Much of the AI application ecosystem, from RAG to tool use to massive token context windows, is optimizing for a single objective: reducing response entropy by feeding the model more information. The more context you provide the more the distribution narrows.

This is a powerful system that produced the co-pilots that summarize papers in seconds and now, AI-powered platforms that run workflows and do bioinformatics analysis, and foundation models that excel at physics/chemistry tasks (e.g. Alphafold, Chai, Boltz). For problems where the “right” answer is derivable from available information, reducing entropy through context works beautifully.

Harvey, the AI for law, is one of the clearest examples of this. Many legal tasks reward fast convergence over a known body of rules, precedent, and language. It is hard to beat a system that can absorb a vast body of statutes, case law, and legal argument and retrieve it instantly. The rules are relatively legible, and much of the relevant context is available. This kind of problem rewards the ability to eliminate response entropy and converge quickly and precisely.

That is why so many knowledge industries are vulnerable to this type of disruption. And the biggest prize is pharmaceuticals. It is the Formula 1 of the knowledge economy. The implicit thesis was simple: instead of legal briefs, feed the system scientific literature, and it will learn biology.

The result has been a torrent of “AI Co-Scientist” and “Neurosymbolic AI” companies entering the most overhyped, overfunded, and epistemologically confused segment in vertical AI. And that confusion is a direct consequence of applying convergent tools to a fundamentally divergent problem.

There is a set of problems where entropy reduction breaks down entirely because it punishes convergence and rewards divergence. The highest-value decisions in drug discovery are exactly this kind of problem.

The decisions that matter, like which targets to pursue, patient populations to bet on, and which mechanisms to trust when the biology is ambiguous, are not made by converging on available information in literature. They are made by someone who has seen something others haven't, interpreted failure differently than the rest, or held a conviction that ran against consensus long enough to be proven right.

Alpha lives in the divergence.

And this is where the Ouroboros appears.

The Ouroboros of Biology.

Drug discovery relies on two principal factors: Asymmetric information and Differentiated Thinking. You win because you know something they don’t and you have smarter scientists who think outside the box. Because of these competitive dynamics, no one is incentivized to share their asymmetric knowledge, let alone publish it.

We train models on the published record of biology, then ask them to generate the next frontier of biological insight from that same record. That is the Ouroboros. A system trying to escape the limits of consensus by recursively consuming consensus. But the published record is not biology.

No one in the history of drug discovery made a drug because they read the literature better.

The scientific literature is a curated, systematically optimistic, self-referential record of what scientists were willing to report, what journals were willing to print, and what funding agencies were willing to support. Papers cite papers that cite papers that no one could reproduce. Positive results get published with a narrative. Negative results disappear into reports, decks, and private team memory.

The top scientific R&D teams know this already. Inside serious drug discovery organizations, this is almost banal. But for the outsiders building in this space, there is this tendency to over index on the value of literature as ground truth. Poor understanding of this reality is most obvious when vendors market their knowledge graph or foundation model to be trained on hundreds of millions of papers and abstracts (yes…even conference abstracts) as if sheer volume somehow makes the system better.

A More Structured Ouroboros.

“Fine. Literature alone is noisy. That’s why we use a knowledge graph to structure the claims. And we also integrate real-world and proprietary customer data into the graph.”

This is directionally correct but hides an insidious trap.

The value of a knowledge graph is the ontology. The ontology is not neutral, it is judgment. Every definition in the schema encodes a view of the world. What counts as the same thing? What makes things different? What gets linked? What gets collapsed? What do words mean?

An ontology is a scientific world view made operational.

The most important question is: whose graph is it?

If it is not your ontology, it is not your knowledge graph.
If it is not your knowledge graph, it is not your science.
If it is not your science, it is not your decision.

The moment a vendor says they “integrate customer data into their graph”, you should ask what scientific judgment you are implicitly outsourcing because you’ve just relinquished control. They are deciding how your internal assays map to external biology. How your translational signals get normalized. How your negative data gets represented. What counts as corroboration. What counts as contradiction. What counts as enough evidence to connect one claim to another.

You cannot have this both ways. You cannot claim that your proprietary ontology-driven evidence graph is valuable and ingesting customer data into that proprietary graph. You are either infrastructure that allows the customer to instantiate, govern, and evolve their own scientific worldview. Or you are a scientific intermediary that aggregates, reshapes, and ultimately substitutes for that worldview.

A simple test is to ask the vendor to change the underlying ontology to accommodate your internal definitions. See what happens.

The Contradiction at the Center

This is why so many of these platforms feel impressive in demo mode and suspicious in principle. They can absolutely help teams search faster, summarize faster, traverse claims faster, and maybe even operate workflows faster. But as soon as you diverge from consensus, everything starts to break.

A pharma company does not win because it has access to information. Everyone has access to information. It wins because it developed a way of seeing that information differently, grounded in data others do not have and judgment others cannot replicate. This is the reasoning layer that cannot be outsourced without consequence. It is where scientists encode scientific reasoning and capture the differentiated thinking a team uses to decide what to believe, what to ignore, what to test, and where to place conviction under uncertainty.

How do we get out of this cycle?

1. Reset on expectations. You are NOT going to get a Harvey for drug discovery. It’s a different problem class (convergent vs. divergent).

2. Stop this. These types of posts are so inflammatory I need a shot of Dupixent.

Final Thoughts.

Drug discovery does not reward systems that converge fastest on the published record. It rewards teams that know something others do not, interpret evidence differently, and maintain conviction under uncertainty.

That means the future of AI in biology will not belong to whoever reads the most papers or structures the most claims. It will belong to whoever helps scientific organizations encode their own ontology, their own context, and their own judgment without surrendering them to a vendor-controlled worldview.

Otherwise, the snake keeps eating its tail.

BioBox is the Decision Operating System for AI-driven pharma.

If you’ve worked in drug discovery, or just biological sciences in general, you’ve encountered the frustration of spending a ridiculous amount of time pulling data together to answer seemingly straightforward questions like “What are the druggable targets in pathway X that are overexpressed in tumor type Y?”.

• Opening six different databases

• Downloading CSV files from PubMed, TCGA, ChEMBL, and your internal systems

• Writing brittle Python scripts to join these datasets

• Discovering that gene names don’t match across sources (HUGO vs. Entrez vs. Ensembl)

• Spending three days on data wrangling before you can even start analysis

• Producing an answer that’s already stale because you can’t easily update it

There are those who believe that agentic systems or “AI scientists” will eliminate these problems, they don’t (which is a topic for another blog post, maybe).

The reality is that this isn’t a tooling problem. It’s a representation problem. We’ve been trying to force inherently graph-structured data into tabular formats for decades and it’s killing productivity and forcing us to build things that don’t need to exist if you just modeled data properly.

The argument for this post is simple: Biology is a graph. Any other representation is a lossy approximation that handicaps our ability to do science at scale.

Why Biology is Fundamentally Graph Structured

Biological Meaning Exists in Relationships

An isolated biological entity has almost no meaning. The information content is encoded in how the parts interact, regulate, and influence each other. Consider TP53, perhaps the most studied gene in all of cancer biology. In any given database you’ll see an entry that looks like this:

Does this tell you anything useful about p53? Does it explain why it’s the guardian of the genome? Does it hint at why 1000s of papers have been published on it?

No. This representation strips away everything that makes p53 biologically meaningful.

Biological meaning is an emergent property of the network.

As a disease driver, p53 is mutated in more than half of all human cancers. But it’s not uniform and the spectra differs dramatically depending on the cancer type. For example, in Lung AD, mutations are mostly missense and cluster within DBD create stable yet dysfunctional proteins. In ovarian cancers, p53 mutations are mixtures of missense and truncating variants that eliminate protein function altogether. These differences have important biological and therapeutic implications. A single mutational frequency column cannot capture this type of nuance.

As a transcription regulator, p53 controls expression of 100s of downstream genes. When DNA damage is detected, TP53 upregulates CDKN1A (p21) to arrest the cell cycle, giving time for repair. It activates BBC3 (PUMA) and BAX to trigger apoptosis if damage is irreparable. It regulates RRM2B and SCO2 to modulate metabolism. It induces SESN1 and SESN2 to manage oxidative stress. The circuits that p53 regulates is not a singular output. It’s a context-dependent network of regulatory relationships that changes based on cell type, stress type, and cellular state.

As a protein, p53 exists in a sprawling web of physical interactions that determine its activity. In unstressed cells, MDM2 targets it for degradation to keep levels low. MDM4 reinforces this repression. Acetylation by p300/CBP enhances transcriptional activity by stabilizing it. Deacetylation by SIRT1 reverses activation. The point is that it’s not a simple on/off switch. These are complex dynamically regulated networks of protein-protein interactions, post-translational modifications, and feedback loops that are, again, context-dependent.

As a therapeutic target, TP53 defines treatment strategies through multiple relationship types. In tumors with wild-type p53, MDM2 inhibitors can reactivate the p53 pathway. In tumors with mutant p53, the mutation creates new vulnerabilities where loss of TP53 function produces synthetic lethality with WEE1 inhibitors, ATR inhibitors, and PARP inhibitors in certain genetic contexts. Mutant p53 can also gain oncogenic functions through interactions with other transcription factors. Understanding how to treat a TP53-altered cancer requires traversing from the mutation through the pathway network to potential therapeutic interventions.

Tables joins become nuclear level explosions.

Try to place them in tables in a relational db and they might look as follows:

• genes table with basic gene info

• gene_disease table linking disease associations

• gen_expression table with transcriptional expression data across tissues/cells/clinical popiulations

• ppi table

• regulatory_relationships table with transcription factor targets

• pathways table with pathway membership

• mutations table with variant data

• drug_targets table with therapeutic relationships

This seems reasonable until you try to answer actual biological questions.

“What are the druggable proteins that interact with genes in the p53 pathway that are overexpressed in lung adenocarcinoma?”

In SQL, this requires:

1. Query the pathways table to get genes in the p53 pathway

2. Join to the gene_expression table filtered for lung adenocarcinoma

3. Join to the protein_interactions table to get interacting proteins

4. Join to the drug_targets table to filter for druggable proteins

5. Potentially join back to the genes table to get protein-coding gene information

That’s 4-5 joins across different tables. Now add in the complications:

• Gene identifiers don’t match across tables (HUGO symbols vs. Entrez IDs vs. Ensembl IDs)

• You need to filter by expression thresholds, but what threshold?

• “Overexpressed” needs to be compared against normal tissue, that’s another join, also what threshold?

• “Druggable” has multiple definitions (ligandable pocket, known ligands, approved drugs)—more branching logic

• Pathway membership is ambiguous (is indirect regulation included? what about post-translational regulation?)

Each of these complications adds more joins, more logic, more query complexity. Change one criterion and you need to rewrite significant portions.

Context-Dependent Relationships Cannot Be Flattened

This is the most fundamental problem with tabular representations. Biological relationships are entirely context-dependent, and context doesn’t fit in columns neatly.

If you want to model Gene X regulates Gene Y, this isn’t binary, it depends on:

• Cell type

• Developmental stage

• Disease state

• Environmental conditions

• Genetic background

• Measurement

• Experimental conditions

In a table, you have three options to choose from, all of them are bad.

1. Ignore the context. Store just “X regulates Y” and lose accuracy.

2. Add context columns to the table. But how many columns do you add? How often does this schema change? How do you query across contexts?

3. Create separate records for each context. Massive duplication.

In a graph, this can be natively modeled and captured by traversing graph nodes. It’s also easy to chain multiple patterns together to get more and more complex relationships as network topologies.

Biological Identity Depends on Network Position

What a specific entity “is” depends on where it sits in the graph (or subgraph).

For example, is PKC an oncogene or a tumor suppressor? Well, it depends. Which tissue? Which pathway context? PKC promotes cell survival in some contexts and apoptosis in others. To understand the functional role, you must examine its network position.

Is CDK4/6 inhibition growth-suppressive or growth-promoting? In ER+ breast cancer with intact Rb, it’s suppressive. That’s why palbociclib works. In Rb-null cancers, blocking CDK4/6 can paradoxically enhance growth by removing cell cycle checkpoints. The therapeutic effect depends on the network context.

This is a fundamental principle: biological function is not an intrinsic property of an entity, it’s an emergent property of that entity’s position in a network of relationships.

Tables force you to assign properties to entities: “CDK4.function = cell cycle progression.” Graphs let you represent the truth: “CDK4 function emerges from its relationships with cyclins, CDK inhibitors, Rb, E2F, and downstream targets, which differ by cell type and genetic context.”

You need a graph.

So what? Why should you care?

Productivity cost: Scientists spend 30-40% of their time wrangling data instead of analyzing it. Most of this time is spent joining, aligning, and integrating datasets i.e., reconstructing the relationship network that should have been preserved in the first place. No, an AI agent doesn’t help you do this either.

Correctness cost: Flattening relationships loses context, which leads to incorrect conclusions. How many drug targets have been pursued based on gene expression data that ignored tissue context? How many pathways are misunderstood because protein isoforms were conflated?

Innovation cost: Complex questions that require multi-hop reasoning are simply not asked because they’re too hard to answer in SQL. Scientists constrain their hypotheses to fit the limitations of their database, not the boundaries of biological possibility.

This is why every major drug discovery organization needs a knowledge graph. Not because graphs are trendy or cool, but because biology is a graph, and representing it otherwise is scientifically incorrect and practically limiting.

Part 1 - Patient Recruitment using Graphs

Introduction

Indication expansion aims to increase the value of your drug by finding additional diseases that can benefit from it. Typically the drug is already in clinic or in market. More indications = more money. This analysis tells you “what” diseases makes sense based on biological principles. Another type of analysis known as indication sequencing describes the “how” and “when” to expand strategically.

All therapeutics, at a very fundamental level, basically translates to “give a substance to a biological thing to modulate a biological event that, when it is abnormal, is part of some disease”.

The key intuition here is to think of the association between a drug and the indication as a function of the biological effects it can modulate within a biological context. For example, you take an Advil when you have a migraine but Advil treats inflammation, which happens to be part of the symptoms of a migraine.

What exactly is a biological event?

A biological event refers to any measurable change within a living system caused by a specific interaction, such as the activation or inhibition of a biochemical pathway, receptor, enzyme, or gene.

How the substance interacts with the target to exert the correct change to a biological event is what we refer to as the mechanism of action (MOA). Biological complexity gives us an opportunity to exploit fundamental processes in different ways through a variety of targets including genes, proteins, protein-protein interactions, specific nucleotides, epigenetic modifications, and so on.

The participants in these effects vary widely, including proteins, small molecules, macro-complexes, and entire tissues. Biological effects also depend on context: the same molecule may behave differently in distinct tissues or under varying conditions, such as stress or infection.

The use of graphs in indication expansion is well-known. To understand why graphs are so powerful for this type of work, we will trace Dupilumab, a biologic that was initially approved for atopic dermatitis (AD) but has since gained approvals across multiple chronic inflammatory diseases such as COPD and asthma.

Graphing Biology

Dupilumab is a monoclonal antibody that acts as a receptor antagonist by binding to the alpha-subunit of interleukin-4 receptor (IL4R).

I’ve actually seen knowledge graphs in the wild that look like this.

This is useless. It gives you no analytical advantage, regardless of how “expertly curated” it is. The graph topology does not enable you to make deductions or inferences. Most of the value in your graph is in the design of the ontology.

Here’s what you should do instead.

Use a concept called Event that captures useful information about biological reactions such as what the input substrate, output and the context of the event.

Next, describe the specific event that the drug is affecting. Dupilumab binds to IL4R and occludes the fibronectin domain, preventing IL4 from binding.

Next, we introduce the Process concept to assign the logical order to a collection of events. Zooming out and expanding the diagram above:

Dupilumab works because it attenuates Th2 pathway activation by crippling STAT6 activation. We can map the impact of the inhibition through evaluating the transcriptional targets of STAT6. To find these targets, you can rely on literature mining or determine these target empirically using epigenetic data sources such as ENCODE. Among these targets include CCL13, IGHE, IGHG1, and IGHG4 that drive type 2 inflammation. Importantly, STAT6 activation leads to IL4R upregulation and increase IL4 secretion, and executes a positive feedback loop to further amplify the response.

Causal Biology Modelling & Inference

Dupilumab is a blockbuster drug with 6 approved indications and more in the pipeline. This “pipeline in a product” type drug is extremely profitable because the drug targets a specific mechanism that is shared across a collection of diseases, in this case, STAT6 activation in type II immune response related diseases.

What happens when STAT6 is activated? What phenotypes are expected? What diseases are relevant?

We can leverage graph algorithms to answer these questions.

First, let’s build and visualize the network.

• Starting from STAT6, mark all its transcriptional targets using ChIP-seq evidence. Specifically, you are looking for co-localization of ChIP-seq peaks for STAT6, H3K4me3, and H3K27Ac.

• For any STAT6 targets, that are known to have DNA binding domains or are known transcription factors, expand out the transcriptional targets.

• Repeat the process 2-3 times outwards

• For all the genes marked - load in disease associations

• Apply circular drawing algorithm for visualization

The positions of nodes in this projection are not arbitrary - they are determined by the topology of the network. Nodes that are more closely related (for example, those that share many edges or are part of the same community) tend to be placed next to one another. You are essentially seeing a one‐dimensional (angular) projection of the network’s connectivity. Nodes placed next to each other are likely to interact or be functionally related, while nodes that are far apart on the circle are less directly connected

In this application, we are doing a sanity check to see if the projection aligns with what we’d expect to see. For example, we know that Dupilumab is indicated for atopic eczema, so we should expect to find it somewhere here in this network. Indeed, we find it placed alongside CLDN1 and a variety of skin related disorders.

So far so good.

An interesting cluster with high convergence around STAT3 and a collection of chronic inflammatory diseases such as ulcerative colitis, psoriasis, and IBD.

The connection become clear when we run a path-finding algorithm between STAT3 and STAT6. Turns out, STAT6 drives the expression of IL4R and IL4 secretion. IL4R in heterodimer with IL13RA1 or IL2RG are both capable of phosphorylating STAT3, representing an alternate mechanism outside of the canonical IL-6 family activation loops. I’d suspect that Regeneron will starting going after treatment refractory UC and IBD soon.

Assessing efficacy for strategic expansion

“Pipeline in a product” are the holy-grail products that every pharma is chasing. When staging the launch sequence, one of the successful ways to build momentum is to use a narrow-first approach. Among the different variables for optimizing this launch strategy, one of them is to ensure you line up indications that you know are going to demonstrate strong efficacy and high market differentiation from standard of care. Translated into systems biology, we can interpret this as saying, does our drug impact the causal biological process the most.

How do you quantify that? Walk through the graph and calculate the path weights. (more on this in part 3 of this series)

To illustrate this, let’s use the recent example of BMS’s Cendakimab dropping out against Dupilumab. This was a wise decision because Cendakimab would have 100% lost this fight.

Cendakimab is a IL13 inhibitor. The goal is still to decrease IL4/IL13 signaling, which is the same as saying STAT6 suppression. There critical event is STAT6 phosphorylation, which happens under IL4R-alpha dependent heterodimerization. We can kick this cascade off at these entry events:

• IL4 binds to IL4R → recruits IL13RA1

• IL13 binds to IL4R → recruits IL13RA1

• IL13 binds to IL13RA1 → recruits IL4R

At first glance, it might look like an IL13 blockade alone could achieve meaningful, or at least, comparable results because it serves as 2 potential entry points. This argument falls apart when we include the biological context in which these events are taking place. Specifically, IL13 binds to IL13RA2 (a decoy receptor) with much higher affinity than IL13RA1. Also, IL4 is still able to exert its effect, independent of IL13 blockade.

At best, Cendakimab might edge out a win on dosing regiment efficiency and a narrower ADR profile, but it’s not enough to drive differentiation and improvement over the current standard of care.

Conclusion

In this post, we dived deeper into how network graph analytics can help inform critical decisions made along the drug discovery landscape. Building a model to capture the complexity of systems biology can only be done in a structured ontology and knowledge graph. The investment in this resource unlocks powerful capabilities that research teams can use to answer questions in indication selection and strategic expansion in a principled way.

What’s Next?

Upcoming in this series, we’ll switch gears into building scoring models and graph neural networks that can help take safety risks off the table when evaluating indications.

BioBox is the knowledge infrastructure for modern biopharma research, built for drug hunters who need to integrate multi-modal data, engineer knowledge, and test hypotheses at scale. To learn more, please visit our website at https://biobox.io or click here to contact us.

Introduction

Indication selection is the process of selecting and ranking potential diseases or conditions (indications) that a drug candidate could target, based on scientific, medical, regulatory, and commercial factors. This strategic decision requires participation from leadership across multiple functions in the company.

By the time we engage with biopharma, they already have a general idea of what therapeutic areas (e.g. CNS, liver disease, auto-immune, etc.) to focus on. Usually, this is set in the company’s mission, a specific area of expertise, and/or something their platform is uniquely positioned to solve for. The asset’s stage of development is also important to consider. The starting points for indication selection strategy are different based on the type of biotech company. Because of these nuances, it was not surprising to find that each customer had a slightly different perspective on how diseases should be ranked. As a result, we’ve ingested dozens of different data sources and built many knowledge graph variants for customers.

In this blog post, we’ll share a blueprint of best practices for modeling data inside a knowledge graph to help biotech teams evaluate their competitive positioning based on feedback from our customers.

Organizing Principles

There are 2 entry points for this discussion: Diseases and Processes. This distinction is important when describing treatment modalities. For example, in immuno-oncology, one of the desired outcomes is to activate T-cell response. These therapies may have broad application across a portfolio of diseases. It is also useful to connect Phenotypes with diseases to improve resolution.

Commercial Attractiveness

Working with business and medical affairs teams broadened my perspective on how much multi-modal data goes into a indication decision. But it basically boils down to this equation:

The goal is to max out this ratio. Let’s talk about some strategies that we’ve seen work.

Patient Recruitment

Your ability to recruit enough patients to satisfy the study criterion is a vital consideration. High competition for patients will make your trial run longer, delaying the time to value. Overextending the trial duration means more staff on payroll. The overall effect is the denominator increases, by a lot, and hurts your commercial attractiveness.

Several factors contribute to patient recruitment potential:

1. Pick an indication that has high prevalence = larger patient supply

2. Large unmet need

3. Competing trials recruiting for the same patient populations

Everyone knows about 1 and 2. There’s not much that can be done there to give you an advantage. But it turns out we can use graphs to avoid setting up competing trials. Every clinical trial will list their inclusion and exclusion criteria for participants. Customers we’ve worked with extracted the unstructured data into a structured set that we then model into the graph.

After repeating this for the exclusion criterion, we can build a graph representation model (mono-partite projection) that calculates similarities of Clinical Trial nodes to help scaffold our own inclusion criterion that will deliver on our scientific/regulatory goals, but minimizes the similarity to other actively recruiting trials.

At decision-time, the process basically works as follows:

1. Enumerate your ideal inclusion/exclusion criteria that will maximize your chance of the trial passing (we’ll cover this topic in part 2 of the series)

2. Run the similarity calculations

3. Work with medical affairs and ClinOps team to drop/augment a criteria

4. Re-run the similarity calculations

5. Repeat until either ClinOps/medical affairs denies another change or similarity scores reaches a global minima

Bonus: Geo-spatial data to minimize site overlap

This next aspect of the graph was not something we expected to be done in our discipline, but yielded information that was remarkably strategic, even outside of drug development (more on this later). The problem was that even when you have the right indication, the right inclusion/exclusion criteria, you still have to physically recruit the patients and bring them to the sites. While the global disease burden is high, if the local recruitment location is highly saturated with similar trials, you’ll still face a recruitment challenge.

Each clinical trial that is active and/or recruiting is required to disclose their study sites. The API response for study locations conveniently includes the geospatial coordinates (longitude, latitude). This adds a new dimension to use in our decision making. We can evaluate how far apart the trials are from a clinical characteristic perspective and we can calculate the physical distance between the trial locations.

Use-cases outside of drug discovery

I recently reconnected with an old colleague who is now in a leadership role at a major Canadian hospital. I shared with him the geo-spatial strategy which brought up a use-case I had not previously considered. Attracting more clinical trials to your research site.

Research hospitals benefit greatly from conducting clinical trials, it’s a significant source of revenue. However, there is an activation time required at the research site to prepare the clinical research operations to execute successfully. It is a very reactive type of work where lucrative opportunities are passed because the timing didn’t align. But if you knew what the trends are for the upcoming trials that are going to start recruiting and have a solid mapping of their clinical characteristics, you can preemptively activate your research operations so that when recruitment opens, your site is ready to go. Furthermore, you can use the spatial data to reverse engineer which trials are competing for the same populations within the same locations and prep/offer your trial site to satisfy their recruitment needs and drive revenue into your institution.

What’s Next?

Upcoming in this series, we’ll continue to focus on indication selection patterns. Specifically, we’ll be sharing best practices to increase the "probability of success" variable in the equation such as mechanism of action analysis and biomarker mapping.

BioBox is the knowledge infrastructure for modern biopharma research, built for drug hunters who need to integrate multi-modal data, engineer knowledge, and test hypotheses at scale. To learn more, please visit our website at https://biobox.io or reach out to one of the team members.

Introduction

Biological states and transitions between these states are controlled and regulated processes. There are numerous chemicals, molecules, complexes that work to coordinate these processes, that together, give rise to form and function. Our task, as scientists and drug hunters, is to understand how these processes work so that we can exploit them and find cures for diseases. To understand causality, we need reliable and up-to-date information from sources that we can trust. For biological pathways and systems biology, the de facto choice for BioBox and our biopharma partners has been the Reactome knowledgebase.

Reactome Knowledgebase

Reactome is a publicly funded, open access, open source database, created in 2002 by Lincoln Stein at CSHL, Ewan Birney at EBI-EMBL, and Suzannah Lewis at Lawrence Berkeley National Laboratory. Reactome is now the leading open resource for human-curated pathway knowledge, and is a key resource in BioBox, used for making reliable predictions.

The Bedrock of Reliability: Reactome’s Curation Practices

In the dynamic realm of bioinformatics and biomedical research, reliable, high-quality data is the lifeblood of scientific advancement. Researchers depend on databases like Reactome to provide accurate and comprehensive information about biological pathways. Central to Reactome's reliability is its meticulous curation process. 

Reactome’s curation process is rigorous and multi-faceted, ensuring that the data it provides is both accurate and relevant:

1. Literature Review and Data Extraction: Curators at Reactome are PhD-level biological experts, who conduct an extensive review of scientific literature to extract knowledge about biological pathways. This step demands a deep understanding of molecular biology and the ability to discern significant findings that can be integrated into pathway models.

2. Data Integration and Standardization: Extracted data is integrated into Reactome according to Reactome’s object-based data model, and standardized using controlled vocabularies and ontologies such as the Gene Ontology (GO). This standardization ensures consistency, making it easier for researchers to query and analyze the information.

3. Expert Review and Validation: Reactome curators work with domain experts to review and validate the curated data, similar to how academic journals conduct peer-review before publication. These experts ensure that the pathways accurately reflect current scientific knowledge and experimental evidence.

4. Community Involvement: Reactome encourages the scientific community to contribute data, suggest corrections, and provide feedback. This collaborative approach helps keep the database current with the latest scientific discoveries. Here is a web page, showing pathways that are ready for external review[1].

Transforming Data with Knowledge Graphs

Knowledge graphs have become a transformative tool in the bioinformatics landscape. By structuring data as interconnected entities and relationships, knowledge graphs provide a more intuitive and powerful way to represent and analyze biological information. Reactome’s adoption of a knowledge graph perspective enhances its curation practices and opens up new avenues for research.

By organizing data within Neo4j, Reactome not only enhances data reliability but also unlocks new opportunities for researchers to explore and interpret complex biological systems. See here for a publication by Fabregat et al, 2018, discussing the Reactome graph database [2].

Data Interconnectivity:

• Entity Relationships: In a knowledge graph, each piece of data is an entity connected by relationships. This interconnected structure allows researchers to see not just isolated data points but the rich web of interactions and dependencies between them. For instance, a knowledge graph can reveal how a particular gene is involved in multiple pathways and how these pathways inter-relate.

• Contextual Understanding: Knowledge graphs provide context by linking related entities. Researchers can understand how a particular biological pathway fits into larger cellular processes and how alterations in one pathway might affect others, leading to more holistic, systems-level insights.
  

  

Enhanced Data Analysis:

• Advanced Querying: Knowledge graphs enable sophisticated querying capabilities. Researchers can formulate complex queries that span multiple entities and relationships, uncovering insights that would be difficult to achieve with traditional database queries.

• Predictive Analytics: By leveraging the interconnected nature of knowledge graphs, AI and machine learning algorithms can predict new relationships and interactions within the data. This predictive capability can lead to the discovery of novel pathways and biological mechanisms.

To try this out, Reactome has provided documentation here that describes how to install and try out Reactome’s Neo4j graph database [3].

AI and Machine Learning:

• AI-Augmented Data Curation: AI and machine learning algorithms can automate aspects of the curation process. For example, natural language processing (NLP) can extract relevant information from scientific literature and integrate it into the knowledge graph, speeding up the data curation process. The Reactome team is actively working toward finding reliable methods to augment curation processes using AI. Stay tuned for more information in the future!

• Pattern Recognition: Machine learning can identify patterns and correlations within the knowledge graph that might not be immediately apparent. Researchers can leverage Reactome’s Neo4j graph database to find new insights, driving new hypotheses and research directions.

Meeting Researchers’ Needs: Reliability, Accessibility, and Discovery

For researchers, the reliability and accessibility of data are critical. By employing knowledge graphs, Reactome enhances these aspects and offers additional benefits:

1. Comprehensive and Accurate Data: The rigorous curation practices ensure that data within Reactome is accurate and comprehensive. The knowledge graph structure further ensures that data is contextually relevant and interconnected, providing a richer understanding of biological pathways.

2. User-Friendly Interface: The knowledge graph framework supports advanced visualization and interaction tools. Researchers can easily navigate through interconnected data, explore relationships, and visualize complex pathways in an intuitive manner.

3. Up-to-Date Information: The dynamic nature of knowledge graphs allows for continuous updates and refinements.

4. Facilitating Discovery: Knowledge graphs not only provide current data but also facilitate the discovery of new knowledge. Researchers can explore uncharted relationships and generate new insights, driving innovation and scientific progress.

Using Reactome Data in BioBox

Loading Reactome data into your BioBox knowledge graph can be done through importing the Reactome data package in the external data package listing.

This data package includes concept and relationship definitions that will incorporate into your custom ontology. Objects loaded through this data package preserves the data and mapping from the Reactome knowledge base and integrates with your private knowledge graph for increased comprehension.

Conclusion

At BioBox, we are thrilled to be partnered with Reactome. Reactome’s commitment to rigorous curation practices, coupled with the power of knowledge graphs, makes it an indispensable resource for researchers. By organizing data in an interconnected, contextual framework, Reactome enhances data reliability and opens up new opportunities for exploration and discovery. As the integration of data analysis and AI technologies continues to advance, Reactome remains at the forefront, providing researchers with the tools they need to push the boundaries of scientific knowledge.

References

[1] Link to Reactome Collaborator Zone webpage: https://reactome.org/community/collaboration

[2] Fabregat A, Korninger F, Viteri G, Sidiropoulos K, Marin-Garcia P, Ping P, Wu G, Stein L, D'Eustachio P, Hermjakob H. Reactome graph database: Efficient access to complex pathway data. PLoS Comput Biol.Jan 29. PubMed.

[2] Link to Reactome Developer Zone documentation: https://reactome.org/dev/graph-database

About BioBox

BioBox is a knowledge infrastructure for modern biopharma research teams to accelerate drug discovery and make better decisions in preclinical studies. Biopharma research teams from startups to top 20 pharms use BioBox to transform multi-omic data into knowledge graphs and use them to drive decision making in target prioritization, indication selection, and biomarker discovery.

To learn more visit biobox.io or reach out to get in touch with a team member.

Knowledge graphs (KGs) are powerful tools for reasoning and analysis. There is a misconception that KGs only apply to text-based information. The most common objection we hear is that KGs are unsuitable for data science. However, with a few changes, thoughtful data modeling, and novel software, we can unlock the power of these KGs and explore information in data-driven and novel ways.

In this post, I’ll share some frameworks we use at BioBox to craft custom multi-modal knowledge graphs built upon datasets spanning various sequencing technologies. Then, we’ll walk through how this graph is used in the research process to answer questions.

Goals of data integration

Using different experimental techniques and combining data logically can deepen our understanding of biological phenomena. For example, you can use ChIP and ATAC-seq to explore the regulatory control that explains the up/down regulation observed in matched RNA-seq data. Interpreting these datasets requires information about genes, molecular functions, biological processes, etc. Each dataset provides thousands of data points, and the search space is enormous. The goal of data integration is to improve our ability to separate signals from noise.

Data schemas

Metadata capture

The best practice is to isolate all the important variables that can be used for logical flows as nodes in the graph rather than storing them as properties in the node.

Why separate disease and cell type?

It’s simple: we care about the relationships that diseases and cell types have.

Specifically, we can exploit the ability to walk from semantically related cell types and diseases. By leveraging the power of connected graphs, we can expand our search space for questions like, “Find all glioma samples,” and walk down the paths. Consider what would happen if we stored these values as properties inside the Sample or Donor nodes. We would need to exhaustively search all samples for some key-value match for every value along the semantic chain of values (glioma, astrocytic tumor, astrocytoma), which becomes prohibitively inefficient.

RNAseq

RNA-seq is relatively straightforward to model. These experiments produce two types of datasets: Gene Expression and Differential Expression. You want to capture some information about the transcriptional state of a gene in a biological context (e.g., disease vs. normal, cell population A vs. cell population B, etc.). The gene expression graph mapping is relatively straightforward once the metadata is appropriately set up. We can describe the expression of a gene within a particular RNA-seq library with a direct edge as follows:

Differential gene expression (DGE) datasets require an additional hyper-edge to capture the comparison information.

Because we use a property graph as the base, we can persist the quantitative values of the log2 fold change, p-value statistics, count values, etc., directly inside the edge. It also becomes useful to incorporate these values as edge weights in other graph-based algorithms (more on that in the future). This is a unique characteristic of property graphs and data graphs that differs from traditional RDF-based knowledge graphs.

ChIPseq

Tracking epigenomic observations requires the graph to comprehend genomic coordinates. To do this, we start by creating 1 kilobase (kb) bins across the entire genome and represent them as beads on a chain.

Then, we can load ChIP-seq observations, such as enrichment peaks, into our graph. Each peak is represented as a distinct node with edges to genomic range nodes that demarcate the genomic region from which the peak was detected.

In our graph, we have also loaded regulatory gene features, such as promoters and enhancers. By combining this information, we can compose more complex and meaningful patterns. For example, in the figure above, the topology describes an EP300 peak bound detected to overlap with gene ABC123’s promoter region, suggesting possible epigenetic control.

Doing useful things with multi-omic graphs.

The usefulness of data comes down to whether or not it can impact decision-making. Ultimately, we are hunting for biological explainability so that we can attribute the importance of a target to the overall etiology of a disease. This forms the basis for a drug program to spend millions of dollars around this therapeutic hypothesis. Combining data modalities enhances the ability of scientists to see the bigger picture and make better decisions.

Extending the graph to answer questions

There are only two requirements for graph model construction:

1. There are two concepts defined (Bipartite graph).

2. There are relationships in your ontology that can directly or indirectly connect these concepts together.

Types of Bipartite Graphs built on the BioBox Platform:

• Gene <> Disease

• Gene <> Gene Module

• Drug <> Adverse Event

• Protein <> Phenotype

Example: Transcriptional Circuits & Pleiotropy

From transcriptomics, we can generate differentially expressed genes and run the count data through GSEA to yield some enriched pathways and gene sets. Typically, we’re hunting for pathways that contribute to an observed phenotype of our disease area. Chances are you will have an overwhelming amount of DE genes and pathway hits. How can we systematically qualify and eliminate targets in a principled way?

One approach is to layer on epigenetic data to identify targets that have the highest impact across multiple dysregulated pathways.

1. Collect the list of genes that were up-regulated and found in enriched gene modules of interest.

2. Evaluate the context-specific binding of transcriptional factors (TF) at the gene list promoter regions.

3. Evaluate the context-specific accessibility of promoter regions for gene list.

4. Rank the TFs based on occurrence

Together, we are filtering for the strongest transcription factors driving the up-regulation of genes involved in disease pathways/mechanisms. It turns out when the data is modeled and connected in the graph, the context-heavy question is trivial to search for.

In practice, this means we are building a graph model to compute the score between two concepts: Disease and Gene. Each layer of data integrated is an additional piece of information that builds up a scientific hypothesis. Each line of evidence can be used to validate or invalidate a hypothesis to varying degrees. In other words, these lines of evidence can be positively or negatively weighted values. The direction depends on the research question and the hypothesis.

For example, in a target prioritization exercise:

• + 0.3, Gene is up-regulated in disease vs. normal

• + 0.1, Gene participates in disease-specific pathways of interest

• + 0.2, Gene is a TF

• + 0.4, Gene product binds to promoters of disease vs. normal up-regulated genes

• - 0.3, Gene is down-regulated in disease vs. normal

• - 0.1, Gene product binds to promoters that are not accessible

The bullets above are reachable graph paths in the knowledge graph as defined by the ontology. An instance of this graph path is an idempotent and unique trace. For example, given 10 differential expression datasets, there can be (up to) 10 paths in the graph connecting a disease to a gene.

This is an important difference from traditional knowledge graphs.

On repeated observations, you should probably choose a scaling factor so it doesn’t dominate the overall score. By default, we use harmonic sums, but you’re free to choose whichever method you desire. Summing across positive and negative weighted paths will give you an overall aggregated score to rank objects from these two concepts together.

Conclusion

Crafting a knowledge graph isn’t just about connecting data; it’s about empowering researchers to uncover new insights, make informed decisions, and drive innovation in ways that were previously unimaginable. By integrating diverse data modalities—whether it's transcriptomics, epigenomics, or beyond—we can construct a rich, interconnected tapestry of information that shines a light on the complex biological mechanisms at play.

At BioBox, we've seen firsthand how a well-constructed knowledge graph can transform the research process, turning what was once a daunting sea of data into a navigable map that leads directly to actionable discoveries. Whether you’re prioritizing drug targets, exploring disease mechanisms, or seeking new therapeutic hypotheses, the power of a multi-modal knowledge graph can be the difference between a promising lead and a breakthrough.

But remember, the true value of a knowledge graph lies not in its complexity but in its utility. It’s not just about storing data in a visually appealing structure—it’s about using that structure to generate meaningful results that can drive your research forward.

So, as you consider building your own knowledge graph, ask yourself: How will it serve your team? What questions will it help answer? With the right design and a clear purpose, your knowledge graph could be the key to unlocking the next big discovery in your field.

If you’re ready to harness the power of knowledge graphs in your research, visit us at biobox.io to learn more and connect with our team of experts. Let’s explore the future of data-driven science together.

Every day new AI and data management tools are hitting the market, promising to revolutionize drug discovery. The majority claim to accelerate insights and facilitate data driven decisions, but it’s often difficult to understand exactly how they are helping teams accomplish this.

In this article I will walk you through exactly how a drug discovery team used the BioBox platform to solve complex data challenges.

Within 8 weeks the data team was able to:

• Build a proprietary evidence ranking and prioritization system that led to the identification of 10 novel targets.

• Prioritize 2 new indications for existing assets.

• Curate a custom representation of disease biology.

• Leverage machine learning algorithms to identify serendipitous data connections.

• Fact check their internal AI solutions that often fall victim to hallucinations.

Challenge: Creating custom representation of disease biology with harmonized public + private data

Building a custom data graph

The foundation of the platform is a data graph, think of this as a custom GPS for navigating disease biology. It enables teams to keep track of important biological relationships such as disease, genes, variants, pathways and their relationships.

Here is where we differentiate from a lot of knowledge graph providers, the foundation of the custom graph curated for each team is quantitive sequencing data such as Single Cell or Whole Genome Sequencing data. In addition to this, we provide a wide variety of curated, versioned and well annotated consortium data such as OpenTargets, Reactome and NIH clinical trials.

Nothing is worse than picking up a new software tool and having to put in countless hours importing data and setting up the new system. Our graph curators worked with the client to make the set up process as simple as possible.

1. Our client described important biological relationships and questions they would like to answer.

2. The client provided us with access to the data they would like to upload to the platform.

3. We took care of the heavy lifting, tailoring their data graph schema and uploading their proprietary data. We ensured the teams were provided with adequate documentation, UI, and API support should they choose to do any data graph management on their own.

Within 48 hours the team had a custom data graph ready for exploration.

Problems solved

• Creating and managing a data graph. We provided the tools and infrastructure to ensure that the graph could be easily updated and maintained.

• Data wrangling and harmonization. We took care of the tedious data wrangling and harmonization to ensure the client’s data was seamlessly integrated.

• Knowledge base versioning and maintenance. We maintain versioned data packages and ontologies, making third party knowledge base integrations effortless.

Challenge: Identifying Serendipitous Data Connections

Traversing the graph

Once the data graph was set up, the team used the graph explorer to traverse their data. They were able to build compelling narrative around targets, diseases, and cell types. Graph explorer sessions were saved and shared with colleagues for real time collaboration.

The team was able to execute graph algorithms like page rank to identify the relative importance of data points and all shortest paths to reveal serendipitous data connections.

Problems solved

• Identifying critical data points that would have been overlooked had the data remained fragmented.

• Building compelling narratives surrounding targets of interest.

Challenge: Building Evidence Ranking and Prioritization Systems

The BioBox platform does not autonomously predict targets or tell teams which indications they should pursue. It enables scientific teams to use their domain expertise to curate custom evidence ranking and prioritization systems by leveraging their fully connected data graph.

The team configured several graph models tailored to specific use cases:

• Variant prioritization. This graph model provided a ranked list of variants identified across public GWAS studies and proprietary WGS studies that would put patients at risk for diseases of interest. Variants were ranked according to the team’s scientific criteria.

• Indication prioritization. This graph model provided a ranked list of indications that would be suitable for an asset that they have in phase 1 clinical trials. Indications were ranked according to the team’s scientific criteria.

• Pathway prioritization. This graph model provided a ranked list of pathways that were perturbed within a disease of interest. Pathways were ranked according to the team’s scientific criteria.

• Target prioritization. This graph model provided a ranked list of genes for diseases of interest according to the scientific criteria that they value.

Problems solved

• Team domain expertise was siloed from quantitative data. Through the curation of a graph model, teams were able to develop a framework of domain expertise for a variety of use cases.

• Prior to the use of BioBox countless hours were spent updating/ rerunning reports and analyses. All graph model reports autoupdate upon the injection of new data.

• Managing data from an active discovery pipeline. All reports are timestamped and versioned. This provided the team with a comprehensive understanding of how data points were changing in priority in as new data was injected.

Challenge: Complex Multi-hop Data Lookups

Intuitive Query Language

Knowledge graphs are notoriously difficult to query. No one wants to write in cypher or SPARQL.

The team used our intuitive query language to traverse their ontology and ask complex multi-hop questions without unnecessary table joins. Multi-hop questions pulled information from multiple data sources within the graph.

Questions such as such as “ Which genes participate in the RNA polymerase II Transcription Pathway and are upregulated in internal tumor samples "? could be answered in seconds.

Important queries were saved so that they could be executed at any time.

Problems solved

• Unnecessary table wrangling and table joins. Countless hours are spent consolidating data before important value generating analyses can begin. Using the query language teams were able to obtain answers within seconds rather than hours.

• Steep coding learning curve typically associated with knowledge graphs. All team members were able to retrieve information from the knowledge graph without the use of code.

Challenge: Fact checking AI solutions

Natural language - Multi-omic GraphRAG

New AI tools are hitting the market each day, many of which fall victim to hallucination. The team was working towards an AI strategy and wanted a ground source of truth derived from data they could trust.

The team leveraged our natural language GraphRAG to converse with their sequencing data.

Problems Solved

• Develop a ground source of truth to fact check their AI. Instead of relying solely on new AI tools that are prone to hallucination, the team was able to use our platform to fact check information against quantitative sequencing data within their graph.

By using the BioBox Platform our client was able to curate a custom data graph without the tedious pain points typically associated with knowledge graph management.

This enabled them to

• Build a proprietary evidence ranking and prioritization system that led to the identification of 10 novel targets.

• Prioritize 2 new indications for existing assets.

• Curate a custom representation of disease biology.

• Leverage machine learning algorithms to identify serendipitous data connections.

• Fact check their internal AI solutions that often fall victim to hallucinations.

The BioBox data graph has given us a 360 view of our data. Data points that could have previously slipped through the cracks are now front and center. This has resulted in us spending significantly less time wondering if we made the right decisions in our discovery pipeline because we are leaving no stones unturned.

Director of Translational Biology

Interested in fact checking your AI solutions and conversing with your data? Send us an email at sales@biobox.io, we enjoy complex data challenges.

As new AI solutions hit the market every day, data scientists are forced to discern the truth from hallucinations.

Knowledge Graphs

It is imperative to have a central source of truth to fact check the information provided by LLMs. One way in which data teams are tackling this challenging is through the use of knowledge graph. In short, a knowledge graph is a semantic representation of data points and their relationships. When it comes to drug discovery, think of it as a GPS for understanding disease biology.

At BioBox, we provide therapeutic teams with the tools to manage and curate a custom knowledge graph composed of private multi-omic sequencing data and public knowledge bases.

For a deep dive on how we leverage them to understand complex biological relationships check out this article below.

Fact Checking LLMs with GraphRAG

It’s no secret that knowledge graphs are typically hard to traverse - no one likes an ontology hairball. Not everyone wants to write in Cypher or memorize their entire graph schema. At BioBox, we have built a suite of tools to make extracting valuable information from the graph as easy as possible, one of which is GraphRAG.

RAG (Retrieval-Agumented Generation) is a framework used to improve the accuracy of information provided by LLMs. It improves LLM accuracy by retrieving information from external sources. This provides the LLM with up to date information grounded on the basis of relevant data rather than solely relying on the data the LLM was trained with. GraphRAG take this a step further, and retrieves information from a knowledge graph. GraphRAG enables teams to use natural language and obtain information directly from their knowledge graph.

Last week, data scientists at Microsoft published GraphRAG to github, enabling users to leverage an LLM to extract knowledge from a collection of proprietary text documents.

At BioBox we are excited to announce GraphRAG support for our multi-omic sequencing based graphs. Instead of spending hours wrangling data and fighting table.joins teams can use natural language and obtain information directly from their graph within seconds.

Check out the demo below.

Our bread and butter is heterogeneous sequencing based data, however we support a wide variety of knowledge bases including NIH clinical trials, Reactome, Alliance Genome, OpenTargets and more.

This allows for questions such as

• What are the genes targeted by drugs used in clinical trials that study the disease Renal cell carcinoma and have an overall status of terminated?

• What are the biological processes that genes that drugs act on that have a clinical precedence for the disease Renal cell carcinoma are involved in? 

• Which variants put you at risk for Renal cell carcinoma?

We have been developing a closed beta with several therapeutic teams. Over the next few months we will be continuing to improve our GraphRAG to support more complex multi-hop queries and will be releasing it to all users.

Interested in fact checking your AI solutions and conversing with your data? Send us an email at support@biobox.io, we enjoy complex data challenges.

Platform therapeutic companies can be divided into two primary categories; those who actively pursue novel targets and those who prioritize licensing their platform technology often opting for well characterized targets.

According to the FDA, platform companies utilize a well understood and reproducible technology that includes a molecular structure, mechanism of action, delivery method, vector, nucleic acid sequence or any combination of the aforementioned.

The technology must:

• Be essential to the structure or function of a drug.

• Can be adapted for, or used by more than one drug sharing similar structural elements.

• Facilitate the development of more than one drug through a standardized process.

A significant value proposition for platform companies lies not only the success of a single program but how many drug programs and indications it can facilitate, thus the desire to be a “designated platform technology”.

To be considered a “designated platform technology” the platform must:

• Be used by a drug approved under section 505 of the FD&C Act or section 351 of the PHS Act.

• Have preliminary evidence that demonstrates the platform technology can be used by more than one drug without adverse effects on quality, manufacturing or safety.

• Data indicates that the technology has a reasonable likelihood to bring significant efficiencies to the drug development, manufacturing or review processes.

Historically, the process for reusing platform technologies in new applications has been long and inefficient. However, new guidelines aim to operationalize this process.

Guidance for Industry : Platform Technology Designation Program for Drug Development

What has changed?

If a platform has an approved ANDA, NDA or BLA companies can request designation of that platform technology to enable leveraging of the technology in new or future applications.

What are the benefits?

Shorter and more efficient timelines

• Engaging early with the FDA enables team to receive timely advice.

FDA prioritization

• If there is significant public health benefit the FDA may prioritize additional engagements leveraging the designated platform technology.

Have an arsenal of data for subsequent programs

• No need to repeat work that was already done, evidence that was used to support the efficacy of a platform technology for a previous submission can be used for subsequent applications. This includes batch, stability and non-clinical safety data.

What is required?

Submissions require the following

• Description of the platform technology and how it meets FDA standards.

• The approved ANDA, NDA or BLA for the technology.

• Identification of shared structural elements between drug products and how the element facilitates the use of the platform technology.

• Scientific support for the use of the platform technology across multiple drugs and how this would not affect safety, quality or manufacturing.

• Risk assessment to evaluate differences between previous and proposed drug product.

• Information to justify why the platform technology would bring significant efficiencies to the drug development , manufacturing or review process.

Read the full document here.

The guidelines are currently in review and the FDA is accepting comments until July 29th.

The road to platform designation became a little easier, now what?

Companies that were traditionally focused solely on licensing their technology might start looking towards developing their own target pipelines. Companies that already have a target pipeline may expand their approach to additional indications and targets.

These processes involve complex decision making and requires a deep understanding of disease biology, patient population, and market needs.

Target selection and indication prioritization challenges

Target selection and indication prioritization are inherently difficult. These challenges become heightened for platform companies who have dedicated the majority of their resources to developing their technology.

Some of these challenges include

1. Complexity of Disease Biology

   • Understanding the intricate mechanisms of diseases and identifying relevant molecular targets is a significant challenge. Diseases often involve multiple pathways and interactions, making it difficult to pinpoint the most effective targets for therapeutic intervention.

2. Multi-Modal Data Integration

   • Integrating multi-omic public and private data to identify actionable insights requires countless hours of data wrangling and database maintenance.

3. Cross functional Collaboration

   • Computational and Translational teams need a medium to share hypothesis, data and collaborate in real time.

4. Designing the right experiments to evaluate a target

   • Identifying the in vivo and in vitro experiments that will best evaluate a target is notoriously complex.

5. Having enough evidence to support a target or indication

   • Combing through thousands of data points to find enough high conviction evidence to support a target often seems like a never ending task.

The BioBox Data Intelligence Platform: A Solution

The BioBox data intelligence platform can help platform therapeutic companies overcome these challenges. Here's how:

Build a GPS to understand the complexity of disease biology: The backbone of target selection is understanding the biology of the disease. Through a fully customizable knowledge graph teams can map out the important biological concepts and relationships that the care about in the context of a disease. Enabling them to identify serendipitous relationships and novel connections. We break down why this is important here.

Seamless Data Integration: With our API and SDK we make multi-omic data integration as easy as possible so you can leave the data wrangling and Table.Joins behind.

Curate custom target and indication prioritization reports: Teams can leverage the knowledge graph to curate custom target and indication prioritization reports. Prioritize data points on the basis of scientific criteria that matters to the team. Check out our case study on how we enable a research team to develop a target prioritization system for renal cell carcinoma here

Searchable and accessible data in one place. Keep track of all of the data supporting a program or a target in a single unified resource that is accessible by all team members. Search for specific relationships and data points across the graph.

Develop an arsenal of targets. It’s no secret that not all targets make it through the validation phase. After substantial wet lab interrogation many targets get cut from the pipeline. Instead of having to start from ground zero, stockpile prospective targets and have backups for your backups.

Scaleable Resource. The data intelligence platform is designed to grow and scale with your team and data. As new data is streamed into the platform reports and analyses automatically update.

If you would like to learn how we enable therapeutic research teams to save time and resources while creating a collaborative ecosystem for target prioritization book a demo with us here.

About BioBox

BioBox is a data intelligence platform for drug discovery. Rapidly build a custom data graph and deploy graph ML models for target prioritization, indication selection, and MOA analysis. Teams using BioBox make faster and better data driven decisions to de-risk drug programs and get assets into clinical studies quicker.

Before utilizing the BioBox platform, the company addressed the following key questions to establish a clear direction:

1. What is our therapeutic approach?

2. What is our biological hypothesis?

3. What internal data will we work with?

4. What external data sources are relevant?

Answering these questions provided a foundation for the structure of their knowledge graph.

Step 2: Configuring a Custom Knowledge Graph

The core of the BioBox platform is a fully customizable knowledge graph. Think of the knowledge graph as a GPS for disease biology where the roads are relationships between destinations such as genes, pathways, variants etc. Explicit directions (data sources) tell you how you can get from point A to B.

Creating and maintaining a knowledge graph and the associated ontologies, is often a complex and time-consuming task. BioBox simplified this process, allowing therapeutic teams to focus on research rather than data wrangling.

Translational and computational teams mapped out the important biological relationships on the BioBox platform. The team started by loading preconfigured data packs from BioBox, which provided a robust initial framework. These data packs contained versioned ontologies and observations from consortiums such as OpenTargets, Alliance Genome, Clinical Trials and more.

They then uploaded their internal data, integrating all biological metadata with predefined concepts and relationships within the knowledge graph.

Wrangling ontologies, consolidating public knowledge bases and integrating internal data are processes that often take weeks to accomplish. Through the use of the intuitive UI and API we were able to accomplish this in a few hours.

- Director of computational biology

For more information on knowledge graphs and their role in drug discovery, refer to our detailed explanation here.

Step 3: Curating a Graph Model

A graph model maps important relationships between two concepts within the knowledge graph. For target prioritization, the concepts used were Genes and Diseases.

Each relationship, or "line of evidence," received a weighted score based on its importance. The team defined scientific criteria linking genes to diseases and categorized lines of evidence. Here are a few examples:

• Gene Expression:

  • Upregulated genes from internal bulk RNAseq and scRNAseq data

  • Upregulated genes from TCGA bulk RNAseq data

• Genomic Observations:

  • Variant frequencies in diseased populations (Internal data)

  • Variants that contribute to increase gene expression (OpenTargets)

  • Variants that are at risk/ protective for a disease (OpenTargets)

• Epigenetic Information

  • Promoter regions that are hypomethylated in disease populations (methylation array (Illumina 850k)

  • Promoter regions there are acetylated in disease populations (H3K27ac ChIPseq)

• Pathways and Biological Function:

  • Genes associated with specific pathways (Internal data + Reactome)

  • Genes that are markers for disease (Internal data + Alliance of Genome Resources)

• Safety and Efficacy:

  • Clinical trial information, including trials with drug failures (ClinicalTrials)

  • Safety liabilities associated with genes (OpenTargets)

  • Drugs approved to target gene of interest (Chembl)

Negative scores were assigned to lines of evidence indicating poor target suitability, such as genes targeted by drugs that failed clinical trials. These scores were used to compute an overall weighted score for each gene, ranking them according to the team’s scientific criteria.

Step 4: Reporting & Ranking Targets

Once the graph model was curated, the team generated reports to evaluate targets. They selected "Renal Cell Carcinoma EFO:0000681" as the disease of interest and left Genes unbiased.

Within seconds, a report of genes ranked according to the scientific criteria that they curated was generated.

By selecting a target and a line of evidence scientists could see the supporting data identified within their knowledge graph.

The initial report highlighted well-characterized targets like BRIC5, AKT, and HIFa.

Using similarity and community detection algorithms, natively enabled in the graph model, the team was able to explore novel targets. Running a Jaccard similarity algorithm yielded targets that scored similar to well-characterized targets along the same lines of evidence. In other words, these targets might behave very closely to well-studied targets like AKT etc. This represented just one example of how the BioBox platform enabled principled and data-driven hypothesis generation through the exploitation of the graph data.

Based on the evaluation, the team adjusted weights and added additional lines of evidence to refine their prioritization system.

As new data was integrated into the knowledge graph, reports and target scores automatically updated, saving the computational team valuable time.

The impact of proprietary data on target prioritization was evident as the team compared reports from graph models with and without their internal data.

To further characterize a target the team explored all of its relationships within the graph explorer. A full break down on our graph explorer can be found here.

Conclusion

By leveraging the BioBox platform, the company efficiently built a custom target prioritization system for renal cell carcinoma. The comprehensive integration of internal and external data sources, coupled with BioBox's customizable knowledge graph and robust analytical tools, provided several key benefits:

• Time Savings: The streamlined ontology management and automated data integration allowed the team to quickly get started and maintain momentum, reducing the time typically spent on data wrangling.

• Cost Efficiency: By efficiently organizing and analyzing data, the company was able to focus resources on the most promising targets, reducing unnecessary expenditures on less viable candidates.

• Enhanced Confidence: The weighted scoring system and robust evidence framework gave the team greater confidence in their target selection, supported by clear, quantifiable data.

• Comprehensive Evidence: The inclusion of diverse lines of evidence, such as gene expression, genomic observations, and clinical trial data, provided a thorough understanding of each target's potential, ensuring well-informed decision-making.

• Cross-Functional Collaboration: The platform’s collaborative features enabled seamless communication and data sharing among team members, fostering a cohesive and efficient working environment.

Note: Proprietary data has been redacted for privacy. Sequencing-based observations in images above have been derived from TCGA data.

About BioBox

BioBox is a data intelligence platform for drug discovery. Rapidly build a custom data graph and deploy graph ML models for target prioritization, indication selection, and MOA analysis. Teams using BioBox make faster and better data driven decisions to de-risk drug programs and get assets into clinical studies quicker.

To learn more about how BioBox can supercharge your discovery pipeline, book a demo with one of our team members!

Biomedical data is highly fragmented. Just for describing genes alone there are at least 3 major systems for identifiers (e.g. Ensembl, Refseq, Entrez). There are countless concurrent efforts in ontology development, dataset harmonization, and public knowledge base creation. Together, they provide the context that scientists use to interpret data. However, in practice, whether you start with literature, or go right to the source data, at some point in your journey you’re going to end up with 100 browser tabs open trying to connect the dots together.

It’s not you, its the data.

To solve this, BioBox developed the Graph Explorer. An interactive application to traverse your custom knowledge graph, just like a map, and develop a GPS for understanding disease biology.

BioBox Graph Explorer

The Graph Explorer is an ontology-aware application that allows scientists to:

1. Discover Hidden Connections: Dive deep into the interrelationships between real-world data objects. Whether you're linking genes to diseases or phenotypes to pathways, the Graph Explorer reveals patterns that are crucial for breakthrough discoveries.

2. Enhanced Data Mining Capabilities: Leverage advanced algorithms such as A*, Dijkstra’s, PageRank, and k-means to navigate and dissect complex data structures. Whether you're finding the shortest paths between two indications of interest or identifying significant gene regulators within your interaction network, our tool simplifies these tasks, making them more accessible and actionable.

3. Collaborative Exploration: Share your findings effortlessly with colleagues. The Graph Explorer allows you to save and share exploration sessions, ensuring that your team builds on collective knowledge and biological insights, fostering a collaborative research environment.

Navigate Through Information Quickly

Hours of valuable research time are wasted trying to figure out how data is connected and sifting through endless webpages and databases. The graph explorer centralizes all these disparate data sources into a singular view.

Load in an object from your data graph and instantly see all the connections associated with it. Importantly, relationships have semantic value and convey a meaningful association between two real-world objects.

For example, in a gene-centric view, starting with KLF4:

You can select one or more nodes by clicking or brush selecting. Active nodes in your selection will display a data card that lists out all data properties annotated on the node. The relationships panel lists the known edges connecting to KLF4 where:

• the direction of the arrow indicating the orientation of the connection

• the badge label representing the Concept associated with the connected object through that relationship, and

• the number in parentheses indicating the quantity of distinct edges of that type.

Exploring Data Connections

Each relationship is backed by either a curated assertion (e.g. from a known ontology) or extracted from a data source. These data sources can be internal proprietary datasets or from literature and public data. In this graph, we’ve loaded in Locus2Gene scores from Open Targets Genetics portal and transformed them into VariantAssociation hyper nodes. By expanding the “has association” edges, we can instantly load in GWAS inferred associations between traits/diseases and KLF4.

Through the directionality and differences in edge types, we can model biological complexity directly in the graph topology. Adding in additional styling helps to visually distinguish these patterns.

Zoom out to see the bigger picture

As we start to walk through the data graph, we can ramp up the data exploration space by expanding all relationships at a Concept level. In our KLF4 example, now that we’ve found some Disease connections through GWAS linkages, we can load in all known disease markers into the canvas.

Next, we can connect the Gene nodes in the canvas through the “activates” edge to reveal a sub-network of gene regulatory interactions.

Run graph algorithms

At this point, the graph is dense, and what we’ve systematically done is the following logic:

• Starting with KLF4

• Find some GWAS associations between KLF4 and diseases

• For those diseases, find all the gene markers

• For all gene markers, connect their activation regulatory patterns

While it is difficult for us to reason over the graph visually, they are the perfect substrate for graph algorithms that can yield interesting data-driven insights. For example, now that we’ve loaded in a gene subnetwork from a given context, we can run community detection algorithms like PageRank to score and detect highly influential genes, with the click of a button!

Sharing a biological narrative

You can take a snapshot of your current session that captures the data, styling, and results of your analysis so that you can continue working on it later, or perhaps, share it with a colleague. There are no limits to how many sessions can be saved, and that means you have an infinite canvas and the tools to paint your biological narrative and share it with your team.

About BioBox

BioBox is a data intelligence platform for drug discovery. Rapidly build a custom data graph and deploy graph ML models for target prioritization, indication selection, and MOA analysis. Teams using BioBox make faster and better data driven decisions to de-risk drug programs and get assets into clinical studies quicker.

To learn more about how BioBox can supercharge your discovery pipeline, book a demo with one of our team members!

If you want to have a successful drug program, target selection is the most important thing you need to get right. The reality is that if you have a bad target, nothing else really matters, it is never going to work. This is why scientists spend a lot of time and resources meticulously analyzing as much data as possible to develop a compelling scientific narrative. We need to better understand the biology of the disease. The more evidence we can find, the more confident we can be in our decision-making. In this post we present our framework for accelerating scientific reasoning in drug discovery and introduce BioBox: The BioData Intelligence Platform.

BioData Intelligence Platform

We envision a platform that acts as the central repository of collective knowledge that integrates multiple data sources to weave together a cohesive data graph. Biologists can interact with and explore this graph in a simple and intuitive interface. Complex scientific logic can be modeled and tuned to assist scientists in making strong, well-reasoned, data-driven decisions.

The platform is divided into 3 layers:

1. Ontology - Defining what things mean in your organization

2. Model - Build custom graph ML models that scores things your way

3. Application - Put models to use in purpose built apps designed for biology

L1-Ontology: Define your data language

Speed and decisiveness is determined by the clarity of communication between different stakeholders when making a target selection decision. In practice, this boils down into getting 3 things right:

1. Getting everyone on your team to agree on what things mean in your organization

2. Improve the readability and interpretability of critical business objects

3. Capture how data enters and gets transformed in your research practices

Your ontology is where you create a single source of truth for your team. Every concept is unambiguously defined and their relatedness explicitly declared. As you begin building and populating your data graph, it automatically keeps track of its history to ensure proper data lineage, governance, and provenance.

L2-Model: Declare your scientific logic

Part of good scientific communication is being able to justify your rationale with data. Certain observations support our hypothesis and some may give us reason to question our assumptions. The strength of the influence depends on the type of evidence we are given. For example, you’d probably give more relevance to sequencing analyses of patient-derived tumor cells than to assays done on immortalized cells. In other words, evidence are valued with different weights. We use these considerations in our mental models that, at its fundamental level, is trying to explain the relatedness between 2 concepts.

In BioBox, we capture this logic into discrete graph models that are expertly tuned by your scientists to automatically compute a score based on all data in your graph. As new information is added to your graph, these graph models recompute to ensure your scores are always up to date. What this means for research teams is that you have a tractable and explainable way to justify scientific reasoning, that is unique to your science.

Because these graph models can be built against any two concepts in your graph, you could use them to score Gene → Phenotype, Drug → Disease, Gene → Biological Process, etc.

Graph models allows your scientists to transparently and explicitly describe data interpretation strategies and facilitates strong data-driven decision making.

L3-Application: Putting the models and graph to work

The platform makes building the data graph and the models easy, but to get real business value, computational and translational scientists must be able to interact with them. The application layer of the BioBox platform refers to the collection of ontology and graph-aware software tools that enable your scientists to interactively mine the graph for novel insights, generate data-driven reports powered by your graph models, and provides API solutions to integrate your data graph into your internal ML models.

Automatic Report Generation From Graph Models

Using your graph models, scientists can rapidly generate scored target prioritization reports in seconds. These reports allows you to:

• generate a scaled score of associations

• get a full trace of every data point used to generate scores

• one-click data snapshot and share findings with cross-functional teams

Interactive Data Graph Exploration

Put the power of your data moat in the hands of your scientists by allowing them to explore it interactively. The BioBox Graph Explorer is a ontology-aware way to navigate the rich connectivity and context of your data graph. Scientists can run graph algorithms on the fly, including community detection and path finding, to discover new insights embedded inside your proprietary and/or publicly available data. Sessions can be saved and shared with cross-functional teams to collaboratively build a scientific narrative.

Who is this for?

This platform caters to teams who

• Are looking to obtain an in-depth understanding of the biology of a disease.

• Need answers to complex biological questions.

• Have multi-modal data (NGS, low throughput, clinical metadata, etc).

• Need a flexible and customizable system that will scale with additional data and team members.

• Have an internal AI strategy.

Who is this not for?

This platform is not for

• Teams looking to strictly mine literature.

• Teams without multi-modal data.

• Teams who do not have specific biological questions they are trying to answer.

• Teams looking for one size fits all knowledge graphs.

Bottom Line

The BioBox Platform is a fully integrated system for research teams to answer complex biological questions using multi-modal data. It helps you solve three things:

1. Clarify Communication: By establishing a shared vocabulary and a clear understanding of data relationships through the ontology layer, BioBox ensures that all team members are aligned, enhancing collaboration and efficiency in decision-making processes.

2. Enhance Decision-Making: Through the model layer, BioBox empowers scientists to quantify their hypotheses and test their theories with precision. This structured approach to scientific reasoning enables you to make data-driven decisions with confidence, backed by a dynamic scoring system that adapts to new data.

3. Drive Innovation: The application layer transforms complex data interactions into actionable insights, enabling scientists to discover novel therapeutic targets and strategies. This not only accelerates the research and development process but also opens up new avenues for innovation within drug discovery.

Today, the challenge often lies not in the acquisition of data, but in its interpretation and utilization. Simply having data is no longer enough. To fully exploit the volume and variety of data available, we need systems to capture the relationships and patterns from observations to form a more holistic view. This is where Knowledge Graphs (KG) truly shine, and when properly used, will completely transform that way data is used in drug discovery. In this article, we describe at a high level, what a Knowledge Graph is and how to get started crafting one.

What is a Knowledge Graph?

At its core, a Knowledge Graph (KG) is a formal structure designed to represent information as a set of entities and the intricate relationships between them. In simple terms, we can think of individuals like BRCA1 or Breast Cancer as nodes, and relationships like mutated_in as descriptive edges between nodes.

When developed properly, it can enable automated reasoning and inference to unearth hidden implicit connections that often belie breakthroughs and discoveries. Many leading biotech and pharmaceutical companies have been ingesting data and constructing internal knowledge graphs to use in a variety of applications from drug repurposing to target identification.

DeepLink, Janssen Pharmaceuticals: Using DeepLink, an internal knowledge graph platform, researchers successfully identified two hallmark targets (one of them is now in their portfolio) for pulmonary hypertension.

ARCH, AbbVie: ARCH is the name of AbbVie’s internal knowledge graph. In a case study, AbbVie scientists used the embedded logic to discover a putative therapeutic in their portfolio that could be used to treat Carney Complex, a rare and deadly disease with no approved treatments.

The essential and most important component of a KG is the underlying ontology. Ontologies are semantic data models which determine the types of concepts that exist in the KG. Importantly, they are distinct from individual (data points) in the KG because they represent entire categories of concepts, not a specific named concept. For example, instead of describing the gene SOX2 and specific properties about SOX2, the ontology focuses on defining the concept of Gene and capturing the characteristics that a Gene should have. Some of these characteristics can be relationships to other concepts in the ontology. For example, we can describe the relationship between a Gene and Pathway with participates_in.

Using this ontology, data can be structured in an interpretable way, and forms the KG.

Tangible Benefits: Drug Repurposing

For biotechnology or pharmaceutical companies that have existing assets in their portfolio with a defined target and mechanism, we can expand the market value of the asset if we can find alternate use-cases for it.

Suppose your asset BX123 selectively inhibits the G1 gene and is an approved treatment for Disease Y by rescuing an overactive pathway P1. When enough data is modeled into a knowledge graph, we can start to make some inferences and uncover new insights. For example, it is also known that P1 activates transcription of G2, which in turn, activates P2 that is implicated in the etiology of Disease X. Through walking the KG, it can be inferred that the asset BX123 could be a viable treatment for Disease X, expanding the use-cases for the asset and ultimately the market value.

Tangible Benefits: Target Identification

KGs can also help researchers make informed decisions when selecting a target to drug. Using sequencing datasets, we can load in observations such as gene expression, differential expression, mutations etc. to enrich the knowledge graph. For example suppose we have normal tissue expression data and disease-specific patient tumor sequencing data.

At first glance, G1 and G2 appear to be viable targets as they both participate in an overactive disease-causing pathway, but after considering information loaded from sequencing data, G2 seems to be most viable. Here’s the reasoning:

• G2 is selectively upregulated in disease

  • G2 is poorly expressed in normal tissue

  • G2 is significantly upregulated in Disease X compared to healthy normal controls

• G1 is not contributing to overactive pathway

  • G1 is abundantly expressed in normal tissues and not found to be differentially expressed in Disease X. It stands to reason that if G1 is implicated in Disease X, all brains would have Disease X.

In this hypothetical scenario, it illustrates how the incorporation of observations from sequencing datasets into the knowledge graph enables researchers to conduct evidence-based reasoning quickly and effectively.

Challenges and Considerations

Despite the tangible benefits that KGs can bring into a biotech research team to accelerate drug discovery, there are significant challenges to making the KG effective. Here are some considerations to make if you are thinking of starting a KG initiative.

Build a shared dictionary

Don’t be an ontology goblin.

The goal is to have shared knowledge. The ontology should be built with all stakeholders, from computational biologists to business teams, in mind. Successful companies implementing KGs have a common understanding of how things are defined using a shared vocabulary in their organization. This massively reduces communication errors and makes data ingestion, harmonization, and reporting a breeze. It matters not where this is done, but that it is visible and available to everyone.

Connect data streams

Before committing to building a KG, be sure to plan out where the data comes from and if it can be reliably and efficiently connected to the KG. Building the perfect ontology and infrastructure is meaningless if you cannot consistently add data to the graph. While they can vary by use-case, the most common ones would be LIMS metadata, bioinformatic workflows and outputs, ELNs etc.

Bottom Line

Knowledge Graphs are tools that can drive massive acceleration in the drug discovery process. There is growing adoption and use of knowledge graphs in big pharma and biotech. When harnessed effectively, it enables research teams to reason over the data at scale and unearth hidden patterns.

However, effective KGs require significant time and capital resources, which may limit adoption in early-stage startups and biotech companies, despite the immense value KGs hold. This is where purpose-built platforms like BioBox can help. We empower smaller research teams with dedicated tools to:

• collaboratively build ontologies, with access controls and versioning

• integrate with data streams to power your knowledge graphs

• explore and mine knowledge graphs for insights

Ready to get started? Book a Demo