Genetic Interactions/Discussion/Version 1 discussions

2006-06-03, Alan Ruttenberg

Motivation for these is to to tighten restrictions, make intention clearer in OWL, validation derivable from ontology.
 * Restriction PARTICIPANTS min 2
 * Subproperty PARTICIPANTS: INTERACTING-GENE
 * Restriction INTERACTING-GENE all gene
 * Restriction INTERACTING-GENE min 2
 * Subproperty PARTICIPANTS: INTERACTING-MOLECULE (can't think of better name)
 * Restriction INTERACTING-MOLECULE all smallMolecule

Gary
 * participants >=2 is definitely good to put. Why create sub properties?

Alan
 * Otherwise you could fill in PARTICIPANTS with 2 smallMolecules and it would be considered valid. This could be ruled out with QCRs but we don't have them yet. The other two reasons are more stylistic. 1) We don't usually directly assert PARTICIPANTS. 2) For querying it it feels clearer to know where to look for genes, versus smallMolecules, since one is often focused on one or the other, though one can still do this query by adding an additional rdf:type clause.

Gary
 * It is legal to have any mix of genes or small molecules - they are effectively equivalent for genetic experiments. Note the examples given in the Biological Questions / Use Cases section

Alan
 * According to your email it is not legal to have any mix of genes or small molecules. There must be at least 2 genes, after which there may be any mix. It is not a genetic interaction experiment if there are no genes involved. Now the current proposal says that all PARTICIPANTS must be genes or smallMolecules, and if you include the min 2 restriction, you could create an instance of geneticInteraction, make the PARTICIPANTS two small molecules and nothing else - no genes and the instance will be well formed. By adding the subproperties and restrictions I propose, it is clear, from the ontology, that this doesn't make sense because we are explicitly represented what you stated in your email.

Gary
 * Sorry, I misunderstood your question in the haste of replying to your e-mail - you must have >2 participants - any mix of genes/small molecules as long as there are at least 2 participants. Having zero genes and 2 small molecules is still a genetic interaction.  Read the Haggarty reference for more information about small-molecule - small molecule genetic interactions.  Also, I am still working on the definitions, etc. and probably won't be done until tomorrow.

2006-06-03 Alan (starting to look at the proposal)
 * Are the phenotype symbols always associated with the participants in a genetic interaction. If not, could you give an example of phenotypeSymbol and associated genetic participant that would not be part of the the interaction participants?
 * Gary: yes.
 * Alan: Why have a separate instance for the phenotype symbol then, rather than having it be a property of the participant?
 * Gary: where would you put it as a participant property? Alan: One the instances that are the value of PARTICIPANTS of the interaction. Probably as an annotation property. But better if (as suggested below) the symbols are dispensed with and the instances become terms in the inequality directly.
 * A number of properties are rather broadly named, e.g. PHENOTYPE. Suggest that the domains on these properties be removed to allow use in other contexts, where appropriate, or make the property name more specific.
 * Gary: Do you have a suggestion for a more specific term?
 * Alan: I'll think about it. But my recommendation is that it is better to drop the domain. See Matthew's japan presentation for rationale.
 * Gary: what do you mean by 'drop the domain'?
 * Alan: The domain of PHENOTYPE is geneticEvidence, which means that anything that has a PHENOTYPE property is inferred to be of type geneticEvidence. I'm suggestion not having a domain for that property.
 * Gary: sorry, I don't understand why you would do this. every property has a domain, otherwise it's not used by any class.
 * Alan: If a property doesn't have a domain it can be used by any class. This is a common error. See http://www.co-ode.org/resources/papers/CommonErrorsInOWL.pdf slide 16. or http://www.w3.org/TR/owl-semantics/syntax.html#2.3: Properties can be equivalent to or sub-properties of others; can be made functional, inverse functional, symmetric, or transitive; and can be given global domains and ranges. However, most information concerning properties is more naturally expressed in restrictions, which allow local range and cardinality information to be specified
 * Gary: ok, I understand that, but what is the rationale for doing that here (or anywhere for that matter)? phenotype is required for the geneticEvidence class, so the geneticEvidence class is the domain of the PHENOTYPE property. (By the way, I'm still thinking closed world - maybe I should make that clear on this proposal)
 * Alan: I Don't understand what you are saying. It says it's an error. Are you asking why you shouldn't have an error here? (for the same reason you don't want to have an error anywhere). The correct consequent of the statement "Phenotype is required for the geneticEvidence class" is "The geneticEvidence class is a subclass of those things which have some value for the PHENOTYPE property", i.e. a restriction . Not sure what closed-world has to do with this.
 * Having the inequality expression be a string is fragile and error-prone. Could easily be expressed in a structured way. For example, you could review the list implementation by co-ode here: http://www.cs.man.ac.uk/~drummond/presentations/owl-lists.ppt
 * Gary: What would be the list element and how would you represent it? What do you think of the alternative representation in the proposal?
 * Alan: Will review when I have some time.
 * For review of these proposals, particularly when they are additive, it would be helpful to either add an annotation property with a change comment to each new class and property, so one can easily identify differences from the surrounding previous version BioPAX, or even better, to structure the ontology proposal as an owl file with just the new stuff, which owl:imports the base version of BioPAX.
 * Gary: good idea, I'll see what I can do.
 * I have some concerns about the confusion that may arise with bother entities and physical entity participants both being allowed to be PARTICIPANTS. For instance with CONTROLLER, which similarly doesn't need annotation, we still use a physicalEntityParticipant. State proposal may speak to this, either helping or making more confusing.
 * Gary: Where is this an issue with this proposal? physicalentityparticipants are not used.
 * Alan: The point is that they are used every where else. What makes the participants in this reaction different than that, for example, pointed to by a CONTROLLER property?
 * Gary: Does note #5 in the proposal answer your question?
 * Alan: I don't think so. Why does one need to annotate the edge of a CONTROLLER property? It's a matter of consistency.
 * Gary: there is no reason by you need physical entity participant for genetic interactions, why complicate the model with added instances that have properties that will never be used. e.g. stoichiometry and cellular location are not useful for genetic interactions.  post-translational modifications may be useful - I will have to see if I can find examples.
 * Alan: by "not complicating the model", you complicate it's use. Now, in order to do queries one has to know, for each class, what kind of thing you will get back. This kind of inconsistency is the bane of users of file formats. If a model doesn't clearly articulated its design principals and then follow them consistently, it is poorly designed. I'm afraid that BioPAX is falling into this trap.
 * Gary: this sounds like a style matter - probably a good discussion point at CSHL.
 * Who are the stakeholders for this proposal - is there a current database which has genetic interaction which will serve as a test bed for this proposal?
 * Gary: good question - other than the databases listed in the proposal + various model organism databases, there are my lab, Boone lab (UofT), Andrews lab (UofT), Sander lab, Roth lab (Harvard) + textmining groups (Rzhetsky lab @ Columbia). Most of these people identified themselves at the Jan 2005 meeting in NYC, however, they are not as active, so I will probably have to follow up with reviewing this proposal with them offline.  Depending on interest of the participants at CSHL, we may elect to just present it with minimal discussion.

Alan: Regarding genetic background. Much of the genetic background discussed which genes have been inactivated, mutated or deleted (http://drnelson.utmem.edu/yeastlect1.html). Wouldn't someone care to search on these, and if so should they not be at least minimally recorded in some representation that encodes the genes as gene instances (even if the string is kept for display purposes). Scanning the reference it seems that it would be adequate to record 1) The strain 2) The mating type 3) A set of geneticBackgroundGeneModification instances, each of which pointed to an instance of gene, with the modification taken from a controlled vocabulary:, partialDeletion, completeDeletion, inactivation, mutation, markerInsertion, etc.

Gary: this is almost exactly the representation used in BIND and it was basically never used, even though 1000s of records were created, so I elected for something simpler: a genetic background string. This will likely hardly be used either (except by BIND converters and for displaying on a web page). There is a lot of genetics not taken into account by this proposal - see the future directions and notes sections.

Alan: What does never used mean? That the nothing was encoded in the fields, or that the fields were not used? Or something else.

Gary: never used for any computational task that data that structured would have been useful for - it was only used to display on a web page and most people were not that interested in having the data displayed in that level of detail. Lots of records had data encoded in these fields.

Alan: How do you know? - weren't there quite a few users of bind? Gary: from speaking to curators who spoke to users - as far as I know, there was never a formal survey done. A better question is can you find uses for this information?

Alan: Could you supply a worked example of a chemical only genetic interaction? I read the Haggerty paper and would like to see how results from it it would play out in the representation.

Alan: There are a couple of questions that occur to me as I think about this proposal, particularly triggered by the inclusion of compounds in the participants of the "interaction". One has to do with the boundary between evidence and conclusion. These findings border more on experiment description than conclusion. For example, the conclusions that one can derive in the three kinds of genetic interactions described here vary. In the compound-gene experiments, the conclusions that the author proposes have to do with what the targets of the compounds are, with the experiment showing that the compound is lethal in combination with the same target that some other gene deletion is synthetic lethal with. So I wonder, does the representation record as participants the the compound and inferred target, a conclusion, or does it record that a compound in combination with a gene deletion is lethal, an experiment which is used to reach the conclusion. (A second hypothesis assigns the function "involved in multidrug resistence" to those genes, that, when knocked out were senstitive to all the compounds)
 * Gary: It records the compound in combination with a gene deletion is lethal. This is conceptually equivalent to the classical gene-gene genetic interactions, as shown in figure 1 of the paper, though with small molecules, there can be multiple unknown genes perturbed. See notes 8 and 12.

The second question has to do with the idea that one might legitimately call an interaction with only two small molecules a genetic interaction. For instance consider two compounds which interact with each other. If the compounds are non lethal alone, but a product of their reaction is toxic, would this be called a genetic interaction involving two compounds? In the Haggerty paper, my read is that the paired compound activity across the various deletion mutants is used as a metric two derive two sorts of conclusions. By showing that, across all the pairs of compounds, two deletion strains behave similarly, we are suggesting that they have related function, and so the relation is between two genes. Conversely, by showing a similar pattern of the compounds effect across different strains one can hypothesis that two compounds have an important similarity. In this case, the relation would be between two compounds, but wouldn't properly be called a genetic interaction (they make analogy that such a relation is similar to molecular feature used to describe similarity of compounds)
 * Gary: The toxic case you bring up is a good one, however, you would expect that to be caught using controls. You are correct about the conclusions drawn, however, this proposal just captures the chemical genetic interactions, shown in figure 1 of the paper. There are 2 types: 1) chemical combination on wild type background - if the cells die, there is an interaction and 2) chemical combination on gene deletion background - if the cells die, there is an interaction between the gene and the 2 chemicals.
 * Alan: To ask you a question you pose yourself, what is the intended use of this information? Can the conclusions that the papers arrive at be derived from your representation proposal in a straightforward way? Would doing so be considered a valid use case for BioPAX?
 * The similarity interactions, among similarly interacting compounds, are not captured by this proposal. It may be interesting to capture those in the interaction class at a future date.
 * On second thought - about the toxic case, there is another case where the chemical could be metabolized to something else, which is the active chemical. This would not be caught by controls and is a general caveat of the method.  Though more genetic analysis would eventually uncover the issue. [There are many cases - Alan]
 * Gary: Can you list the cases you have in mind? This is interesting for general background information for the proposal.
 * Alan: Following up on my point about what can be called a genetic interaction, I spoke to some biologists at work and they also think that for something to be called a genetic interaction, it should involved genes. Sometimes the interaction of a compound and a gene might be called a genetic interaction, when the compound is known to target a specific gene. In that case, the interaction should have the genes as participants and the fact that the compound is a proxy for the gene part of the evidence.
 * Gary: correct, however you mostly don't know the genes - how do you create an interaction with an unknown participant, but a known state (if the state corresponds to the perturbation)? I think this would work better with state variables that the current implementation.
 * Alan: Couple of things. First, you do you want to mix arbitrary phenotypes with state? The state proposal is currently oriented towards molecules, not e.g. hair color of mice. 2) The issue isn't that you don't know the genes, it's that you don't know whether the genes. Another way to put it is this: Suppose you know the mechanism is that a small molecule binds to a an essential kinase. Do you call that a genetic interaction? We currently would call this a modulation. Should the user have the choice of representing the same things both ways? I think that the thing that characterizes genetic interactions is partially the type of experiment, and partially the fact that we know the genes that are involved, but we don't necessarily know how they are involved - by regulation, or by action on a product etc. Otherwise I worry that the term genetic interaction is meaningless - virtually all interactions could be considered genetic interactions, after all, what doesn't have to do with the genes. (you could answer: interactions between epigenetic phenomena and compounds, but then, in this proposal, those would be classified as genetic interactions, in a similar manner to the toxicity case mentioned below)
 * It's not that it's not important information to know these other kinds of things, it's just that they ought to be called something else. I think they are part of a more general type of indirect control, where the thing controlled, instead of reaction, is some phenotype. As another example, another name one gives to the interaction of two molecules which express a phenotype is a synergistic effect. For example, one of the types of screens we and other people do is try to find drugs which interact synergistically, as candidates for combination therapies. The only difference that I can see between this sort of screen and the chemical genetic screen is the intention of the investigator.
 * Definitions of genetic interactions on the web seem to agree with this interpretation. Some citations (4 of the top 5 google hits give a definition)
 * http://genomebiology.com/2005/6/4/R38 A genetic interaction is the interaction of two genetic perturbations in the determination of a phenotype. Genetic interaction is observed in the relation among the phenotypes of four genotypes: a reference genotype, the 'wild type'; a perturbed genotype, A, with a single genetic perturbation; a perturbed genotype, B, with a perturbation of a different gene; and a doubly perturbed genotype, AB. Gene perturbations may be of any form (such as null, loss-of-function, gain-of-function, and dominant-negative). Also, two perturbations can interact in different ways for different phenotypes or under different environmental conditions.
 * http://www.nature.com/nbt/journal/v23/n5/abs/nbt1096.html Genetic interaction analysis,in which two mutations have a combined effect not exhibited by either mutation alone, is a powerful and widespread tool for establishing functional linkages between genes.
 * http://www.yeastgenome.org/help/PhenoGenIntDetailsHelp.html Genetic interaction data are presented in a table that contains the following columns: Bait/query, Interaction Type, Hit, Database source, Phenotype, Notes, and Reference(s). Bait/query indicates the gene that was used as the Bait when the interaction was detected. In genetic interaction experiments, the Bait would be the starting strain or construct. Gene names are hyperlinked to their respective locus pages.[...] Hit indicates the gene participating in the genetic interaction, and the types of mutant alleles involved (e.g., null or point). In genetic interaction experiments, the Hit would be the gene identified in the screen (i.e., a suppressor would always be the Hit). Gene names are hyperlinked to their respective locus pages.
 * http://en.wikipedia.org/wiki/Genetic_interaction Epistasis takes place when the action of one gene is modified by one or more others that assort somewhat independently.[...] Epistasis and genetic interaction refer to the same phenomenon.
 * Gary: these definitions are a little out of date - any genetic perturbations that causes a phenotype change in combination different from individual effects is a genetic interaction, including from small molecules (even if you don't know the genes involved). I asked a lot of other biologists here and they all agreed.  However, maybe using state variables will make gene and small molecule genetic perturbation participants fit together better conceptually.
 * Alan: Do you have a citation giving the definition you are referring to?

From Sven Nelander at MSKCC: I have only two comments, and both apply to the use of the Galitski nomenclature of gene-gene interactions. This way of categorizing gene-gene interactions may be quite problematic in a couple of cases:

1) Response surfaces

Looking at the work of for example Combinatorx, interaction is defined in terms of how the experimentally derived ‘response surface’ deviaties from a ‘zero response model’. In a typical yeast genetics setting, the response surface would be a 2x2 matrix containing the phenotypic value the gene A and gene B wild-type and mutant cases. It cannot be excluded that future experiments will emphasize perturbation gradients, also involving more than two genes. In these cases, the Galitski nomenclature for prototype interactions in yeast genetics, using very simple phenotypes might not apply. As a computational biologist, I would probably prefer to analyze a database that stores the actual surfaces, rather than interpretations/labels of prototype surfaces. I hope that the BioPax gene interaction format will be able to capture this. See: Keith CT, Borisy AA, Stockwell BR. Abstract 	Multicomponent therapeutics for networked systems. Nat Rev Drug Discov. 2005 Jan;4(1):71-8. Review. PMID: 15688074 [PubMed - indexed for MEDLINE]

2) Multivariate readouts

One situation where your classification of interactions may be problematic is when the phenotype is multivariate. For an extreme example, see the Van Driesche et al paper, where epistasis analysis is carried out using a TP readout. In these cases, the Galitski way of classifying interactions - using inequalities - will have to be replaced by some generalized framework. I haven’t really seen any multivariate classifications of gene-gene interaction though.
 * Gary: these are recorded in the future directions section

> Gary: One question - do you think it makes sense to store interactions involving chemicals in the same place we're storing genetic interactions involving genes?

Sven: I think it does make sense to store perturbation information at the same place whether it occurs at the gene level (knockdown/out, overexpr etc) and at the level of gene products (drugs), especially if the drug really can be said to target a specific gene product. I guess one could record the interaction and then classify it as gene-gene, gene-drug, drug-environment, gene-drug-environment etc?


 * Alan: Classifying them as such would be a welcome improvement, I think. But there is still some confusion in my mind about what the "interactors" should be, versus the "genetic-background" in, e.g. the Haggarty type experiments.
 * Alan: Just had a chat with Sven to get some more input. He is in agreement that the common use of the term genetic interaction is among gene participants. He also noted boundary cases, which arguably could go in this or another class (latter is my preference). Cases
 * From population genomics. If gene A is mutated 50% of the time in a tumor and gene B is mutated 50% of the time in the tumor, then one expects that 25% of the time both will be mutated. If this occurs at a higher or lower rate, then these can be called "interactions".
 * The kind of correlation derived connection between genes that have similar profiles in the Haggerty experiment, suggests interaction, but wouldn't be distinguised in kind from a correlation observed in microarray experiments. So if one is present, then the other would be too.

Comments from Ranjani Ramakrishnan: Genetic Interactions Proposal

1. Should the participants for the geneticInteraction class be restricted to gene and small Molecule. Couldn’t you perturb the system in other ways such as using siRNAs, ncRNAs etc? The mechanism by which they will perturb the system will still be consistent with the definition of the geneticInteraction class.

2. Regarding the existing property ORGANISM for the new gene class. Trying to capture the viral infection of a cell, the gene is clearly a part of the virus’ genetic makeup but is present in the host cell. How should this be represented? Same question for proteins that are coded for by the virus. Should they refer to the host or to the virus?

3. Property EXPERIMENTAL-FORM-TYPE: Would you consider adding the terms hypermorph and antimorph to the controlled vocabulary list?

4. In the class geneticEvidence, the property confidence allows for multiple values. Shouldn’t it just be one value? Even if you have multiple lines of evidence in support of the entity, it should be synthesized to one value.

5. In the geneticEvidence class, how will the data on scoring a phenotype to generate the phentoypeInequality relationship be captured? The Drees paper has an algorithm to do it but a user could potentially use some other criteria to generate the inequality. Also, if the method involves some computation, the parameters required and their values need to be captured.

General items:

4. You mention that BioPAX will be searching and supporting the creation of controlled vocabularies for phenotypes. Any guidelines on how often it will be done and if you will maintain lists of deprecated terms and the source of these terms etc.

Comments from Anastasija: Introduction Genetic interactions between genes occur when two genetic perturbations (e.g. gene mutations) have a combined phenotypic effect not caused by either perturbation alone. I think this definition might not be completely correct. In an epistatic interaction, infact, the phenotypic effect of a double perturbation is caused by one (but only one) of the perturbations involved (WT > A > B = AB or WT > A = AB > B). As in my presentation yesterday, I would suggest the following definition: A genetic interaction occurs when the phenotype of a simultaneous perturbation of 2 genes cannot be predicted from the phenotypes of the single perturbations alone. From Howard’s comments on my committee meeting I understood that the formula W(ab) = W(a) x W(b) + e is widely accepted to quantify genetic interactions. So for quantitative phenotypes the best definition might be: A genetic interaction occurs when e <> 0. But I don’t know whether it is possible to combine these two definitions in a general one. I was also thinking that information about “no interaction” should also be included in the ontology. In this way a combination of 2 or more mutations will have either of these annotations: not tested, tested (no interaction), tested (type of interaction). Other types of non-lethal genetic interactions are possible. Suppose absence of gene Z shows no phenotype, while absence of A causes a non-lethal fitness defect. Absence of both genes may either mitigate ('buffer') or exacerbate ('aggravate') the phenotype observed when only A is missing. If that is the case, A and Z are said to have a epistatic (genetic) interaction. You’re speaking about “epistasis” as a synonym of “genetic interaction”. It is risky because in genetics these are two different things. Basic requirements Very common genetic interaction types more specific than the ones above, such as synthetic-lethal and synthetic-sick. I don’t think we should mix phenotype, interaction type and interaction level. In case of SGA synthetic lethals or sick, the interaction type is “synthetic”, the phenotype is “colony size” and the interaction level can be defined qualitatively (full, partial) or quantitatively (colony sizes of the two single mutations and of the double - all compared to wt).

Phenotype	Interaction type	Interaction level SL (SGA)	colony size	synthetic	full SS (SGA)	colony size	synthetic	partial In the treatment of synthetic sick/lethals we have an additional problem: the SLs published by Charlie Boone and others do not correspond to the definition of SLs by Drees et al. Otherwise, we can think that “death” is a separate phenotype (like PSI-MI): Phenotype	Interaction type SL (SGA)	death	synthetic SS (SGA)	altered growth (colony size)	synthetic

These are the types of interaction that Charlie Boone’s lab considers SL or SS (in the last column the Drees class is indicated): AB	<	A	<	B	<	WT	additive AB	<	A	=	B	<	WT	additive AB	<	A	<	WT	=	B	conditional AB	<	A	<	WT	<	B	single-nonmonotonic AB	<	B	<	A	<	WT	additive AB	<	B	<	WT	=	A	conditional AB	<	B	<	WT	<	A	single-nonmonotonic AB	<	WT	=	A	=	B	synthetic AB	<	WT	=	A	<	B	conditional AB	<	WT	<	A	<	B	double-nonmonotonic AB	<	WT	<	A	=	B	double-nonmonotonic AB	<	WT	=	B	<	A	conditional AB	<	WT	<	B	<	A	double-nonmonotonic Basically the difference is due to the fact that Drees et al. name “synthetic” only interactions where A and B have no phenotype as single mutants. Boone’s lab, on the other hand, says that sinthetic sick is when AB is sicker than both A and B. The PSI-MI ontology is consistent with Drees et al.: Synthetic = Two silent mutations show an altered phenotype when they co-occur on the same cell. Proposed implementation The INTERACTION-TYPE property uses the PSI-MI interaction type controlled vocabulary (CV). The PSI-MI interaction type CV needs more genetic interaction types to support this proposal. Ensure that PSI-MI controlled vocabulary contains the following genetic interaction types (requires communication with the PSI-MI group, via their sourceforge feature request tracker). PSI-MI mixes interaction-type, phenotype, mutation type (knock-out, partial alteration, expression alteration), experimental conditions (temperature, nutrients). I think it would be better to keep these aspects separate because, e.g., one particular phenotype (growth rate) can be involved in many different interactions (potentially, in all of them). Intergenic non-complementation interaction : it involves comparison of diploid heterozygous phenotypes and maybe it could be considered next (more complex) step.

Property: GENETIC-BACKGROUND, range = String (optional), property definition: "The genetic background of the experiment, defined according to the standards of the field of study (e.g. R0013: MAT-alpha hsc82delta::LEU2 hsp82ts-URA3 can1delta::MFA1pr-HIS3 ura3delta0 leu2delta0 his3delta1 lys2delta0 met15delta0. Note: the e.g. C. elegans community may have different nomenclature than the S. cerevisiae community." I think that the nomenclature problem is becoming more and more important. Even in yeast the existing rules don’t cover all possible cases. In the genotype that you use as an example I see these problems:	MFA1pr-HIS3 = the promoter of the MFA1 gene has been placed upstream of HIS3 gene. There’s no standard way to indicate that.	Mat-alpha = Mat-α = Matα. delta = Δ. Isn’t it possible to avoid the use of greek symbols? This symbology is tradition, but for storage and exchange purposes isn’t it better to replace α with @ (for example) and Δ with something else? It might be a little bit out of the purposes of this proposal, but I think it is important to bring attention to the problem.

Add a new subclass of the existing utilityClass, called "phenotypeInequality", defined as follows Wouldn’t it be useful to include also quantitative measurements of the phenotype, when these are available?

Alan: For instance consider two compounds which interact with each other. If the compounds are non lethal alone, but a product of their reaction is toxic, would this be called a genetic interaction involving two compounds? I think it is an important point. Are we sure we want to mix these two things (gene-gene interactions and compound-compound interaction)? In my opinion, the interpretation of compound-compound interactions is much less straightforward than gene-gene interactions. It seems that compound-compound interactions can be compared to gene-gene interactions only when 1) compound 1 and compound 2 target two distinct and specific gene products; 2) they don’t react with each other and they don’t physically interact; 3) there’s no synergy in the import of the two compounds to the cell. If at least one of these conditions is not satisfied, two chemicals will still be interacting but their interaction won’t be comparable to a gene-gene interaction.