Once in a rare while, unexpected events leave a major impact on society such as the current COVID-19 pandemic, necessitating additional urgency to create, innovate and be more efficient. That is true for established enterprises and emerging ones alike.
ReSurfX is an ‘outcomes intelligence’ company focused on improving innovation, outcomes, and ROI of enterprises from data-intensive activities, with primary focus on healthcare (Pharma, Biotech and patient care). ReSurfX provides value through significant improvements in accuracy, robustness, novel insights, and advance prediction of outcome directions by leveraging a data-source agnostic novel machine learning (ML) approach we invented – Adaptive Hypersurface Technology (AHT). ReSurfX::vysen is an enterprise software product that delivers these functionalities.
Here we demonstrate the predictive power of AHT based functionalities through an extremely powerful SyRTOP configuration of ReSurfX::vysen product and outline several applications. The rare form of validations provided here are in the form of FDA actions based on Real World Evidence and Post Market Surveillance that are difficult to achieve even with expensive targeted studies.
Nuanced and extensive evaluations of AHT based solutions have demonstrated markedly higher predictive capacity and other advantages in extracting utility from data than available alternatives. The advantages of ReSurfX::vysen include: (i) robustness – often it is known that even the best alternatives do not perform well even across datasets of the same category, (ii) differentiation – in many applications, the data scale in which AHT based solutions operate to deliver significant value, other ML do not even converge, hence providing unique advantages, (iii) novel insights that are currently missed - including due to technological differences that are causal in the previous two points, and (iv) predicting direction of outcomes in advance - resulting from improvement of both sensitivity and accuracy.
In sectors for which ReSurfX currently provide solutions, some problems are so deeply ingrained that even experienced experts and leaders charged with driving innovative and effective operations often think there are no solutions for those problems. Hence improving or solving these types of problems are not even part of their mission and priorities. In addition to validation of the powerful predictor we discuss relevance to applications to such complex problems in these sectors.
Artificial Intelligence (AI), ReSurfX and Differentiation
Modern Artificial Intelligence (AI) solutions, the term often used loosely, increase the ability to gain information from data beyond what was previously possible. AI solutions are built with ML and statistical approaches at the core. Although these ML and statistical methods of analysis have been in use for a long time, they have not been uniformly successful even across datasets similar to ones they were developed (or customized) for, hence leading to many known as well as other shortcomings affecting conclusions and ROI. The improvements in most modern AI solutions are achieved from new combinations of analytical approaches of the past and creative addition of algorithmic layers to ML (e.g., back propagation) made possible by immense advancements in computational technologies. In contrast to other AI solutions, the ReSurfX::vysen product leverages the novel data-source agnostic ML approach AHT in turnkey functionalities already incorporated into the product as well as those being built. ReSurfX::vysen offers markedly improved advantages over alternatives in terms of accurate predictions and novel insights.
System Response based Triggers and Outcomes Predictor (SyRTOP) of ReSurfX::vysen
Recently ReSurfX released SyRTOP, a new powerful predictive use configuration of functionalities incorporated into ReSurfX::vysen. A pictorial outline of the ReSurfX::vysen SaaS product in a powerful System Response based Triggers and Outcomes Predictor (SyRTOP) configuration is shown below (Fig. 1). The blocks in green are currently incorporated turnkey solutions in the ReSurfX::vysen product. Below, we outline some of the modules that facilitate the SyRTOP application and then demonstrate the unprecedented predictive power of SyRTOP as well as the ability of individual component functions to improve existing workflows by integrating with current solutions in use in an enterprise. Each of the FOUR major functionalities of ReSurfX::vysen outlined here are best-in-class based on our comparisons to widely used public-domain and commercial solutions. Importantly, AHT and the ReSurfX::vysen functional modules are data-source agnostic and have wide applications. Below we highlight the value of SyRTOP through its use in predicting drug classes from response profiles in a rigorous format. Even the specific use case used here has direct utility across the product life cycle (PLC) of the healthcare sector (i.e., Pharma/Biotech and patient care).
The four major components of that lead to accurate predictions by SyRTOP
- vysen-Comp: AHT based solution for accurate and robust detection of differences between states of systems in large datasets.
- vysen-Pred1: One of several predictors from the ReSurfX suite that takes total response of a system state based on one or more reporter variables (e.g., data reporting state of the system of a whole organism, tissue, cells etc. in response to diseases, treatments etc.). In the example used below, system state is defined by total changes of a single reporter variable – i.e., gene expression. The vysen-Pred1 ‘a predictor’ component is designed as a plugin to accommodate other solutions that will aid use of solutions built for serving special business goals. We initially built in this optionality to help the ReSurfX team and early customers compare different solutions with SyRTOP to evaluate value.
- vysen::KR – An accurate Knowledge [Representation] Repository with accurate representation of system responses from data used as prior knowledge to enable predictions.
The ability of the best expertise (human and machines) to consistently make accurate decisions of high value as in expensive and critical patient care or where value and errors from each stage are cumulative (such as a long PLC or complex workflows) depends on the quality of knowledge repositories of prior knowledge. ReSurfX has demonstrated that current backend data repositories used (as ground truth) have over 30% errors. AHT is designed to overcome many of the inherent problems in ML and statistical approaches that current AI solutions suffer from. In this way, SyRTOP helps enterprises maximize the value of their data and expert manpower.
- vysen-KR-Gen: A “Big Data solutions” based module that assembles extremely accurate KRs from enterprise-specific [often data unique to enterprises are key differentiators from their competition] or other large public domain data assets.
vysen-KR-Gen and vysen-KR serve many purposes that continuously improve the ability to maximize the value of enterprise specific assets, data, knowledge, and infrastructure. This happens in many forms, including continuous addition of information that enterprises generate, inherent as well as incorporated improvements of ReSurfX::vysen functionality modules, and through evolving data processing capabilities. The ReSurfX::vysen product will also be expanding built-in vysen::KR minimally as part of product and functionality tests at-scale. Currently, the product incorporates carefully sourced drug-response KR from 7000 responses to drugs at different doses and time of treatment.
Nuanced demonstration of value in SyRTOP predictions
To demonstrate the predictive power and accuracy of SyRTOP and its components, we provide an example on drug class prediction given a particular system response. In this case, the system response is total gene expression changes following treatment with commercial drugs as well as chemicals acting through the same mechanisms of action (MOA). We specifically demonstrate marked improvement of SyRTOP to predict correct and/or relevant drug class(es) for “query responses” in comparison to a widely used commercial solution from NextBio® (now part of Illumina®) as well as a workflow based on the best public domain solutions. The drug class predictions were performed for an extensively studied dataset generated by the US FDA. We combined data on system-wide gene expression changes generated by sequencing (RNA_Seq) and microarray platforms upon treatment with many drugs (at different doses and time-points) from the US FDA study as well as from the US National Toxicology Program (NTP). This “knowledge repository” that was used as learning dataset with metadata, had highly related (and often comorbid) diseases, drugs for unrelated disease indications, and drugs with subtle variations in chemical structures with different disease use indications. This dataset thus allowed in-depth evaluation of performance characteristics expected of an “exceptional predictor”.
Specifically, we focused on input queries representing two drug class combinations with the following properties:
- SIX ‘fibrate’ and TWO ‘statin’ drugs for the comorbid diseases hyper-cholesteraemia (often also associated with diabetes and obesity) and heart diseases, respectively. This combination is one of the most prevalent pre-existing conditions in the developed world.
- FIVE ‘azole’ chemical based drugs used in fungal treatments and TWO ‘estrogen class’ drugs (called estrogen class for simplicity) approved for birth control and for treatment of menstrual problems in women (that are estrogen dependent). The latter are chemical variants of a specific class of azoles (triazole).
The reference KR used for learning and prediction included 675 response states after 196, 51, 111 and 66 treatments with fibrates, statins, fungal azoles, and estrogen class of drugs, respectively. In addition, other drugs and chemical treatments in the KR used in this demonstration included known toxins, suspected toxins and carcinogens, NSAIDs, chemotherapeutics, and neurological agents. In all responses were from 163 independent drugs and compounds used for treatment at various doses and/or for different time periods.
*We simplified the KR metadata for the four drug classes as follows: we labelled three related compounds with same biological effect (on diabetes, cholesterol and acting primarily through the same biological pathway) as ‘fibrates’.
In Figure 1, the alternative approaches compared with ReSurfX::vysen functionalities are represented in yellow and pale orange colors. ReSurfX can help customers incorporate their solutions similarly.
The pictorial in Fig. 2 summarizes the number of each of the four input query drug classes (i.e., fibrates, statins, azoles, estrogens) in the top 25 predictions for each of the 15 query drug responses - using:
- SyRTOP – that used vysen-Comp and vysen-Pred1 top panel,
- input query and KR were processed using a workflow using some of the best public domain solutions further improved by us (we refer to as SWF1) - bottom left panel, and
- input query and KR processed using a widely used commercial approach of NextBio® (now part of Illumina®) (we refer to as SWF3) - bottom right panel
The RFA approach from Illumina® was used for drug class prediction with SWF1 or SWF3.
In Fig. 2, each of the four query drug classes is indicated by a different color. For example, fibrates are indicated by green, and in the case of the first query drug in the top panel, which is a fibrate, 25 of the 25 predictions of SyRTOP in the KR are also fibrates. The striking improvement in prediction accuracy and quality from SyRTOP is evident in Fig. 2 - despite the complexity in the form of many closely related and confounding data in the KR.
The drug class prediction result shown in Fig. 2 is not of breakthrough importance per se as the query set is an extensively studied dataset. However, this use case highlights how the AHT based SyRTOP application captures important previously validated and sophisticated insights in a highly nuanced data design and evaluation of results.
* A note on the number of predictors used in this example: though top 25 used here is a reasonable choice for simplicity of the purpose, the ReSurfX::vysen product offers multiple options for the users either to choose by top N predictions or an empirically calculated threshold for the metric of trust (in the range 0 – 100) provided with each prediction.
[Edit 02/14/2021: the result shown in this version is further improved than indicated in the introduction of SyRTOP ReSurfX::vysen Yields Remarkably Accurate and Actionable Insights Using System Response to Triggers – A Drug Response Study dated January 21, 2020 that already had marked improvement over other alternatives.]
Several cases of complex and significant value uncovered in the above results are outlined below. These results necessitate this section to dive a bit deeper into the science, as the use case is biologically and medically oriented. The key conclusions and evidences are in brown font.
- The first and most obvious value is the remarkable cleanliness of prediction of drug classes by SyRTOP – as evident from a visual observation of the number of predicted drug classes that are the same class as that of the query (each represented by a different color in Fig. 2). Whereas, the bottom two panels (RESULTS OBTAINED WITHOUT USING SyRTOP) are often riddled with WRONG predictions in all cases. This result does not need further elaboration – except to know if there are known drug interactions that are missed or if any drug class prediction that is different from the query input is valid [see next set of observations].
- The following nuanced validations were uncovered from predictions of SyRTOP through close evaluation of the results and deep sector knowledge. The insights, as it turned out, happened to be proven by Real World Evidence (RWE) and Post Market Surveillance (PMS). Such results are difficult to achieve through quick bench or animal experiments, hence are of higher value in the larger context of a Pharma enterprise or patient care needs.
- Fibrates and statins are used for the comorbid conditions hyper-cholesteraemia and heart disease. Fibrates reduce triglycerides (bad fat) and to some extent raise HDL (good) fat. Both effects of fibrates reduce the risk of heart disease and stroke. Although there were SIX fibrates among the query drugs, interestingly only in the case of GEMFIBROZIL (GEM) (drug #3 of x-axis in Fig. 2 and Fig. 3) dis SyRTOP predict one of the four drug classes different than of the query. In the case of GEM, SyRTOP also predicted the input query as statins multiple times in the top 25 predictions used here. Importantly, it turns out that RWE has shown that GEM over compensates triglycerides. As a consequence the US FDA and the American Heart Association (AHA) released safety guidance in 2012 not to prescribe GEMFIBROZIL with statins.*
*RWE is one class of data from large number of patients typically collected from medical claims, complaints etc. after the drug is approved for a medical condition by FDA (or other medical regulatory authority in other parts of the world) and released to the market. This class of RWE and their use is called Post Market Surveillance (PMS), to know about previously undetected effects from clinical trials (especially negative effects, contraindications) that become evident when used in large number of patients. Often based on such data, regulatory agencies mandate drug manufacturers to add extra precautionary or restrictive labels OR in some cases withdraw a drug from the market.
- The other important observation comes from drug class prediction on the azole type of drugs that are approved and used as anti-fungal creams. Of the FIVE azoles among the query drugs, SyRTOP predicted FLUCONAZOLE (drug #14 of x-axis in Fig.2 and Fig. 3) as belonging to the estrogen-class of drugs more frequently than with other azoles. FLUCONAZOLE is the only triazole, the other four were imidazoles.
This is another valuable observation because:
The estrogen-class of drugs (norethindrone and ethinylestradiol) approved for use as birth control and menstrual problems etc. are derived through chemical modification of triazole.
SyRTOP’s classification of fluconazole as having estrogen-like properties is also consistent with the US FDA safety announcement in 2011 that high-dose FLUCONAZOLE (the triazole mentioned above) administered to pregnant women can cause birth defects.
SyRTOP predictions are (i) very clean and (ii) when it did predict a query drug as belonging to a different drug class, it captured validated insights. This provides strong support for customers to use ReSurfX products and functionalities and rely on the new insights predicted by SyRTOP. Such validations of value of SyRTOP predictions demonstrated above are difficult to achieve or expensive to prove even with targeted experiments. The goal of ReSurfX is to improve such predictions, reduce failures, prevent unwanted resource spend, and improve decisions at each step (in drug development, patient care and more). The evidence of the sort summarized here provides strong incentive to our customers for adoption of ReSurfX products in the form of strong validations of value and ample prior results, thus lesser effort would be required for adoption.
SyRTOP predictions also have other important ramifications on many potential applications (besides those highlighted above):
- The SyRTOP predictions uncovered valuable and complex insights that are usually difficult to validate as in (i) the fibrate (GEMFIBROZIL) and statins prediction, and (ii) the structural similarity of a chemical structure likely showing up in the total system response - both cases also had known contra-indication that lead to FDA guidance based on RWE and PMS.
- The earlier such negative effects of a drug in development are identified, effort can be made to modify a chemical, biologic etc. and reduce undesired properties OR kill a drug development program with clear rationale before spending lot of time and resources.
- A worse case is where such negative effects are identified after release to market (as in PMS) with enormous cost to the society and Pharma. One well-known example is the pain killer drug Vioxx (from Merck) that was found to cause a high number of heart attack and stroke – resulting in withdrawal from the market and expensive lawsuits. Research later published in the medical journal Lancet estimates that 88,000 Americans had heart attacks from taking Vioxx, and 38,000 of them died [Source: NPR, 2007].
- The results and applications of SyRTOP can be further extrapolated from the recent use of dexamethasone in a rapid prospective clinical trial on COVID patients during this ongoing pandemic and recognition that it reduced death in COVID patients with respiratory distress by up to 50%. Prior RWE based support for dexamethasone was supported by limited but reasonable effectiveness when used to treat patients with ARDS (Acute Respiratory Distress Syndrome). The rapid use was possible because the toxicological and other several relevant properties were known as dexamethasone was approved for other disease treatments over 50 years ago, thus excluding a protracted but very important step in early stages of a typical clinical trial. Further the pandemic made this near real-time patient reassignment trial reassignment of patients between without and with dexamethasone treatment possible. The above example highlights many more uses for SyRTOP often not prioritized today, as solutions that offer the right combination of insight and robustness do not exist (hence lack reliability and pose expensive risks). The results from SyRTOP also provide strong basis for its applications in DRUG REPURPOSING – i.e., identification of new uses for preapproved drugs (so reducing time and cost).
- Lastly, the interactions identified could also be previously unknown positive effects, a major goal of combination drugs strategies being attempted for a while now. These efforts are expected to make a strong impact on the market in the near future. It is realistic and reasonable to expect that the predictive insights from SyRTOP on drug classes having similar system behavior should be valuable with such efforts as well.
Though not shown here, the data used in the results used in this example alone provided an opportunity to confirm extensibility of such performance with 500,000 comparisons.
While the above examples relate to drug development, the same can be equally explained in the context of patient care.
THESE VALUE OF SyRTOP HIGHLIGHTED CAN BE CRISPLY ENCAPSULATED AS:Robustness, reliable and novel insights with proof =Trust in ReSurfX::vysen outputs, predictions =Significantly better innovation and ROI.
Component functionalities used in SyRTOP improves existing workflows with other solutions – adding extra value:
We used two major components vysen-Comp and vysen-Pred1 based on AHT for SyRTOP. Remarkably, we find that both: (i) vysen-Pred1 [component 2] alone applied to poor input data that is the current standard (as mentioned above, contains over 30% errors), as well as (ii) vysen-Comp [components 1, and central to 3 and 4 providing input system response profile and knowledge-base to other ML and statistics solutions used (as AI or not)], significantly improve the quality of results of drug class prediction. The above is evident by comparing Fig. 3 that uses either vysen-Comp or vysen-Pred1 with bottom two panels of Fig.2 where no ReSurfX functionality leveraging AHT was used.
Integration, interoperability and reuse of components
The ReSurfX::vysen product is architected to be integrative with the workflow of customers and to facilitate this advantage in multiple ways, including through:
- APIs with well-defined end-points and SDKs often bundling many APIs into functional modules to enable automation in data-intensive efforts,
- user friendly UI that enable subject matter experts to easily test, confirm and guide informatics teams on large scale applications,
- having properties necessary to integrate with IT infrastructure and compliance needs, and finally
- features that cater to manpower continuity and regulatory compliance.
In addition, results shown in Fig. 3 demonstrate that component modules of SyRTOP in ReSurfX::vysen improve outcomes with existing solutions used in an enterprise. Specifically:
- query system state generated by vysen-Comp and vysen-KR generated using vysen-Comp improved prediction from other ML, AI or statistical learning approaches, as well as
- vysen-Pred1 when used with existing lower quality of data (compared to those processed by vysen-Comp) improves predictive ability.
With this knowledge, we designed the generation of high accuracy KR based on vysen-Comp as a module and vysen-Pred1 as a plugin to facilitate use of other modules by customers. This modularity is in line with integrative product design facilitating interoperability and reusability to improve existing workflows and for uses in complementary or other special business needs.
Breadth of applications and value
It is essential that a novel technology based product of this nature (with new solutions continually developed and functionalities incorporated into the product) be highly integrative with alternatives that have been developed and customized over decades for specific applications. We highlighted a lot of uses for SyRTOP (one use configuration). Importantly, solutions of ReSurfX have a lot of other important applications because: (i) AHT is a data-source agnostic solution, (ii) using more than one reporter variable of a system will only improve the predictive power and robustness, and (iii) AHT is applicable to generation of data, knowledge extraction from each data source and for combining different data sources for effective downstream operations (outcomes). Some other salient advantages of ReSurfX offerings leveraging AHT not elaborated here also include the ability to deliver better outcomes through inherent personalization in outcomes as well as the ability to extract valuable insights from lesser data even in cases where ML and most statistical approaches fail which opens up the many new valuable opportunities. Those advantages are outlined in ReSurfX website in examples in previous posts at ‘vysdom’ section (including ‘Overcoming the Curse of Dimensionality with Combinatorics’).
In summary, ReSurfX provides differentiated value compared to other available alternatives and ReSurfX offerings improve innovation and outcomes, complementing in-house expertise, thereby opening new commercial options. One key differentiator is the use of a different core approach by ‘Adaptive Hypersurface Technology’ (AHT) that has inherent advantages, including increased accuracy with higher sensitivity with robustness, and the ability to predict outcome directions in advance. ReSurfX provides targeted applications of high-value where few alternatives work well to maximize the innate value of enterprise resources and expertise for healthcare and for drug discovery and development enterprises. With these advantages, the applications of data-source agnostic AHT based solutions span efficient and innovative early discovery, clinical trials, care delivery to patients (including decisions on treatment choices), knowledge of efficacy all of which have direct consequences on business models and practices that aid the outcomes-based healthcare economy. ReSurfX value offerings for downstream applications (clinical trial and after) in drug development enterprises overlap with those for patient care in healthcare enterprises. Besides obvious improvements in innovation and efficiency in data-intensive activities, some other areas of high-value applications of ReSurfX across business operation include:
- Drug Repositioning and Combination Drugs.
- Choice of drug/treatment for a given patient/time.
- Development of capabilities for ‘Advance Prediction of Outcome Directions’ that will help:
- Monitoring clinical trials.
- Evidence based market approach for high cost drugs.
- Patient treatment choices and monitoring.
- Quantitative support for Evidence Based Medicine.
We continue to build some of the most accurate products and solutions for ‘outcomes intelligence’ leveraging the power of AHT beyond SyRTOP. Let 2021 be the year where one of your major enterprise decisions is to partner with ReSurfX to maximize the value of your assets as well as expand your business goals. Initiate a conversation with our team by clicking HERE that links to contact page in ReSurfX website.
We appreciate the contribution of all members associated with ReSurfX in various capacities. Special thanks to: Suresh Gopalan, Fred Ausubel, David Kuttler, John Janakiraman, QC-General (pseudonym), Joseph Sabatini, Jeff Killian, Sanjay Jain, Saranya Balaji, Sandhya Kshirsagar, Brendan Koenigar, Sandeep Kochhar, Shahin Gharakhanian, many talented software developers and other contract personnel. We also thank our investment partners who wish to remain anonymous at this time.
**June 17, 2021: added the word ‘algorithmic’ near the first use of ‘back propagation’ to explicitly differentiate from neural network layers.