the number of GTEx samples in which the isoform was observed; (6) the maximum expression of the isoform in any tissue; (7) an indicator based on alignment of whether all introns are conserved in the mouse genome; (8) an interactive table with functionality to search and sort transcripts to find isoforms of interest; and (9) a Foldseek (van Kempen et al., 2022) interface to search any given protein structure against all 237,295 transcripts presented here. These predictions can be mined to discover, for example, cases where a known protein gets a surprisingly low pLDDT score, or where alternative isoforms have structures that get higher scores and appear more stable than previously reported forms of the same protein.

Though the complete sequence of the human genome has been revealed (Nurk et al., 2022), the annotation of the human genome, by far the most comprehensively studied, remains far from finished. The use of accurate predicted protein structures for gene annotation, as we have done here, represents a new paradigm, not only for human gene discovery but for all other species as well. For decades, the scientific community has relied principally on two methods to discover and validate protein-coding genes at the genome scale: sequencing of transcripts (or cDNAs), and alignment of DNA and protein sequences to detect evolutionary conservation in other species. Protein structure holds valuable information regarding biological functionality, providing an independent and powerful tool to complement these methods. The analysis described here takes a first step toward improving genome annotation of humans using structure prediction, but specific methods deploying this powerful new tool on a broader scale will require fundamentally new computational strategies. As such, development of comprehensive genome annotation protocols incorporating protein structure prediction will remain an area of active investigation for years to come.

We folded all transcripts in the CHESS annotation less than 1000 amino acids in length. Similar to the initial effort to fold the human proteome (Tunyasuvunakool et al., 2021), the length limit was chosen to make the overall computational runtime feasible. This yielded 233,973 transcripts representing 127,398 unique protein-coding sequences at 20,666 loci. Coding sequences for CHESS transcripts were determined with ORFanage (Varabyou et al., 2021). For each protein sequence, we generated a multiple sequence alignment by aligning them with ColabFold’s MMseqs2 (Steinegger and Söding, 2017) workflow (colabfold_search) against the UniRef100 (Suzek et al., 2015) (2021/03) and ColabFoldDB (2021/08) database. Structure predictions were made with ColabFold (commit 3398d3) using AlphaFold2 and MMseqs2 version 13.45111. To speed up the search, we set the sensitivity setting to 7 (-s 7). We predicted each structure using colabfold_batch and stopped the process early if a pLDDT of at least 85 was reached by any model (--stop-at-score 85) or if a model produced a pLDDT less than 70 (--stop-at-score-below 70). All models were ranked by pLDDT in descending order. Runtime was estimated from a sample of 500 proteins randomly selected from the 127,398 structures. Prediction of all structures on 8 × A5000 GPUs required 34 days. Multiple sequence alignment took 34 hr using MMseqs2 on an AMD EPYC 7742 CPU with 64 cores.

To generate the 940 protein isoforms in Supplementary file 3, we used the following filtering criteria and procedures. For each isoform with a distinct coding sequence located at a MANE v1.0 locus and with coding sequence overlapping a MANE annotated protein, we compared the pLDDT score to that of the associated MANE protein. As described previously, pLDDT is a reliable measure of the confidence in a structure, where predictions with 70 ≤ pLDDT ≤ 90 are confident, those with pLDDT >90 are highly confident, and those below 50 represent low-confidence structures and may be disordered proteins (Varadi et al., 2022). We only considered alternative isoforms that had a pLDDT score ≥70, indicating a generally well-folded protein, to avoid including any intrinsically disordered proteins (Ruff and Pappu, 2021). Filtering and general analysis was performed in Colab (https://colab.research.google.com) with Python version 3.8.15.

We selected isoforms that, when compared to the MANE isoform for the same gene, scored at least 5% higher as measured by pLDDT and were at least 90% as long. Additionally, to capture cases where the MANE transcript might be missing functional sequence elements, we selected alternate isoforms that were at least 5% longer than the MANE isoform, that had equal or higher pLDDT scores, and that were assembled in an equal or higher number of GTEx samples. Finally, to capture cases where an alternate isoform might be functional despite being substantially shorter than the MANE protein, we selected isoforms at least 50% as long as the MANE protein where the alternate isoform scored at least 5% higher and was assembled in an equal or higher number of GTEx samples. After applying these filters, we observed that in some cases, a processed pseudogene (Zhang et al., 2002) contained within an intron outscored the associated primary transcript. To eliminate such cases, we used GFFcompare (Pertea and Pertea, 2020) to ensure that isoforms overlapped their MANE transcript’s coding sequence. When multiple alternate isoforms contained the same coding sequence, and thus received the same pLDDT score, we selected the isoform assembled in the highest number of GTEx samples.

Transcript annotations for MANE, CHESS, RefSeq, and GENCODE were retrieved from the following sources. The MANE v1.0 database was downloaded from NCBI at https://www.ncbi.nlm.nih.gov/refseq/MANE. Annotations from the CHESS v3.0 database were retrieved from http://ccb.jhu.edu/chess. Additional CHESS annotations came from an unpublished set of transcript assemblies created as part of the process of building CHESS v3.0; these were assembled from approximately 10,000 GTEx RNA-seq experiments across 31 tissues using StringTie2 (Kovaka et al., 2019). Transcripts from this set were given a CHESS ID starting in ‘hypothetical’ if the locus was missing from CHESS v3.0, or else given an ID ending in ‘altN’ if the locus was present in CHESS v3.0 but the exact isoform was not. Note that many of these, particularly those with a poor protein folding score, were not retained in the final CHESS v3.0 database. RefSeq annotations (releases 109 and 110) were downloaded from https://www.ncbi.nlm.nih.gov/projects/genome/guide/human/index.shtml. GENCODE (v38, v39, v40) annotations were collected from https://www.gencodegenes.org/human/.

All visualizations and 3D protein structure atomic alignments were performed in PyMol (Schrödinger, 2015) version 2.5.2 using non-orthoscopic view, white background, and ray trace 1200,1200. RMSD were calculated without excluding any outliers. Ramachandran plots were created using PyRAMA version 2.0.2 with Richardson (Lovell et al., 2003) standard psi and phi values for all amino acids excluding glycine and proline. Intron-exon structure plots were produced with MISO (Katz et al., 2010) (commit b714021) and TieBrush (Varabyou et al., 2021) (commit e986d64).

RNA-seq data were downloaded from NCBI for run SRR5756467 from BioSample SAMN07278516, a pineal gland from a patient who died at midnight. A detailed summary of the experimental protocol used to generate these data can be found in NCBI BioProject PRJNA391921. Isoform-level quantification was performed using Salmon (Patro et al., 2017) version 1.8.0.

RNA-seq data were downloaded for run SRR18337982 from BioSample SAMN26725167, a tissue sample from the testis of a control mouse. A detailed summary of the experimental protocol used to generate these data can be found in NCBI BioProject PRJNA816862. cDNA reads were aligned to the mm39 reference genome using HISAT2 (Kim et al., 2019) version 2.1.0 then assembled into transcripts using StringTie2 (Kovaka et al., 2019) version 2.2.1.

We assessed the conservation of GT-AG intron boundaries between CHESS human transcripts and transcripts from the GRCm38 mouse reference genome. Data for the human-mouse alignment was extracted from a 30-species alignment anchored on GRCh38 that was downloaded from the UCSC genome browser (Navarro Gonzalez et al., 2021). We used MafIO in BioPython (Cock et al., 2009) version 1.71 to check if all intron boundaries were conserved between mouse and human transcripts. In the supplementary files, a value of ‘TRUE’ in the ‘introns in mouse’ column indicates that all boundaries were conserved, while ‘FALSE’ means that at least one splice site (either a GT at a donor site or an AG at an acceptor site) was not conserved in the alignment.