Comment: "A proteome-wide atlas of drug mechanism of action"

 

It's been a while, but I found a very interesting paper that I want to comment.

Proteomics is an invaluable tool for unlocking the secrets of compound polypharmacology: by quantifying the protein expression changes caused by different compounds, researchers can gain valuable insight into compound interactions at the cellular level. Proteomics in application to the understanding of compound polypharmacology is relatively well established: there are tons and tons of publications on this topic about single compounds or some SAR exploration with 10s of compounds max… and now this goldmine was published – “A proteome-wide atlas of drug mechanism of action”. 875 compounds, 875 proteomics profile after 24h treatment (in duplicate) on ~9000 proteins in HCT116 cell line. To be honest, nothing radically new was done in this paper, but the robustness, richness, and accessibility of data that were generated in this paper are, in my opinion, unprecedented. Articles like this have a global impact on drug design and the discovery process. The paper is also accompanied by a well-designed website https://wren.hms.harvard.edu/DeepCoverMOA/.

Notes:

Expected findings.

  • Every compound has polypharmacology: this paper indicates clearly that compounds, by default, have some form of polypharmacology – see cases 1-4 below. Rephrasing a famous quote “In chemical tool we trust. All others must bring data”. The complexity of a compound interacting with a cellular system cannot be described in full by biochemical assays, even by the best possible set of them.
  • Some compounds cause a very heavy expression impact on >25% of the proteome (and some are very specific). Those are common suspects like inhibitors of proteasomal degradation, transcription, and histone acetylation. E.g. vorinostat, pan-HDAC inhibitor,  modulates hundreds of genes in L1000 in variety of cell lines, same about bortezomib, proteasome inhibitor.

Not-so-obvious findings

  • Only 15% of the compounds regulate the expression of their main protein targets “by a median of around 1.56-fold, with upregulation occurring more frequently than downregulation”. I had a long conversation with a more experimental-savvy friend that confirmed that this finding is an expected one, under one important condition that I will discuss in a separate blog post.
  • Only 15 proteins are modulated on average by the compound “with upregulation occurring more frequently than downregulation”. That, to be honest, is way below my expectations of 42 or so.
  • “Most of the proteome is accessible to regulation by small molecules” and >4000 proteins found to be downregulated by at least one compound. So, the curse of inhibition and downregulation is hunting us again. The number of compounds increasing the expression of specific proteins is very low.
  • Most of the upregulation events were enriched with stress-related pathways linked to “overcome exogenous stressors through upregulation of autophagy and sterol metabolism”. 

The best way to describe this paper is to cite it directly: “…we defined a roadmap for how library-scale annotation of small-molecule fingerprints can be used in drug discovery”. There are two use cases meticulously described in the paper, + I added two more of my particular interest. I highly recommend reading this paper in depth.

Case 1: JP1302


Numbers are Tanimoto similarities to JP1302

JP1302 was developed as a “selective alpha-2C adrenoceptor antagonist” [ref]. ADRA2C is not detectable in HCT116 but causes an expression profile similar to RNA transcription inhibitors. Actually, adrenergic receptor subtypes are not detectable in HCT116 (Check notebooks for HCT116 overview). Authors found that JP1302 correlates with RNA transcription inhibitors: flavopiridol, AZD5438 and PHA767491, with main targets sharing no similarity to ADRA2C. Yet, all of top scoring compounds were downregulating “RNA polymerase II (Pol II) phosphorylation”. While correlation with a full profile (~9000 proteins) is a first choice to rank compounds, it always carries much noise that can be pruned to a specific set of genes that authors did, finding a CBL0137, a FACT complex inhibitor. Further comparison with a “positive control” for the global RNA transcription inhibitors showed that several protein expression changes of JP1302 are unique, thus confirming its selectivity. In the next steps, a dose-response study was done for several concentrations that revealed that low concentrations of the JP1302 “had no effect on Pol II phosphorylation and induced an increase in p21 expression and H2AX phosphorylation—a marker of DNA damage” and “dose-dependent proteome remodeling and found that the relative numbers of proteins up- and downregulated upon treatment diverged as JP1302 concentrations decreased, suggesting concentration-dependent differences in pathway activation”.

Case 2: RN486


Numbers are Tanimoto similarities to RN486

RN486 is a BTK inhibitor, and BTK is not expressed in HCT116 cell line, but “RN486 caused dramatic proteome remodeling, suggesting a non-BTK off-target”. Ibrutinib is another example of BTK inhibitor that causes significantly fewer expression changes (thou it’s not a very selective BTK inhibitor). And here comes into play a pathway-centric analysis showing that RN486 significantly enriched in “autophagy markers coupled with a decrease in lysosomal proteins reveals dysregulation of autophagy as the main phenotype in RN486-treated cells, which is a common mechanism of many kinase inhibitors”. The authors did not stop here: they did target-fishing using a solubility alteration assay (PISA) that revealed a TEX264 receptor that mediates ER-autophagy, showing a good agreement with pathway analysis.

 Case 3: close analogs

Numbers are Tanimoto similarities to Nutlin3a

Like with any biosignature readouts (HCI, gene expression, metabolomics): it helps to find unusual similarities between compounds. Highly correlated compounds by proteomics profiles might share no chemical similarity. For example, nutlin3a has top 2 hits (idasanutlin, MI-773) with a very similar substructure and pharmacophore, but the third hit (UNC2250) has no chemical similarity, and I really doubt any pharmacophore similarity, while all proteomics correlations are above 0.8. On a side note - nutlin3a is a “hurricane” compound, it induces up/down expression changes in > 365 proteins with an average number of 15 proteins.

Case 4: Fishing out activity cliffs



The only acceptable compound I found for this case is Tazarotenic Acid (AGN 190299) vs Tazarotene. Both compounds are RAR-β modulators; Tazarotene is being FDA approved for treating psoriasis and acne and metabolized to Tazarotenic Acid in vivo. Surprisingly, Tazarotene has two copies of the compound with identical MW and Formula, but not sharing proteins at 1.0 cutoff – a really strange bug. RAR-β has a low expression level in the HCT116 cell line, which is likely why we do not have a characteristic signal (transglutaminase, involucrin, and various keratins) for any of those compounds. But still, it is interesting to see an overlap between the drug and active metabolite: Only 4 proteins of a total of 44 are overlapping (abs value cutoff = 0.5) between Tazarotenic Acid and Tazarotene, indicating that small changes in chemical structure can induce a lot of changes on an expression level.

This publication and all publications in this blog are solely my own and do not express the views or opinions of my employer.

Comments

Post a Comment

Popular posts from this blog

Paper comment: "MELLODDY: cross pharma federated learning at unprecedented scale unlocks benefits in QSAR without compromising proprietary information"

Image
 Intro I finally got my hand on carefully reading papers by the MELLODDY consortium led by my ex-supervisor (and fantastic visionary) - Hugo Ceulemans. I participated in some very-very early discussions and saw the painful process of the project's initiation. What Hugo did is colossal, and he helped to build a PRECEDENT! Honestly, the amount of PR the project got is too little! (Compared to PR of some other ML-for-chemistry companies). The project united ten big pharma companies to build a federated multi-task model for activity assays (activity on proteins, PK/PD, toxicity):  40000+ assays on 21+ million compounds with 2.6+ billion end-points . MELLODDY used federated learning to improve individual models of the consortium partners by confidential sharing of the "model"  without sharing private data itself. An exciting approach that has a lot of tiny, small, and big challenges related to fairness, data security, MLOps, ML engineering, ML algorithms, and legal questions. ...
Read more

How many actual bits in unfolded Morgan fingerprint?

Image
The easiest  possible descriptors used in ML for chemistry are fragments, and they are usually used in a folded form. The default folded size is 1024, with some recommendations to use higher dimensionality (>=4096) for computational chemogenomics tasks.  Folding of descriptors is done for several reasons: to save space, get a uniform length, with a price of  bit collision , and sometimes insensitivity to small changes.  Let's check how many actual bits are in drugs as an example of FDA-approved chemical space vs random selection of 300K compounds from Enamine REAL virtual space of 3B compounds commonly used in virtual screening. Number of unique bits So, as we can see, on average, for radius 3, there are 16 bits per compound on average for drug space that's going down 4 and 2 bits for increased chemical space of 100K and 300K diverse compounds from the Enamine REAL library.  Enamine virtual space is, of course, repetitive, but it still indicates that with a l...
Read more

Kaggle Solubility competition: overview

Kaggle Solubility competition is over. It did draw the attention of 100 participants with some strong issues underlined by the participants. I decided to share my dive into this... Data: 70 711 compounds in train set: A very good and diverse dataset.   30 308 compounds in test set . Quality m etric: Kappa coefficient  (was it changed later?). Results Initial winner (by leaderboard) was a Ohue laboratory of Tokyo Institute of Technology with a combination of the BART transformer language model and GNN with focal loss  ( solution ).  The dramatic difference between the winner and the rest of the participants was noticeably big and raised a lot of questions. Bernhard Rohde (Novartis) was the second leaderboard winner. He underlined the same issue and showed that “plate/row/column” has a drastic effect on model quality (explicit IDs were provided: EOS12286). Still, the Bayesian Instance-Based Offset Learning solution he used draw my attention, because it can be app...
Read more