Key Points:
- Polygenic scores (PGS) give insights into a patient’s inherited risk for coronary artery disease (CAD) but it is unclear how this information should inform clinical care.
- Participants in this study had their lifetime risk of major coronary events (MCS) estimated by a PGS and were then divided into deciles based on that risk. A causal AI algorithm that integrated data from Mendelian randomization studies as well as randomized trials was used to estimate the reduction in LDL or systolic blood pressure (SBP) needed in each decile to reduce their risk to that of the participants with the average PGS. These estimated were then validated against the observed events in each group.
- This study found that most patients could overcome their polygenic risk for CAD with small lifetime reductions in LDL and systolic blood pressure (SBP), however the later in life that therapy is initiated, the more that lipids and blood pressure needed to be lowered to achieve the same effect.
Polygenic scores (PGS) combine multiple genetic variants from multiple pathways that contribute to varying degrees and mechanisms to a given disease state. While a PGS can give an assessment of risk, in isolation they do not provide information on the specific etiology of that risk or what specific actions could be taken to lower that risk. Therefore, their clinical utility in guiding individual treatment decisions is not well defined.
Artificial intelligence (AI) systems designed to elucidate causality, called Causal AI, seek to provide information on causal relationship that purely predictive AI models trained on historical data might lack. The algorithms created by Causal AI are designed to encode biological mechanisms of cause and effect in order to be able to predict outcomes as well as prescribe the specific actions need to change these outcomes. These models can then be tested against observed or empirical evidence to determine validity.
On March 5, 2023, Dr. Ference of the University of Cambridge presented the results of Translating Polygenic Risk for Coronary Artery Disease into Clinically Actionable
Information Using Causal A.I. at the Late Breaking Clinical Trial section of ACC.23/WCC.
This study sought to bridge the gap between PGS and its implications for individual clinical decision making by using Causal AI to estimate how much each person much lower their LDL, SBP, or both in order to overcome their inherited polygenic risk of a major coronary event (MCE).
First, a Causal AI ensemble was trained on 1.8 million participants of either Mendelian randomization studies or randomized trials of LDL or BP lowering treatments. A PGS for coronary artery disease (CAD) was created from over 4 million variants from genome wide association studies. Next 445,765 participants from the UK biobank had their PGS calculated, and Causal AI was used to estimated how much their LDL or SBP needed to be lowered to get to the average polygenic risk of CAD. Then these estimated were validated by comparing the observed event curves of patients randomized by nature (using genetic instrumental variables) into those who lowered their LDL, SBP, or both by the amount estimated by Causal AI to overcome their polygenic risk versus those in the same percentile of polygenic risk but with mean population LDL and SBP levels. The primary outcome was time to first major coronary event (myocardial infarction or coronary revascularization).
They found that for most of the population, Causal AI predicted that polygenic CAD risk can be overcome with relatively modest lowering of LDL, SBP, or both (e.g. the 99th percentile of risk would require a a 20.7 mg/dL reduction LDL C and a 6.0 mmHg reduction of SBP to reach the equivalent of the 50th percentile of risk). The Causal AI output was validated by the fact that patients with the PGS in the 80th percentile that were randomly allocated by nature to the magnitude of LDL or SBP lowering prescribed by Causal AI had the same risk of MCE as those participants randomized by nature to average polygenic risk, LDL, and SBP levels. They also found that the later in life therapy was started, the greater the reductions in LDL and SBP were needed to overcome the same polygenic risk. A family history of CAD provided independent and additive information to the polygenic risk score; one was not a replacement for the other, and the overall inherited CAD risk depends on both. The Causal AI was also accurate in the amount of LDL and SBP lowering needed to overcome a patient’s overall inherited CAD risk based on both PGS and family history.
This trial is unique in that it demonstrated how causal AI can be used to take a PGS and turn it into clinically actionable information, in this case by quantifying the amount of LDL or SBP lowering needed to overcome excess polygenic risk for CAD. According to this study, polygenic risk can be overcome by modest changes in LDL and SBP, but the later in life these risk factors are addressed, the more they need to be lowered in order to achieve the same amount of risk reduction. It also showed that for most people, compared to a family history of CAD, “polygenic risk is a relatively weak contributor to their overall risk,” reported Dr. Ference. “Causal AI can be used to estimate how much each person must lower their LDL and SBP to overcome their overall inherited risk based on both their family history and polygenic predisposition.” This novel application of using Causal AI to convert polygenic risk into an actionable clinical plan can help providers fine-tune the intensity of preventative care provided to the individual patient across the spectrum of age and cardiovascular risk. It also has the potential to change the way we approach primordial prevention in the future.
