Illuminating the Druggable Human Proteome with an AI Protein Profiling Platform

Creating a proteomewide atlas of reversible and covalent protein-ligand binding sites would transform our understanding of pro- tein functions and accelerate therapeutic dis- covery; however, current approaches face significant challenges. Chemical proteomics is constrained by limited proteome coverage and data heterogeneity, while existing ma- chine learning models have limited predictive power due to structural dependencies and inadequate handling of heterogeneous ex- perimental labels. Here we developed AiPP , an AI protein profiling platform that provides comprehensive annotations of ligand inter- action sites directly from sequence. AiPP is powered by the evolutionary scale protein large language models (ESM3 and ESMC) and two new databases, LigCysABPP (cys- teine ligandability from activity-based pro- tein profiling or ABPP) and LigBind3D (re- versible binding sites in co-crystal structures). LigCysABPP comprises >700,000 cysteine- site records from 15 ABPP studies covering >10,000 human proteins. We developed an evolutionary representation-based clustering framework followed by consensus analysis to reconcile and augment heterogeneous ex- perimental annotations. Two complementary iterative data expansion protocols were im- plemented to enhance model performance and generalization. AiPP recovers 80% (Top- 1) of all cysteine liganding events from the Protein Data Bank with AUPRC of 84% and AUROC of 89%. Furthermore, it recapitu- lates the consistently and heterogeneously liganded cysteines determined by a recent ABPP study using over 400 cancer cell lines. AiPP provides a starting point for creating a proteome-wide atlas of protein-ligand interac- tions and systematic discovery of druggable sites. The integrative approach that leverages evolutionary information to harmonize large- scale proteomic data can be broadly applied to protein research.

The human proteome comprises ∼20,300 distinct genes that encode over 200,000 pro- teoforms1 (proteins and their variants aris- ing from splicing, post-translational modifica- tions, single-nucleotide polymorphisms and mutations); however, fewer than 900 proteins have been therapeutically targeted by the FDA-approved drugs (https://go.drugbank. com/).2 This gap underscores both the urgent need and vast potential to expand the drug- gable proteome. First developed as a chem- ical proteomic approach that utilizes small- molecule covalent probes (known as the activity-based probes) to report on enzyme 1 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint activities in complex biological systems, 3,4 activity-based protein profiling (ABPP) has evolved into a powerful and versatile tech- nology. By combining activity- or reactivity- based probes with mass spectrometry (MS) based quantitative proteomics, ABPP has en- abled several major advances, 5,6 including the discovery of reactive cysteines7 and char- acterization of small-molecule–protein inter- actions (ligandability) in cells, 8 as well as the development of covalent inhibitors. 9 In par- ticular, the isotopic tandem orthogonal pro- teolysis (isoTOP) ABPP experiments utilizing competition ratios (CRs) between covalent probes and fragments first introduced by Cra- vatt and Weerapana 7,10,11 have led to quan- titative assessment of cysteine ligandability across over 10,000 proteins using different human cell lines 8,12–24 Recently, the ABPP technology has been further developed for high-throughput screening through minimiz- ing the amount of starting protein materials and instrument time 15 as well as using label- free quantification.18,25 Despite the enormous potential of ABPP and rapid progress made in recent years, sev- eral challenges remain, hindering efforts of employing ABPP to create a comprehensive, accurate map of cysteine-directed ligandable sites across the proteome for covalent drug design. One major challenge is limited pro- teome coverage due to the dependence on protein abundance, probe chemistry, and cell type (e.g., membrane-bound proteins are un- derrepresented in lysates due to insolubil- ity).5,25–30 A second major challenge is the variability in detected proteins, cysteines and their assessed ligandability across the pub- lished ABPP datasets (see refs 25,30–32 and later discussions). This variability may at least in part be attributed to the diverse cellular con- texts25,31 and to specific experimental condi- tions,30 including probe chemistry and con- centration, fragment concentration, treatment duration, cell type (e.g., cell lysates vs. live cells), cellular state (e.g., mitotic vs. asyn- chronous conditions; protein mutations; post- translational modifications; redox state),33 cell line (e.g., different cancers 31), specific elec- trophilic fragments used, 25 and quantification protocols (e.g., MS protocol and CR threshold for defining a liganding event). An additional challenge arises from the inherent limitations of shotgun proteomics, 34 which prevents the unambiguous identification of a subset of lig- andable cysteines. 18,25 To address these lim- itations, recent studies implemented label- free and data-independent acquisition proto- cols,18,25 demonstrating improved proteome coverage depth and enhanced data complete- ness compared to traditional approaches. In contrast to proteomics approaches which offer broad coverage across the proteome but with significant data variability, X-ray crys- tallography delivers high-resolution structural information on protein-ligand interactions for a limited number of complexes. Drawing from the cysteine-liganded co-crystal struc- tures deposited in the Protein Data Bank (PDB), several databases (e.g., CovPDB, 35 CovBinderInPDB,36 and LigCys3D 37) have been developed, which annotate nearly 800 proteins with chemically modified cysteines. Based on these databases and experimen- tal or AlphaFold 38 structure models, ma- chine learning (ML) models, including support vector models (SVMs), 39 boosted decision trees,37,40 graph neural networks (GNNs), 41 and convolutional neural networks (CNNs), 37 have been developed for structure-based pre- diction of ligandable cysteines, achieving re- call and precision as high as 95%. How- ever, these models are not yet deployable for proteome-wide mapping of ligandable cys- teine sites due to three critical limitations: (i) the extremely small training dataset and lack of protein diversity; (ii) the absence of reliable controls for unligandable cysteines; and (iii) the requirement for protein structures, which often cannot be met even with state-of- the-art structure prediction tools such as Al- phaFold3.42 Recently, the ABPP-derived databases (CysDB43 and TopCysteineDB 32) have been developed and used to train structure-based ML models for predicting cysteine reactiv- ity or ligandability. 32,44 Based on CysDB 43 and other ABPP data (∼800 reactive cys- 2 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint teines) and associated crystal structures, Forli, Backus, and coworkers developed the random forest model CIAA to predict reactive cysteines with an accuracy of 68% in exter- nal validation.44 Using TopCysteineDB, which combines PDB- (nearly 800 liganded cys- teines) with ABPP-derived cysteine ligand- ability data (over 9,000 ligandable cysteines compiled from three publications 13,15,31 ) and experimental or AlphaFold structures, Gohlke and coworkers 32 developed the XGBoost model TopCySPAL to predict cysteine ligand- ability. A unique feature of TopCySPAL is that it is trained with negative cysteines defined as those undetected in liganded proteins quan- tified by ABPP. 32 In the 20% hold-out test, TopCySPAL achieved AUROC and AUPRC of 96.4% and 91.4%, respectively. Another structure-based ML model was the CNN 31 developed using the DrugMap database (an ABPP dataset obtained with 400 cancer cell lines), which achieved AUROC of 69.3% in external evaluation. Despite recent advances, existing ML mod- els for predicting cysteine ligandability remain severely constrained in both performance and applicability. First, current approaches for developing ML models inadequately leverage ABPP datasets and lack rigorous frameworks for assignment of ligandability labels. These models fail to reconcile the inherent variabil- ity in experimental labels, leading to incon- sistent training data and unreliable predic- tions. Second, this reconciliation challenge is compounded by the absence of standardized or rigorous evaluation metrics and bench- marking protocols, making it difficult to rig- orously assess model performance or com- pare different approaches in a meaningful way. Finally, and perhaps most critically, exist- ing models require high-quality protein struc- tures as input, rendering them inapplicable to substantial portions of the proteome that re- main structurally unresolved even with state- of-the-art structure prediction tools such as AlphaFold3,42 creating a significant coverage gap in ligandability assessment. To overcome the ML challenges and un- lock the full potential of ABPP , we developed a sequence-based AI Protein Profiling (AiPP) platform for annotating reversible ligand bind- ing (LigBind) and (cysteine-directed) cova- lently ligandable residues (LigCys) across the proteome. AiPP leverages the state-of-the- art protein large language models (pLLM) ESM Cambrian (ESMC) 45 as the foundation model. We curated a comprehensive ABPP database LigCysABPP comprising 15 inde- pendent datasets published since 2016. We developed a representation-based clustering strategy to reconcile and augment the lig- andability labels. A similar strategy is ap- plied to refine the labels in LigBind3D. We further developed two iterative learning pro- tocols to systematically expand the ML train- ing data, achieving enhanced model accu- racy and generalization. Beyond predict- ing ligandable sites, AiPP incorporates RIDA, a disorder annotation platform that identifies short disordered molecular recognition fea- tures (MoRFs) with a propensity for binding- induced folding Cysteine-containing MoRFs are particularly interesting, as they may serve as anchor points for protein-protein interac- tion (PPI) drug design. AiPP provides a unified platform, paving the way for creating proteome-wide annotation of druggable pock- ets and their adjacent covalently ligandable residues.

and Discussion Curation of an ML-ready cysteine ligand- ability database LigCysABPP. From 15 independent ABPP datasets (also referred to as sources) published between 2016 and 2025,8,12–25 we extracted 683,192 peptide- derived records of cysteine ligandability for 58,704 unique cysteine residues across 14,417 proteins in human proteomes. Fol- lowing data clean-up steps, including protein sequence reconstruction, record validation, and consolidation of UniProt IDs (UIDs) refer- ring to identical protein sequences, the raw entries were converted to 669,908 structured, site-specific cysteine ligandability records for 53,867 cysteine residues across 11,017 pro- 3 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint a Other liganded 215 (203) ABPP quantified proteins 10,649 (15,559) Drug targets 70 (87) LC3D 285 (290) unliganded proteins: 4,147(30,663Cys) db c Figure 1. Figure 1: Overview of the ABPP-quantified proteins and cysteines in LigCysABPP database. a. The Lig- CysABPP database is manually curated from a collection of 15 published cysteine-directed ABPP studies. 8,12–25 Peptide-derived records are extracted from each publication, validated against UniProt, consolidated across iden- tical sequences and subsequences, augmented with unquantified (unseen) cysteine-sites on otherwise quantified proteins, and filtered for compatibility with downstream processing steps. b. LigCysABPP contains the cysteine ligandability records of 10,649 ABPP-quantified proteins, of which 6,502 are liganded and 4,147 are unliganded. A liganded cysteine is defined as having at least 1 pos ABPP record while a liganded protein is defined as having at least 1 liganded cysteine. Among the liganded proteins, 931 are known drug targets (with 2,293 liganded cysteines), while the rest of 5,571 other liganded proteins contain 13,266 liganded cysteines. The LC3D database contains 285 unique proteins (290 cysteines liganded in the co-crystal structures), among which only 70 proteins (87 drug targets) are also identified as liganded by ABPP .c. Functional classes of drug targets, other liganded proteins, and unliganded proteins according to The Human Protein Atlas classifications46(https://www.proteinatlas.org/). d. Histogram of the number of ABPP liganded cysteines per protein shows that most liganded proteins contain 1-3 liganded cysteines. teins. Each record contains the ligandabil- ity label (pos for liganded and neg for un- liganded) given by the study authors for a specific cysteine site in a protein under a set of experimental conditions and cellular con- text. To fully characterize the cysteines in ABPP-quantified proteins, we added unquan- tified (provisionally neg) records for cysteines belonging to the ABPP-quantified proteins but undetected by the probes. To accommo- date the ESM-based ML, protein sequences shorter than 30 or longer than 2046 amino acids were removed. This procedure results in 703,135 total curated records describing ligandability (pos, neg, or unquantified) of 140,459 unique cysteine sites across 10,649 distinct proteins (Fig. 1a). Detailed extrac- tion, validation, and consolidation logs are provided in Supplemental Data (SD6–SD11). Functional analysis of ABPP-quantified proteins. We define a liganded cysteine as one with at least one positive ABPP record. A supporting source for a liganded cysteine is one that contains at least one pos record for that cysteine. Among the 10,649 ABPP- quantified proteins, 6,502 are liganded (con- taining at least one liganded cysteine) and 4,147 are unliganded (Fig. 1b). Among the liganded proteins, 931 (with 2,293 pos cys- teines) are known drug targets, while the rest of 5,571 proteins (with 13,266 pos cys- teines) represent unexplored potential ther- apeutic opportunities (Fig. 1b). Among the liganded drug targets and other proteins, enzymes constitute the largest functional 4 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint family, followed by transporters, nuclear re- ceptors, and others (Fig. 1c). It is note- worthy that transporters and nuclear recep- tors are substantially more prevalent among other ligandable proteins, each represent- ing over one-third the abundance of en- zymes (Fig. 1c). Interestingly, GPCRs are sparsely represented, likely because most cysteine residues are not positioned near ligand-binding pockets. It is intriguing that the majority of unliganded proteins (2,593 out of 4,147) do not belong to any functional class identified by The Human Protein Atlas46 (https://www.proteinatlas.org/), suggest- ing that a significant fraction of the human proteome remains to be functionally charac- terized. The overlap between the ABPP-liganded proteins and those evidenced by co-crystal structures in PDB is extremely small, only 60 proteins (42 are known drug targets, Fig. 1b). This limited overlap may be attributed to the historical focus of structural studies on bacte- rial and other non-human proteins, which are overrepresented in the PDB. The limited over- lap also demonstrates the potential of ABPP to greatly expand the known ligandable pro- teome. The vast ABPP data also allows us to ask a pertinent question: how many lig- andable cysteines are in a single protein? Interestingly, while the majority of proteins contain only one ligandable cysteine, a siz- able number of proteins harbor 2–4 ligandable cysteines (Fig. 1d), suggesting that cysteine- directed probes or inhibitors may be designed to target allosteric and other ligandable pock- ets. Analysis of ABPP sources and records to assess consensus in cysteine ligand- ability. In order to devise an ML label- ing scheme based on LigCysABPP records, we first analyzed the consensus among 15 sources for liganded cysteines, which are grouped by the number of sources that quan- tified them (n =1,...,12). For instance, 2,996 cysteines were quantified by a single source, while 2,644 and 2,154 were quantified by 2 and 3 sources, respectively. We then exam- ined the consensus levels by determining how many liganded cysteines are supported by m number of sources (m =1,...,8). This analysis highlights the challenge of reconciling source- specific labels, as considerable variability ex- ists. For example, among cysteines quanti- fied by 2 sources, only 310 (11%) are sup- ported by both sources, while 2,334 (89%) showed conflicting results – pos in one source and neg in the other. Among cysteines quan- tified by 3 sources, only 82 (4%) are unani- mously supported, while 449 (21%) are pos in 2 sources and 1,623 (75%) are pos in only 1 source. This pattern of decreasing consensus becomes more pronounced as the number of sources increases (Fig. 2a and Supplemental Fig. S1). In the above analysis, a supporting source can have any number of pos records. To fur- ther dissect the ABPP noise, we also exam- ined the number of supporting records for lig- anded cysteines (Fig. 2b). For liganded cys- teines supported by 1 source, 7,787 have only 1 pos record, 1,374 have 2 pos records, and 816 have 3–4 pos records. For liganded cys- teines supported by 2 sources, 814 have 3–4 positive records. As the number of support- ing sources increases, the number of liganded cysteines with positive records decreases – with 5 sources, the number of liganded cys- teines with at least 5 pos records is below 30. The steep decline in well-supported liganded cysteines suggests that requiring agreement among more than 4 sources may be too strin- gent for practical ML applications, as it would severely limit the training data size. The significant variation in cysteine ligand- ability likely reflects differences in experimen- tal conditions (e.g., probe and fragment con- centrations) and cellular context (e.g., cell line-specific factors) as shown in SI Data Ta- ble S1. The influence of cellular context on cysteine ligandability was recently 31 demon- strated by an ABPP study of over 400 can- cer proteomes, which identified nearly 400 cysteines exhibiting varying degrees of en- gagement by an identical fragment (“hetero- geneous ligandable”). While cellular con- 5 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint c a 2,996 2,334 1,623 449 310 275 106 72 235 95 60 92 82 146 164 1168787 162 75 100 94 101 179241 105 115 50 53 82 37 436 1,010 754353 141 n=10 (479) n=11 (355) n=12 (234) 119 24 36 n=1 (2,996) n=2 (2,644) n=3 (2,154) n=4 (1,589) n=9 (572) n=5 (1,285) n=6 (1,002) n=7 (877) n=8 (680) 549 416 306 b Cys quantified in n sources (publications) and liganded in: 1 source 2 sources 3 sources ≥4 sources In Pos Cluster In Neg Cluster Unlabeled 94,128 Unquantified Cys 74,190 11,665 8,273 Supporting sources Positive records 1 3 5 7 92 4 6 8 10 5 10 15 20 Figure 2: Analysis of cysteine ligandability across ABPP sources and records. a. Each pie chart displays the number of liganded cysteines quantified by n sources. A liganded cysteine is defined as one that has at least one pos ABPP record. Each pie segment represents the number of cysteines supported by 1 (blue), 2 (green), 3 (yellow), and ≥4 (purple) sources. b. The number of liganded cysteines with the number of supporting sources and positive records. A supporting source for a specific cysteine is defined as one that contains at least one pos record for that cysteine. A total of 7,787 cysteines are labeled pos by only 1 record (black). No liganded cysteine is supported by >10 sources. Liganded cysteines with >25 records are negligible and therefore excluded from the plot. c. Following representation-based clustering, 8,273 unquantified cysteines are labeled pos, 11,665 labeled neg, while 74,190 cysteines remain unlabeled. texts will be addressed in our future stud- ies, the present work focuses on cysteines that demonstrate consistent labeling patterns across different experiment conditions and cell lines. Using representation-based clustering to derive cysteine ligandability labels for ML. The variability in cysteine ligandability across the ABPP datasets presents a significant challenge for ML. Encouraged by our recent work47 demonstrating that residue-level rep- resentations (also known as per-token em- beddings) from ultra-large pLLMs such as ESMC45,48 encode not only evolutionary infor- mation but also structural and local biochemi- cal environments, we leveraged the residue- specific representations to reconcile the la- bels for liganded and unliganded cysteines. Specifically, we performed clustering analysis on all cysteines in ABPP-quantified proteins using pairwise similarity scores computed from ESMC-derived residue-specific embed- ding vectors. For each cluster, we identified a representative cysteine as the node exhibiting the highest average similarity to all other clus- ter members. This approach yielded compact, non-overlapping clusters of cysteine sites with high embedding similarity, each anchored by a representative node used for consensus la- beling and subsequent model training. To pre- vent leakage, clusters were never split across data partitions and all clusters sharing a UID were assigned to the same partition. To derive high-confidence ligandability la- bels from the ABPP data, we performed cluster-based label assignment on the full set of representation-based clusters accord- ing to varying levels of consensus, defined as a combination of supporting sources and 6 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint records (denoted nS–mR). For example, un- der the 4S–4R criteria, a cluster is labeled pos only if it contains at least four positive records from four different sources. Conversely, a cluster is labeled neg if it contains no posi- tive records and at least four negative records from unique sources. Clusters that did not meet either pos or neg labeling condition were excluded from the training set. To augment coverage, unquantified cysteines in quantified proteins cast default negative votes and in- herit a cluster-based label. Applying the 4S–4R criteria yielded 936 pos and 4,590 neg labeled clusters, cov- ering 143,914 site-level records describ- ing 11,869 distinct cysteine residues across 11,541 unique proteins. Importantly, this ap- proach enabled us to assign labels to 21% of ABPP-unquantified cysteines, classifying 8,273 as pos and 11,665 as neg (Fig. 2c). The cluster representatives were then used for downstream model training. Additional de- tails are provided in the Supplemental Meth- ods, including data validation, representation- level deduplication, and procedures with representation-based criteria to prevent data leakage between training, validation, and ex- ternal evaluation. Note, while conventional sequence identity filtering methods, such as CD-HIT, 49 focus on eliminating sequence similarity, they fail to account for functionally convergent or struc- turally similar ligandable sites that may exist within otherwise divergent proteins. In con- trast, clustering based on pLLM representa- tion goes beyond simple sequence compar- ison by integrating local biochemical, struc- tural, and evolutionary information. Thus, representation-based criteria are more strin- gent and effective in reducing the risk of data leakage. Support from 4 sources may be an opti- mal tradeoff between label confidence and data coverage. To examine how label con- fidence depends on the number of quantify- ing sources, we measured variability using Shannon entropy (a measure of labeling un- certainty) of the binary pos/neg distribution for cysteine clusters quantified byn sources. The number of clusters falls sharply as the thresh- old increases; for example, requiring at least 4 quantifying sources reduces the set to 11,913 clusters compared to 27,410 with a 2-source threshold (Fig. 3a, top). Entropy is high for clusters supported by 2 or 3 sources, reaches its minimum at 4 sources (Fig. 3a, bottom), and then rises again at thresholds ≥ 7, peak- ing at 12 sources where only∼150 highly het- erogeneous clusters remain. Fig. 3b complements this analysis by show- ing how positive cysteines accumulate as ad- ditional sources are added under different consensus rules (1S–1R, 2S–2R, . . ., 5S– 5R). At the 4-source threshold, the num- ber of positives normalized by the na ¨ıve to- tal (all cysteines reported as liganded in any record) qualitatively plateaus: each additional source contributes relatively few new positives (Fig. 3b, top). Together, these results sug- gest that requiring evidence from at least 4 supporting sources (defined here as quanti- fying sources that reported a cysteine as lig- anded at least once; the 4S–4R rule) provides a practical balance between label confidence and data coverage for annotating ligandable cysteines with the currently available data. Development of the sequence-only LigCys models. Inspired by our recent demonstra- tion that ESMC-based, sequence-only mod- els can accurately predict residue-specific pKa values, particularly for cysteines despite limited training data, 47 we developed the Lig- Cys model for predicting cysteine ligandabil- ity. The model employs ESMC as the founda- tional pLLM, using frozen 2,560-dimensional per-residue embeddings passed through a three-layer multilayer perceptron (MLP) clas- sifier trained on the ABPP-derived data de- scribed below. Embeddings from layer 76 of ESMC were used (Methods), as prelimi- nary testing showed this layer provided opti- mal performance for LigCys predictions. Con- sistent with prior observations, 47,50 different transformer layers capture distinct evolution- 7 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint a b 4S-4R Figure 3: Analysis of how source threshold affects label variability and data coverage. a. Top: Bar plot showing the number of cysteine clusters supported by at least n independent sources, regardless of the number of underlying records. For clarity, bins corresponding to odd values ofn are unlabeled. Bottom: Label variability within each cluster, stratified by the number of supporting sources. Variability is quantified using the Shannon entropy of the binary label distribution ( p, the fraction of sources labeling a cysteine as positive), scaled to a 0–100% range. Boxes denote the interquartile range; medians are shown as horizontal dashed lines; whiskers span the 10th–90th percentiles. A minimum in entropy is observed at n ≥ 4, suggesting this threshold maximizes labeling consistency across sources. b. Top: Incremental gain in the number of positive cysteines when increasing from n−1 to n sources. Bottom: For each consensus threshold (e.g., 1S–1R through 5S–5R), we randomly sample n sources from a pool of 15, repeat 200 times, and compute the average number of positive cysteines. This is normalized by the na¨ıve total (i.e., all cysteines reported as positive in any single record). The 95% confidence interval is shown for the na¨ıve curve. The 4S–4R consensus curve (highlighted in purple) plateaus after four sources, supporting its use as a data-driven threshold for robust labeling. ary and structural features at different rates; layer 76 embeddings were therefore also used for LigBind models. We also evaluated pLLM fine-tuning approaches using ESM2, 51 but the resulting models did not outperform the representation-only ESMC approach for cysteine ligandability prediction. LigCys models trained on the 4S–4R ABPP dataset show the strongest generaliza- tion. Due to the highly variable nature of ABPP data, both model training and validation are extremely challenging. To address val- idation challenges, we used crystallography- validated ligandable cysteines from the LC3D database as an external test set. LC3D cysteines and any representation-based clus- ters containing these cysteines were removed from the training set to prevent data leak- age (Supplemental Methods). Therefore, the LC3D test set is not only orthogonal to the ABPP training set but also provides the most stringent validation for model’s ability to gen- eralize. We evaluated how ABPP data consensus requirement affects the model performance by training LigCys models with ligandability la- bels derived from increasingly stringent con- sensus criteria (Table 1). These LigCys mod- els are compared using the standard metrics, including the most pertinent metric Top-1 re- covery, which is defined as the probability that the cysteine with the highest predicted classi- fication score in a protein is a true pos given that the protein has at least one pos cys- 8 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint Table 1: External evaluation of LigCys models trained using ABPP datasets with varying con- sensus thresholdsa ABPP training data External LC3D evaluation (%) Criteria Pos Neg Prot PRC ROC Prec Rec Top-1 Top-1* 1S–1R 13430 55441 7051 60.0 72.2 38.4 37.8 40.5 44.5 1S–2R 6078 16791 4216 70.2 80.9 55.9 55.1 56.8 59.6 1S–3R 3396 8600 2829 76.6 85.8 62.3 61.9 65.3 68.2 1S–4R 2094 5048 2004 78.1 86.3 65.3 64.9 68.3 72.9 1S–5R 1418 3354 1495 80.1 87.4 70.0 69.6 71.7 75.9 2S–2R 3467 6838 2617 77.0 84.2 65.3 64.7 66.2 70.3 2S–3R 2198 4765 1883 77.3 85.5 64.3 64.1 66.6 70.3 2S–4R 1450 3326 1404 78.8 86.4 69.6 68.8 70.9 75.1 2S–5R 988 2258 1041 79.5 86.6 71.7 70.9 72.2 76.8 3S–3R 1115 1920 997 78.8 86.6 68.8 68.1 70.0 74.2 3S–4R 867 1583 837 80.2 88.3 68.3 67.5 71.3 75.5 3S–5R 649 1244 669 80.8 87.7 72.6 71.7 73.5 77.6 4S–4R 384 642 389 81.2 87.3 74.3 73.2 75.2 79.4 4S–5R 347 590 358 80.3 86.1 74.3 73.2 74.3 78.9 5S–5R 142 236 150 79.8 85.6 71.3 70.3 72.6 76.8 aAt a consensus threshold nS–mR, the label assignment requires at least m records from n different sources (see main text). Each model is an ensemble of 200 models (see Supplemental Methods). Performance is measured on the orthogonal LC3D testset on the per-protein basis. Cysteines without covalent ligation are provisionally labeled neg. Precision and recall of predicting ligandable cysteines are calculated with a classification threshold of 0.5. Top-1 recovery refers to the probability that the cysteine with the highest classification score is a true pos given that the protein has at least one pos cysteine. Top-1* refers to the evaluation using LC3D augmented with labels from representation-based clustering (see Supplemental data). teine. Since LC3D negatives reflect the ab- sence of observed liganding event rather than validated inactivity, AUROC/AUPRC should be interpreted with caution; we therefore em- phasize ranked recovery metrics (e.g., Top- 1). Confirming the substantial heterogeneity across ABPP data, the 1S–1R model, despite being trained on the largest dataset (13,430 pos and 55,441 neg cysteines), delivers the worst performance, with AUROC, AUPRC, and Top-1 recovery values of 72.2%, 60.0%, and 40.5%, respectively. At the single-source level, performance improves dramatically as the number of required pos records increases, reaching AUROC, AUPRC, and Top-1 recov- ery of 87.3%, 81.2%, and 71.7%, respectively, despite a ten-fold reduction in training data. A similar trend of significant performance gain with increasing number of pos records is ob- served at the 2- and 3-source levels. How- ever, at the 4-source level, performance with 5 pos records (4S–5R) is very similar to that with 4 records (4S–4R), which has the AU- ROC, AUPRC, and Top-1 of 87.3%, 81.2%, and 75.2%, suggesting that model perfor- mance may be plateaued. Indeed, the 5- source model (5S–5R) shows decreased per- formance across all three metrics by 2–5% compared to the 4S–5R model, which may be attributed to the substantial (more than 50%) reduction in training data (only 142 pos and 236 neg cysteines). This comparison shows that the 4S–4R training set yields the best model performance (Table 1, highlighted row) for generalization to the completely orthogonal external vali- dation data. This observation is consistent with the analyses in Fig. 2 and 3, suggest- 9 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint ing that the 4S–4R training set balances la- bel confidence with dataset size for effective generalization. To further assess model per- formance, we computed Top-1* recovery for the LC3D test set that received additional label assignments through representation- based clustering among themselves (referred to augmented LC3D or LC3D*). The 4S– 4R model remains the top-performing one, achieving AUROC, AUPRC, and Top-1* of 87.3%, 81.2%, and 79.4%, respectively. Enhanced model performance through it- erative training data expansion guided by LC3D-based evaluation. Although the 4S– 4R dataset achieves superior performance relative to others, its small size may limit the model’s generalization capability. To ad- dress this limitation, we hypothesized that us- ing LC3D as a validation benchmark would allow us to identify additional cysteines (and proteins) with high-confidence labels from the broader ABPP dataset to expand the training data. To test this hypothesis, we developed an iterative procedure guided by LC3D Top-1 recovery, starting from a truncated version of the high-confidence 4S–4R dataset contain- ing cluster representatives from proteins with both positive and negative cysteines. In this truncated version (denoted 4S–4R t), 23 pro- teins quantified by 10 or more sources with at least one pos and one neg cysteine are withheld for model evaluations. The candidate training pool included cysteines from all 4S– 4Rt clusters, extending beyond the baseline to encompass non-representative members as well as proteins supported exclusively by positive or exclusively by negative evidence. In iterations 1–3, we generated 100 batches, each consisting of 80, 100, and 150 randomly selected cysteines from the can- didate pool, respectively. From iteration 4 onward, we generated 100 batches of 175 randomly selected cysteines. In each itera- tion, batches were evaluated by incorporating them into the training set and training ensem- bles of 24 models (6 data splits with 4 random seeds per split). The best-performing batch was selected for further validation using 10 new random splits with 20 models per split, with Top-1 recovery serving as the evaluation metric. If Top-1 recovery exceeded the pre- vious iteration’s baseline, the best batch was accepted. Through six iterations, Top-1 recovery in- creased from 70.9% to 79.9% on LC3D and from 75.1% to 84.5% on LC3D*, accompa- nied by steady improvements in AUROC from 85.5% to 88.9% and AUPRC from 78.1% to 84.2% (Table 2). In iteration 7, no further gain was observed, indicating that performance had plateaued under this random-batch ex- pansion scheme. Hereafter we will call the Iter6 model LigCys-S. Relative to the base- line 4S–4R t dataset, the expanded training set nearly tripled in size to 1,099 proteins, driven by over twofold increase in negative cysteines (1,345) and a nearly 30% increase in positives (441). Using the same data, we trained a structure-aware model (6-SA) to test if adding additional structural features would enhance predictive power (Table 2, row 6-SA). Interestingly, along with other calculated met- rics, the Top-1 recovery of LC3D liganded cys- teines is decreased by 1.7 %. This suggests that the evolutionary driven pLLM-based rep- resentations capture the majority of informa- tion pertinent to cysteine ligandability. Enhanced model performance through it- erative training data expansion guided by ABPP-based evaluation. As random sam- pling in the LC3D-guided procedure may leave parts of the candidate pool unexplored, we hypothesized that a systematic, cross- validation (CV) driven approach could more efficiently identify additional high-confidence data. To test this hypothesis, we developed an expansion strategy guided by CV perfor- mance on the ABPP data, starting from the same baseline dataset (4S–4R t) as in the LC3D-guided expansion. The same candi- date data pool was also used. At each iteration, candidate batches of fixed composition (25 positives and the remain- der negatives) were assembled, with batch 10 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint Table 2: Performance metrics for iterative training data expansion guided by LC3D-based evalu- ation (top block) and ABPP-based cross validation (bottom block). ABPP training dataa ABPP valid. (%)b External LC3D evaluationc (%) LC3D-guided expansion Iter Pos Neg Prot PRCE ROC PRC ROC Prec Rec Top-1 Top-1* 0 354 574 366 74 74.6 78.1 85.5 70.5 69.6 70.9 75.1 0-SA 354 574 366 79 76.5 80.1 86.9 71.3 70.5 72.6 76.8 1 359 649 442 92 78.2 79.5 85.8 71.3 70.3 72.2 76.8 2 369 739 534 94 76.1 80.4 86.8 73.2 73.4 73.5 78.1 3 384 874 674 92 74.2 81.8 87.8 73.9 72.8 74.8 79.8 4 401 1032 827 115 77.5 82.9 88.0 78.2 76.9 78.2 82.8 5 425 1186 966 139 79.4 83.0 87.8 76.9 75.6 78.2 82.8 6 441 1345 1099 141 78.2 84.2 88.9 77.3 77.7 79.9 84.5 6-SA 441 1345 1099 128 76.7 83.8 88.7 76.9 75.6 78.2 83.2 ABPP-guided expansion Tourd Pos Neg Prot PRCE ROC PRC ROC Prec Rec Top-1 Top-1* 0 354 574 366 74 74.6 78.1 85.5 70.5 69.6 70.9 75.1 1 378 643 452 93 80.2 80.1 86.8 72.6 71.7 73.1 77.3 2 468 967 817 122 82.1 81.5 88.3 78.1 72.1 74.4 79.4 3 555 1432 1258 148 82.6 80.1 87.5 71.3 70.3 71.8 76.0 4 645 1998 1744 187 83.5 79.3 87.1 69.8 69.2 70.5 76.0 4-SA 645 1998 1744 178 80.9 80.5 87.7 69.8 69.2 71.7 76.8 Blendede - - - - - 83.2 88.4 75.4 74.5 76.9 85.5 a In the LC3D-guided data expansion, Top-1 recovery of LC3D liganded cysteines was used to select candidate data batches (complete data in Supplemental Data). The total number of pos, neg cysteines, and distinct proteins in the training set is given after each iteration. The start of the expansion (Iter 0) used the 4S–4R t dataset, which excludes 23 proteins quantified by 10 sources from the original 4S–4R dataset. b Model validation metrics are AUROC (ROC) and AUPRC enrichment (PRCE), defined as(AUPRC−random)/random. Here random represents the AUPRC of a random classifier (% positive samples in the dataset). c For external evaluation against the LC3D dataset, precision (Prec) and recall (Rec) were calculated using the default classification threshold of 0.5. Top-1 and Top-1* were calculated using LC3D and LC3D* (label augmented by clustering) datasets, respectively. d In the ABPP-guided data expansion, AUPRC enrichment (PRCE) was used to select candidate data batches. Tour denotes the number of successive tournaments (effectively iterations 1, 5, 9, and 13; see Supplemental Data). All other details follow the notes above. e Blended denotes ensemble prediction using both Iter 6 and Tour 4 models. size set to 100 cysteines in early iterations and increased to 175 as the pool expanded. Batches were prioritized by ensemble uncer- tainty, and a batch was admitted only if it sig- nificantly exceeded the AUPRC enrichment (PRCE) of the previous iteration baseline in the CV. PRCE (enrichment of AUPRC over random guess) was used due to the vari- able pos:neg Cys ratio in each iteration. To broaden the search, advancing batches were further refined by simple genetic algorithm steps, preserving the best performers while generating crossover and mutation variants. Expansion proceeded through four tourna- ments, each comprising four iterations. Here, a “tournament” refers to a series of eval- uations in which candidate batches com- pete against the baseline: top performers 11 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint are retained and recombined, and only the strongest batch is accepted at each itera- tion of the tournament. Expansion terminated when the final tournament ended without any accepted batch. This acceptance scheme yielded steady performance gains: Top-1 re- covery stayed constant at 70.9% to 70.5% on LC3D and at 75.1% to 76.0% on LC3D*, with corresponding increases in AUROC and AUPRC (Table 2, under ABPP-guided expan- sion). Hereafter we will refer to the Tour4 model as LigCys-A. The final expanded set comprised 1,744 proteins and 2,633 cysteine residues, representing a 1.8-fold increase in positives and a 3.4-fold increase in negatives relative to the 4S–4Rt baseline. Performance gains plateaued after the fourth tournament, indicating that the available candidate pool had been exhausted. As in the LC3D-guided expansion, we also trained a structure-aware model using the final expanded data. The resulting model (4-SA) has a small increase of 1.2% for AUPRC and Top-1 recovery com- pared to the sequence-only model. Beyond the LC3D test, we evaluated the LigCys-S and LigCys-A models against the ABPP hold-out set comprised of 23 proteins quantified by 10 or more sources, each con- taining at least one positive and one negative cysteine. From this hold-out set, we created 9 test sets using the nS-nR consensus thresh- olds (n =1,...,9). As expected, the number of labeled cysteines decreases from 1S-1R (∼300 cysteines) to just under 100 at 4S-4R and 7 at 9S-9R (Supplemental Fig. S2). The AUPRC of both models increases with the consensus level of the test set, from ∼68% (67 for LigCys-A model) at 1S-1R to 73% (75 for LigCys-A model) at 4S-4R and nearly 100% at 9S-9R (Supplemental Fig. S2). Sim- ilarly, the model recall steadily increases from below 40% at 1S-1R to 100% at 9S-9R (Sup- plemental Fig. S2). These improvements stem from increased TPs and decreased FNs with growing consensus level (Supplemental Fig. S2). The analysis demonstrates that the models perform better on consistently lig- anded cysteines, corroborating our initial find- ings (Table 1). Consistent with the LC3D test, the ABPP hold-out test suggests that LigCys- S and LigCys-A models perform similarly, with LigCys-S showing slightly fewer false posi- tives when predicting highly ligandable cys- teines. Development of an ML-ready LigBind3D database and the sequence-only LigBind models for predicting reversible ligand binding residues. Since truly ligandable cysteines are located at or near a reversible binding pocket, we developed sequence-only LigBind models (LigBind-seq) to predict re- versible ligand-binding residues complemen- tary to cysteine ligandability predictions. To develop the LigBind models, we created an ML-ready database called LigBind3D based on BioLiP2, 52 an updated manually curated dataset of experimental structures from bio- logically relevant protein–ligand complexes in the PDB. BioLiP252 is significantly larger than commonly used datasets for protein-ligand affinity predictions (e.g., BindingDB 53) which contain only protein–ligand interactions with experimentally determined binding affinities. We selected BioLiP2 entries of monomeric proteins containing small molecules with molecular weights between 150–600 Da and applied a custom structural annotation pipeline (PickPocket, unpublished) to vali- date biologically relevant protein-ligand inter- actions. Following the various filtering steps, we reconstructed the full protein sequence di- rectly from the atomic coordinate file. Ligand- binding residues were labeled based on a 4.5 ˚A heavy-atom distance cutoff. These residues were converted into structured, site- level records. Details are given in Supplemen- tal Methods. The final LigBind3D database comprises 687,712 site-level records across 1,998 unique proteins. This database offers a direct structural perspective on reversible protein-ligand interactions, complementing the LigCysABPP and LC3D databases. In contrast to LigCys3D, which restricts to cys- teines, LigBind3D labels every residue in a protein, reflecting the broader diversity of reversible protein-ligand interactions. Sim- 12 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint ilar to the treatment of LigCysABPP data, representation-based clustering was applied to the LigBind3D data for creating the training, validation, and test datasets. The LigBind-seq model employs a simpler single-layer feedforward network architecture that processes fixed 2,560-dimensional per- token representations extracted from ESMC layer 76. Preliminary screening showed that the simpler architecture was sufficient to cap- ture ligand-binding patterns from sequence alone. In addition to the sequence-only mod- els, we also developed structure-aware Lig- Bind models (LigBind-SA) by incorporating 3D geometric features through a pretrained transformer model derived from the PeSTo ar- chitecture54 alongside the pLLM embeddings. Details are given in Supplemental Methods. In the hold-out test, LigBind-seq achieves AUROC, AUPRC, and Top-10 scores of 93.9%, 74.5%, and 78.1%, respectively, while LigBind-SA attains 95.1%, 77.5%, and 80.1% (Table 3). The modest performance improve- ments of LigBind-SA relative to LigBind-seq indicate that the ESMC-derived representa- tions provide the majority of the model’s pre- dictive power. Table 3: Performance of the sequence-only and structure-aware LigBind models in pre- dicting reversible ligand-binding residuesa Metrics (%) LigBind-seq LigBind-SA AUROC* 93.9 95.1 AUPRC* 74.5 77.5 Prec* 72.1 71.1 Rec* 66.4 71.5 Top-5* 79.2 81.2 Top-10* 78.1 80.1 Top-20* 76.6 79.5 a All models were trained and evaluated* using cluster- propagated labels derived via representation-based clustering. Performance metrics reflect a post hoc ensemble of 200 models (Supplemental Methods). Per-residue ligandability probabilities were averaged across all 200 models and thresholded at 0.5. All met- rics were computed on a per-protein basis using the held-out test set of 129 proteins. LigBind predictions offer complementary validation for LigCys predictions. Cova- lent ligandability of cysteines requires prox- imity to a reversible binding pocket. There- fore, accurate LigBind predictions not only provide reversible pocket information but may also serve as complementary validation for LigCys predictions. To test this complemen- tary relationship, we applied both LigBind and LigCys models to the LC3D test set with pro- teins in the LigBind training set excluded. We then calculated the radial distribution function (RDF) of predicted ligand-binding residues surrounding TP , TN, FP , and FN cysteines (Fig. 4). Both LigBind-seq and LigBind-SA models were used. Except for TNs, all RDFs exhibit a peak around 4–6 ˚A. Notably, both LigBind-seq and LigBind-SA RDFs for TPs show the highest peak, consistent with the ex- pectation that ligandable cysteines are proxi- mal to reversible binding pockets. The peak of the LigBind-SA RDF (dashed) is a bit higher compared to the LigBind-seq RDF (solid), suggesting that incorporating structural fea- tures enhances pocket prominence. Reas- suringly, both LigBind-seq and LigBind-SA RDFs for TNs lack peaks, consistent with the expectation that unligandable cysteines are not located near ligand-binding pockets. Re- markably, the LigBind-seq RDF for FNs ex- hibits the second highest peak, suggesting that LigBind-seq predictions could help iden- tify overlooked ligandable cysteines based on pocket proximity. Conversely, the LigBind- seq and LigBind-SA RDFs for FPs display very low peaks, suggesting that incorrectly predicted ligandable cysteines can be filtered out based on their lack of proximal binding pockets. These results demonstrate that inte- grating LigCys and LigBind predictions within a unified framework would further enhance the reliability of covalent ligandability assess- ment. Development of the AiPP platform for sequence-based predictions of reversible and covalent ligand binding sites across the proteome. Building on the LigCys and 13 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint 𝑔 𝑟 Distance 𝑟 from Cys of interest (Å) 0 20 255 10 158 100 200 (tail normalized) TP (n=15) TN (n=167) FP (n=8) FN (n=122) LigBind-seq LigBind-SA 300 400 500 600 700 Figure 4: Radial distribution function of LigBind- predicted ligand-binding residues around cysteine sites, stratified by the LigCys prediction outcome (TP, TN, FP, FN).Distance is the minimum all-atom Eu- clidean distance between the cysteine of interest and each predicted ligand-binding residue. Radial distribu- tion function g(r) is normalized by the mean shell den- sity over the last 20% of the distance range. Solid and dashed lines represent the LigBind-seq and LigBind- SA predictions, respectively, on the pruned LC3D test set that excludes proteins in the LigBind training set. LigBind models, we developed the Artificial Intelligence Protein Profiling (AiPP) platform as a unified, sequence-based framework for proteome-wide ligandable site identifica- tion (Fig. 5). AiPP integrates reversible and covalent ligand-binding predictions as well as a disorder predictor to support rational drug discovery. The LigBind module detects residues that spatially stabilize ligand binding; the LigCys module identifies cysteines sus- ceptible to modification by a covalent ligand; and the RIDA module (based on the RIDAO server56) predicts per-residue disorder and molecular recognition features (MoRFs) that can fold upon binding. Together, these pre- dictions map ligandable binding sites capa- ble of retaining cysteine-reactive compounds sufficiently long for covalent bond forma- tion—guiding targeted covalent drug design, while simultaneously identifying dynamic or induced-fit binding regions amenable to PPI inhibitor development. Additionally, the re- cently developed KaML-ESM 47 and KaML- CBtree57 models are incorporated to offer reactivity information for cysteines of interest based on sequence only or structure. Starting from a protein sequence, AiPP gen- erates a predicted 3D structure (via a local version of ESM3 48), a structural and func- tional disorder annotation (via a local ver- sion of RIDAO server 56), and the residue- specific representations (from the pLLM ESMC45). The ESMC-based representa- tions are processed by LigBind and LigCys modules to generate predictions of reversible and cysteine-directed covalent ligand binding sites, respectively. LigCys generates predic- tions from the LigCys-S and LigCys-A mod- els, as well as a blended prediction that en- sembles results from both models. Option- ally, structure-aware versions of the predic- tions are provided. Results are visualized on the ESM3 predicted structure with potential MoRFs. Optionally, the predicted structure is processed by the pretrained transformers de- rived from the PeSTo architecture54 to gener- ate geometric representations, which are then integrated with the ESMC-based representa- tions as well as hand-engineered features to enhance predictions by LigBind and LigCys. Additionally, cysteine reactivity (using pKa as a proxy) and solvent accessibility (buried frac- tion) are annotated to inform ligand design. Evaluation on the consistently and hetero- geneously liganded cysteines from a re- cent ABPP study. 31 A recent ABPP study using over 400 cancer cell lines demonstrated that while a majority of the quantified cys- teines were consistently liganded across cell lines, a subset of cysteines showed varied ligandability depending on the cell line (re- ferred to as heterogeneous cysteine ligand- ability).31 We made predictions for 14 pro- teins containing the consistently and hetero- geneously liganded cysteines featured in the publication31 (DrugMap dataset) but absent in the LigCys training set. Note, although LigCysABPP database does not contain the DrugMap dataset, there is a significant over- lap in proteins/cysteines. LigCys predicted 14 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint UniProtID: P05141 ADP/ATP translocase 2 Reversable Binding ResidueCovalently Ligandable Cysteine C129 (Unligandable) C257 (Top-3) pKa=8.2; exposure=0% C57 (Top-1) pKa=7.7; exposure=0% C347 (Top-1) pKa=6.5; exposure=25% C160 (Top-2) pKa=8.1; exposure=0% UniProtID: P0DPH7 Tubulin alpha-3C b c a Figure 5: Architecture of the AI protein profiler (AiPP) for annotating covalent and reversible ligand binding sites given a protein sequence. a. Starting from a protein sequence, AiPP generates a predicted 3D structure (ESM3), structural and functional disorder annotations (RIDA), and ESMC-based representations. These representations are processed by three specialized modules: LigCys, LigBind, and KaML-ESM to predict covalent, reversible ligand binding residues, cysteine reactivities, respectively. Results are mapped onto the ESM3 predicted structure. Optionally, the predicted structure is processed by a transformer model (PeSTo) to generate geometric representations, which are then integrated with the pLLM-based embeddings as well as hand-engineered features to generate LigCys-SA and LigBind-SA models. b. For TUBA3C, LigCys predicted C347 as Top-1 ligandable, supporting the finding that it is consistently liganded across cell lines. 31 C347 is at the binding interface with β- tubulin in the adjacent dimer of the protofilaments that comprise microtubule structures. For visualization, we added a surface rendering of β-tubulin based on the electron diffraction model (PDB ID: 1TUB). 55 c. For SLC25A5, LigCys predicted C57, C160, C257, and C129 as Top-1, Top-2, and Top-3 ligandable, respectively, while C129 was predicted as unligandable, consistent with the order of engagement frequency across cell lines.31 tubulin alpha-3C (TUBA3C, UID P0DPH7) Cys347 and SOX10 (UID P56693) Cys71 as Top-1 ligandable (Fig. 5b), corroborating the consistent engagement across different can- cer cell lines. 31 Although this comparison is limited, the data suggest that LigCys models can confidently predict consistently ligandable cysteines. Based on the electron microscopy model of α/β-tubulin dimer, 55 Cys347 is positioned at the binding interface with β-tubulin from the adjacent dimer in the protofilaments that com- prise the microtubule structures. Given the disease relevance of tubulin proteins, the con- vergence of computational predictions and experimental evidence for Cys347 ligandabil- ity suggests potential drug design opportu- nity for disrupting or stabilizing this protein- 15 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint protein interaction (PPI). SOX10 transcription factor is a major dependency of melanoma cells.31 Notably, the compelling ligandability of Cys71, which is located at the dimerization domain, facilitated a proof-of-concept ligand design that stabilizes the PPI, leading to the disruption of melanoma transcriptional signal- ing.31 Among the heterogeneously liganded cys- teines31 not in the LigCys training set, Lig- Cys predictions vary. While PCBP1 (Q15365) Cys109, LARS1 (Q9P2J5) Cys70, and RPS3A (P61247) Cys201, are predicted as non-ligandable, SLC25A5 (P05141) Cys257 and ACAT1 (A0A5F9ZI66) Cys413 were each predicted as the Top-2 ligandable cysteine in their respective proteins. Interestingly, closer examination of our prediction and the pub- lished ABPP data 31 for SLC25A5 revealed that the LigCys ranking of the 3 predicted lig- andable cysteines (Fig. 5c) is identical to the order of the positive engagement frequency across cell lines. Note, a positive engage- ment is defined by ≥60% engagement, cor- responding to the commonly used CR thresh- old of 4. 31 While this level of agreement may be anecdotal given our limited comparison, the data suggest that LigCys models may al- ready capture some information about hetero- geneous ligandability through the prediction ranking. While our LC3D-based evaluations focused solely on the Top-1 predictions, users can explore additional candidates by exam- ining cysteines ranked 2nd, 3rd, or lower in their ligandability. Concluding Discussion We present the first unified AI platform to en- able sequence-based, scalable annotation of reversible and covalent ligand-binding sites across the proteome. This platform is en- abled through the new ML-ready databases and methodological advances that address several critical challenges in proteome-wide ML. By incorporating positive, negative, and un- quantified ABPP records from 15 publica- tions, the LigCysABPP database substan- tially exceeds the scope and comprehensive- ness of existing ABPP databases, including CysDB,43 DrugMap, 31 and TopCysteineDB.32 LigCysABPP comprises 608,898 quantified and 94,237 unquantified cysteine records of 10,649 unique proteins. Of these proteins, 6,502 contain at least 1 liganded cysteine (a total of 15,559 liganded cysteines) while 4,147 contain only unliganded and unquan- tified cysteines. Building upon BioLip2 52 the LigBind3D database offers more comprehen- sive coverage of protein-ligand interactions than commonly used binding databases such as BindingDB. 53 LigBind3D contains anno- tations for reversible ligand-binding residues across 1,998 unique proteins derived from co-crystal structures. Together with LC3D (a refined database of covalently liganded cysteines evidenced by co-crystal structures), LigCysABPP and LigBind3D provide solid foundations for developing ML models aimed at proteome-wide discovery of covalent (Lig- Cys) and reversible (LigBind) ligand-binding sites. A major bottleneck in developing ML mod- els for predicting ligand binding residues is experimental label variability (e.g., conflicting annotations for ligandable cysteines) and in- completeness (lack of high-confidence nega- tives for covalent or reversible ligand-binding residues). We addressed these issues by de- veloping a pLLM representation-based clus- tering approach to reconcile and augment la- bels. To develop robust LigCys models from ABPP data, we established a consensus anal- ysis framework evaluating models trained on data at varying consensus thresholds, using LC3D as an orthogonal validation dataset to identify the most reliable baseline. Building from this optimized baseline, we implemented two iterative approaches – LC3D- (orthogonal validation) and ABPP-guided (inherent valida- tion) – to systematically increase the train- ing set size and diversity, achieving substan- tial improvements in model performance and generalization. The structural requirement of ML mod- els significantly limits practical applications. 16 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint We addressed this constraint by leverag- ing state-of-the-art pLLM ESMC 45 to de- velop LigCys and LigBind models capable of rapid, sequence-based prediction of co- valent and reversible ligand-binding sites for proteome-scale exploration. The LigCys models achieved 79.9% Top-1 recovery of all known liganded cysteines from co-crystal structures. The LigBind models demonstrated strong performance on the hold-out test set of 129 proteins, achieving an AUROC of 93.9%, AUPRC of 74.5%, and Top-10 recovery rate of 78.1%. LigCys and LigBind models are integrated into AiPP , a unified AI platform designed for proteome-scale annotation of covalent and reversible ligand-binding sites. Addition- ally, AiPP incorporates several complemen- tary predictive tools: a meta disorder pre- dictor RIDA (based on RIDAO 56) for identi- fying MoRFs critical to PPI drug design, a cysteine reactivity predictor 47,57 for additional functional studies, and ESM3 48 for protein structure prediction. The integration of struc- tural information (LigBind-SA) enhances bind- ing site predictions and provides users with vi- sualization capabilities. Importantly, AiPP can be systematically improved (e.g., by incorpo- rating continuously growing data from ABPP experiments) and expanded (e.g., by adding prediction capabilities for lysine and other co- valently ligandable residues). Current work has several caveats that can be addressed in the future. Although limited comparison to the recent DrugMap work 31 suggest that LigCys models may already capture some information about heteroge- neous ligandability through the prediction ranking system, current LigCys models are currently agnostic to cellular context, e.g., redox conditions, post-translational modifi- cations, and mutational states. Two recent ABPP studies revealed that phosphoryla- tion33 and specific cancer-cell environment 31 can significantly modulate cysteine reactiv- ity and/or ligandability in a subset of cys- teines. The underlying mechanisms may in- volve phosphorylation-induced protein con- formational changes, shifts in cellular redox state, and accumulated cysteine-proximal mutations in malignant cells. 31,33 Molecular dynamics (MD) simulations 58,59 have demon- strated that the nucleophilicity of cysteines (and other residues such as lysines) can be conformation-dependent, with kinases pro- viding clear examples of how reactivity varies between functionally relevant conformational states (e.g., DFG-in vs. DFG-out forms). We envision that the contextual information can be added to the training data and at inference stage to provide proteoform-specific predic- tions in the future. This enhancement may re- solve the divergent performance observed be- tween models trained on datasets expanded using LC3D- vs. ABPP-guided strategies. An- other limitation of current AiPP is the LigBind model which was trained using a database consisting of monomer proteins. In addition to addressing these limitations, we also envision coupling AiPP with ligand docking and gener- ation models to advance proteome-wide drug design. Finally, the representation-based la- bel analysis and augmentation approaches developed in this work are broadly applica- ble for ML-experiment integration, establish- ing a generalizable framework for develop- ing proteomics-derived ML models across di- verse applications. Supplemental Materials Supplemental Methods contains detailed pro- cedures for constructing the LigCysABPP , LC3D, and LigBind3D databases; pLLM- based data clustering, label assignment and reconciliation; LigCys and LigBind model ar- chitectures, training, and validation; and Lig- Cys training data expansion. Supplemental Figures contains a more detailed analysis ac- companying Figure 1; model performance analysis using the ABPP hold-out set. Supplemental Data A Supplemental Data Excel file contains the following 20 sheets: 17 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint Benchmarking studies: ESMC representa- tion (SD2), consensus (SD3), source abla- tion (SD4), source prioritization (SD5), option- ablation (SD20) and split-level noise (SD21). Data extraction logs: ABPP extraction log (SD6), log of ambiguous and multivalue entry splitting (SD7), ABPP statistical report (SD8), ABPP orphaned records (dropped due to lack of unambiguous support) (SD9). Sequence and clustering records: validated UID sequences in FASTA format (SD10), UID deduplication log (SD11), precluster records (SD12), cysteine (Cys) clusters re- port (SD13). Production distillation: 4S–4R distillation re- port (SD14). Support reports: UID-level (SD15), ROI-level (SD16), and representative-level (SD17). ABPP hold-out set details: UID-level (SD18) and ROI-level (SD19). Note: Production runs refer to combined ABPP and LC3D data processed in a single distillation workflow, as illustrated in Fig. 1a. Data Availability Databases and models are available at https://github.com/JanaShenLab/AiPP/, which also contains a link to the cor- responding web server https://aipp. computchem.org/. Two data files accom- pany this manuscript: the Supplemental Data spreadsheet (SD2–SD21) and the LigCys- ABPP v1.xlsx spreadsheet; both are mirrored with code, models, and data in the repository. Acknowledgments Research reported in this publication was supported by the National Institute Of Gen- eral Medical Sciences (R35GM148261) and National Cancer Institute (R01CA256557) of the National Institutes of Health. The con- tent is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. We thank Benjamin Cravatt from the Scripps Re- search Institute for insightful comments and discussion. We thank Chetan Mishra and Marius Wiggert from Evolutionary Scale for facilitating the usage of ESM Cambrian and ESM3. This work was supported by an Evolu- tionaryScale compute grant.

(1) Aebersold, R. et al. How Many Human Proteoforms Are There? Nat. Chem. Biol. 2018, 14, 206–214. (2) Wishart, D. S. et al. DrugBank 5.0: A Major Update to the DrugBank Database for 2018. Nucleic Acids Res. 2018, 46, D1074–D1082. (3) Liu, Y .; Patricelli, M. P .; Cravatt, B. F . Activity-Based Protein Profiling: The Serine Hydrolases. Proc. Natl. Acad. Sci. U.S.A. 1999, 96, 14694–14699. (4) Greenbaum, D.; Medzihradszky, K. F .; Burlingame, A.; Bogyo, M. Epoxide Elec- trophiles as Activity-Dependent Cysteine Protease Pro¢ling and Discovery Tools. Chem. Biol. 2000, 7, 569–581. (5) Chan, W. C.; Sharifzadeh, S.; Buhrlage, S. J.; Marto, J. A. Chemo- proteomic Methods for Covalent Drug Discovery. Chem. Soc. Rev. 2021, 50, 8361–8381. (6) Cravatt, B. F . Activity-Based Protein Pro- filing – Finding General Solutions to Spe- cific Problems. Isr. J. Chem. 2023, (7) Weerapana, E.; Wang, C.; Simon, G. M.; Richter, F .; Khare, S.; Dillon, M. B. D.; Bachovchin, D. A.; Mowen, K.; Baker, D.; Cravatt, B. F . Quantitative Reactivity Pro- filing Predicts Functional Cysteines in Proteomes. Nature 2010, 468, 790–795. (8) Backus, K. M.; Correia, B. E.; Lum, K. M.; Forli, S.; Horning, B. D.; Gonz´alez-P´aez, G. E.; Chatterjee, S.; 18 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint Lanning, B. R.; Teijaro, J. R.; Ol- son, A. J.; Wolan, D. W.; Cravatt, B. F . Proteome-Wide Covalent Ligand Dis- covery in Native Biological Systems. Nature 2016, 534, 570–574. (9) Roberts, A. M.; Miyamoto, D. K.; Huff- man, T. R.; Bateman, L. A.; Ives, A. N.; Akopian, D.; Heslin, M. J.; Contr- eras, C. M.; Rape, M.; Skibola, C. F .; No- mura, D. K. Chemoproteomic Screening of Covalent Ligands Reveals UBA5 As a Novel Pancreatic Cancer Target. ACS Chem. Biol. 2017, 12, 899–904. (10) Weerapana, E.; Speers, A. E.; Cravatt, B. F . Tandem Orthogonal Proteolysis-Activity-Based Protein Pro- filing (TOP-ABPP)—a General Method for Mapping Sites of Probe Modification in Proteomes. Nat. Protoc. 2007, 2, 1414–1425. (11) Wang, C.; Weerapana, E.; Blewett, M. M.; Cravatt, B. F . A Chemo- proteomic Platform to Quantitatively Map Targets of Lipid-Derived Electrophiles. Nat. Methods 2014, 11, 79–85. (12) Bar-Peled, L. et al. Chemical Proteomics Identifies Druggable Vulnerabilities in a Genetically Defined Cancer. Cell 2017, 171, 696–709.e23. (13) Vinogradova, E. V. et al. An Activity- Guided Map of Electrophile-Cysteine In- teractions in Primary Human T Cells. Cell 2020, 182, 1009–1026.e29. (14) Cao, J.; Boatner, L. M.; Desai, H. S.; Burton, N. R.; Armenta, E.; Chan, N. J.; Castell´on, J. O.; Backus, K. M. Multi- plexed CuAAC Suzuki–Miyaura Label- ing for Tandem Activity-Based Chemo- proteomic Profiling. Anal. Chem. 2021, 93, 2610–2618. (15) Kuljanin, M.; Mitchell, D. C.; Schweppe, D. K.; Gikandi, A. S.; Nusinow, D. P .; Bulloch, N. J.; Vino- gradova, E. V.; Wilson, D. L.; Kool, E. T.; Mancias, J. D.; Cravatt, B. F .; Gygi, S. P . Reimagining High-Throughput Profiling of Reactive Cysteines for Cell-Based Screening of Large Electrophile Li- braries. Nat. Biotechnol. 2021, 39, 630–641. (16) Y an, T.; Desai, H. S.; Boatner, L. M.; Y en, S. L.; Cao, J.; Palafox, M. F .; Jami- Alahmadi, Y .; Backus, K. M. SP3-FAIMS Chemoproteomics for High-Coverage Profiling of the Human Cysteinome**. ChemBioChem 2021, 22, 1841–1851. (17) Tao, Y .; Remillard, D.; Vino- gradova, E. V.; Y okoyama, M.; Banchenko, S.; Schwefel, D.; Melillo, B.; Schreiber, S. L.; Zhang, X.; Cra- vatt, B. F . Targeted Protein Degradation by Electrophilic PROTACs That Stereos- electively and Site-Specifically Engage DCAF1. J. Am. Chem. Soc. 2022, 144, 18688–18699. (18) Y ang, F .; Jia, G.; Guo, J.; Liu, Y .; Wang, C. Quantitative Chemopro- teomic Profiling with Data-Independent Acquisition-Based Mass Spectrometry. J. Am. Chem. Soc. 2022, 144, 901–911. (19) Y an, T.; Boatner, L. M.; Cui, L.; Tontonoz, P . J.; Backus, K. M. Defining the Cell Surface Cysteinome Using Two- Step Enrichment Proteomics. JACS Au 2023, 3, 3506–3523. (20) Burton, N. R.; Polasky, D. A.; Shik- wana, F .; Ofori, S.; Y an, T.; Geis- zler, D. J.; Veiga Leprevost, F . D.; Nesvizhskii, A. I.; Backus, K. M. Solid- Phase Compatible Silane-Based Cleav- able Linker Enables Custom Isobaric Quantitative Chemoproteomics. J. Am. Chem. Soc. 2023, 145, 21303–21318. (21) Koo, T.-Y .; Lai, H.; Nomura, D. K.; Chung, C. Y .-S. N-Acryloylindole-alkyne (NAIA) Enables Imaging and Profiling New Ligandable Cysteines and Oxidized Thiols by Chemoproteomics. Nat. Com- mun. 2023, 14, 3564. 19 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint (22) Burton, N. R.; Backus, K. M. Functional- izing Tandem Mass Tags for Streamlin- ing Click-Based Quantitative Chemopro- teomics. Commun. Chem. 2024, 7, 80. (23) Njomen, E.; Hayward, R. E.; De- Meester, K. E.; Ogasawara, D.; Dix, M. M.; Nguyen, T.; Ashby, P .; Si- mon, G. M.; Schreiber, S. L.; Melillo, B.; Cravatt, B. F . Multi-Tiered Chemical Pro- teomic Maps of Tryptoline Acrylamide– Protein Interactions in Cancer Cells. Nat. Chem. 2024, 16, 1592–1604. (24) Tian, C.; Sun, L.; Liu, K.; Fu, L.; Zhang, Y .; Chen, W.; He, F .; Y ang, J. Proteome-Wide Ligandability Maps of Drugs with Diverse Cysteine-Reactive Chemotypes. Nat. Commun. 2025, 16, 4863. (25) Biggs, G. S. et al. Robust Proteome Profiling of Cysteine-Reactive Frag- ments Using Label-Free Chemopro- teomics. Nat. Commun. 2025, 16, 73. (26) Abegg, D.; Frei, R.; Cerato, L.; Prasad Hari, D.; Wang, C.; Waser, J.; Adibekian, A. Proteome-Wide Profiling of Targets of Cysteine Reactive Small Molecules by Using Ethynyl Benziodox- olone Reagents. Angew. Chem. Int. Ed. 2015, 54, 10852–10857. (27) Shannon, D. A.; Weerapana, E. Covalent Protein Modification: The Current Land- scape of Residue-Specific Electrophiles. Curr. Opin. Chem. Biol.2015, 24, 18–26. (28) Long, M. J. C.; Aye, Y . Privileged Elec- trophile Sensors: A Resource for Cova- lent Drug Development. Cell Chem. Biol. 2017, 24, 787–800. (29) Wang, S.; Tian, Y .; Wang, M.; Wang, M.; Sun, G.-b.; Sun, X.-b. Advanced Activity-Based Protein Profiling Applica- tion Strategies for Drug Development. Front. Pharmacol. 2018, 9. (30) White, M. E.; Gil, J.; Tate, E. W. Proteome-Wide Structural Analysis Identifies Warhead- and Coverage- Specific Biases in Cysteine-Focused Chemoproteomics. Cell Chem. Biol. 2023, 30, 828–838.e4. (31) Takahashi, M. et al. DrugMap: A Quanti- tative Pan-Cancer Analysis of Cysteine Ligandability. Cell 2024, 187, 2536– 2556.e30. (32) Bonus, M.; Greb, J.; Majmudar, J. D.; Boehm, M.; Korczynska, M.; Nazemi, A.; Mathiowetz, A. M.; Gohlke, H. TopCys- teineDB: A Cysteinome-wide Database Integrating Structural and Chemopro- teomics Data for Cysteine Ligandabil- ity Prediction. J. Mol. Biol. 2025, 437, 169196. (33) Kemper, E. K.; Zhang, Y .; Dix, M. M.; Cravatt, B. F . Global Profiling of Phosphorylation-Dependent Changes in Cysteine Reactivity. Nat. Methods 2022, 19, 341–352. (34) Mann, M.; Kulak, N. A.; Nagaraj, N.; Cox, J. The Coming Age of Complete, Accurate, and Ubiquitous Proteomes. Mol. Cell 2013, 49, 583–590. (35) Gao, M.; Moumbock, A. F . A.; Qaseem, A.; Xu, Q.; G ¨unther, S. CovPDB: A High-Resolution Coverage of the Covalent Protein–Ligand Inter- actome. Nucleic Acids Res. 2022, 50, D445–D450. (36) Guo, X.-K.; Zhang, Y . CovBinder- InPDB: A Structure-Based Covalent Binder Database. J. Chem. Inf. Model. 2022, 62, 6057–6068. (37) Liu, R.; Clayton, J.; Shen, M.; Bhatna- gar, S.; Shen, J. Machine Learning Mod- els to Interrogate Proteome-Wide Co- valent Ligandabilities Directed at Cys- teines. JACS Au 2024, 4, 1374–1384. (38) Jumper, J. et al. Highly Accurate Protein Structure Prediction with AlphaFold. Na- ture 2021, 596, 583–589. 20 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint (39) Zhang, W.; Pei, J.; Lai, L. Statisti- cal Analysis and Prediction of Covalent Ligand Targeted Cysteine Residues. J. Chem. Inf. Model. 2017, 57, 1453–1460. (40) Reimer, B. M.; Awoonor-Williams, E.; Golosov, A. A.; Hornak, V. Cov- CysPredictor: Predicting Selective Co- valently Modifiable Cysteines Using Pro- tein Structure and Interpretable Machine Learning. J. Chem. Inf. Model. 2025, 65, 544–553. (41) Du, H.; Jiang, D.; Gao, J.; Zhang, X.; Jiang, L.; Zeng, Y .; Wu, Z.; Shen, C.; Xu, L.; Cao, D.; Hou, T.; Pan, P . Proteome-Wide Profiling of the Covalent-Druggable Cysteines with a Structure-Based Deep Graph Learn- ing Network. AAAS Research 2022, 2022, 9873564. (42) Abramson, J. et al. Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3. Nature 2024, 630, 493–500. (43) Boatner, L. M.; Palafox, M. F .; Schweppe, D. K.; Backus, K. M. CysDB: A Human Cysteine Database Based on Experimental Quantitative Chemoproteomics. Cell Chem. Biol. 2023, 30, 683–698.e3. (44) Boatner, L. M.; Eberhardt, J.; Shik- wana, F .; Holcomb, M.; Lee, P .; Houk, K. N.; Forli, S.; Backus, K. M. CIAA: Integrated Proteomics and Struc- tural Modeling for Understanding Cys- teine Reactivity with Iodoacetamide Alkyne. ACS Chem. Biol. 2025, 20, 1669–1682. (45) ESM Team. ESM Cambrian: Re- vealing the Mysteries of Pro- teins with Unsupervised Learning. https://www.evolutionaryscale.ai/blog/esm- cambrian. (46) Uhlen, M.; Oksvold, P .; Fagerberg, L.; Lundberg, E.; Jonasson, K.; Fors- berg, M.; Zwahlen, M.; Kampf, C.; Wester, K.; Hober, S.; Wernerus, H.; Bj¨orling, L.; Ponten, F . Towards a Knowledge-Based Human Protein Atlas. Nat. Biotechnol. 2010, 28, 1248–1250. (47) Shen, M.; Dayhoff, G. W.; Shen, J. Pro- tein Electrostatic Properties Are Fine- Tuned Through Evolution. 2025. (48) Hayes, T. et al. Simulating 500 Mil- lion Y ears of Evolution with a Language Model. Science 2025, 387, 850–858. (49) Fu, L.; Niu, B.; Zhu, Z.; Wu, S.; Li, W. CD-HIT: Accelerated for Clus- tering the next-Generation Sequencing Data. Bioinformatics 2012, 28, 3150– 3152. (50) Vig, J.; Madani, A.; Varshney, L. R.; Xiong, C.; Socher, R.; Rajani, N. F . BERTology Meets Biology: Interpreting Attention in Protein Language Models. ICLR 2021. 2021. (51) Lin, Z.; Akin, H.; Rao, R.; Hie, B.; Zhu, Z.; Lu, W.; Smetanin, N.; Verkuil, R.; Ka- beli, O.; Shmueli, Y .; Fazel-Zarandi, M.; Sercu, T.; Candido, S.; Rives, A. Evolutionary-Scale Prediction of Atomic- Level Protein Structure with a Language Model. Science 2023, 379, 1123–1130. (52) Zhang, C.; Zhang, X.; Freddolino, L.; Zhang, Y . BioLiP2: An Updated Struc- ture Database for Biologically Relevant Ligand–Protein Interactions. Nucl. Acids Res. 2024, 52, D404–D412. (53) Gilson, M. K.; Liu, T.; Baitaluk, M.; Nicola, G.; Hwang, L.; Chong, J. Bind- ingDB in 2015: A Public Database for Medicinal Chemistry, Computational Chemistry and Systems Pharmacology. Nucl. Acids Res. 2016, 44, D1045– D1053. (54) Krapp, L. F .; Abriata, L. A.; Cort ´es Ro- driguez, F .; Dal Peraro, M. PeSTo: Parameter-Free Geometric Deep Learn- ing for Accurate Prediction of Protein 21 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint Binding Interfaces. Nat. Commun. 2023, 14, 2175. (55) L ¨owe, J.; Li, H.; Downing, K.; No- gales, E. Refined Structure ofαβ-Tubulin at 3.5 ˚A Resolution 1 1Edited by I. A. Wil- son. J. Mol. Biol. 2001, 313, 1045–1057. (56) Dayhoff II, G. W.; Uversky, V. N. Rapid Prediction and Analysis of Protein In- trinsic Disorder. Protein Sci. 2022, 31, e4496. (57) Shen, M.; Kortzak, D.; Ambrozak, S.; Bhatnagar, S.; Buchanan, I.; Liu, R.; Shen, J. KaMLs for Predicting Protein p K a Values and Ionization States: Are Trees All Y ou Need? J. Chem. Theory Comput. 2025, 21, 1446–1458. (58) Liu, R.; Yue, Z.; Tsai, C.-C.; Shen, J. As- sessing Lysine and Cysteine Reactivities for Designing Targeted Covalent Kinase Inhibitors. J. Am. Chem. Soc. 2019, 141, 6553–6560. (59) Liu, R.; Verma, N.; Henderson, J. A.; Zhan, S.; Shen, J. Profiling MAP Kinase Cysteines for Targeted Covalent Inhibitor Design. RSC Med. Chem. 2022, 13, 54– 63. 22 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted September 8, 2025. ; https://doi.org/10.1101/2025.09.07.670677doi: bioRxiv preprint