Jim Abraham, PhD
Chief Data Officer
SVP, Cognitive Computing
Cancer research has traditionally involved the use of a handful of markers to understand the origin of tumors and to predict therapeutic response. Increasingly, multi-marker patterns and more complicated rules have been developed that enable clinicians to customize therapeutic strategies for each patient. However, cancer discovery has arguably been limited to what a human can comprehend. Applications of artificial intelligence to large cancer datasets have been hampered by the varying degrees of assay quality in public datasets. The need for large amounts of high-quality data to independently validate the algorithms provides an additional roadblock to large scale deployment of findings as these data sets are fixed.
To address these needs, Caris Life Sciences® built one of the largest clinico-genomic database in the world, now called CODEai™. This massive database is a substantial resource for discovery as it contains more than 215,000 molecular profiles of tumors with consistently gathered clinical grade molecular data along with immense amounts of clinical outcomes data. CODEai also excels in that it is ever-growing, which provides researchers with completely independent validation sets from the exact same assay for their algorithms. Our team has leveraged CODEai to generate multiple artificial intelligence algorithms aiding in diagnosis and therapy selection.
Carcinoma of Unknown Primary (CUP) represents a clinically challenging heterogeneous group of metastatic malignancies in which a primary tumor remains elusive despite extensive clinical and pathologic evaluation. Furthermore, CUP is associated with poor outcomes, which might be explained by use of suboptimal therapeutic interventions since there is a general agreement that CUP tumors retain the biologic properties of the putative primary malignancy. MI GPSai™ was trained on genomic data from 34,352 cases and genomic and transcriptomic data from 23,137 cases, and was validated on 19,555 cases in order to aid in diagnosis. MI GPSai can predict the correct tumor type out of 21 possibilities on 93% of cases with 94% accuracy. When considering the top two predictions for a case the accuracy increases to 97%.
For therapeutic decision support, first-line treatment for metastatic colorectal cancer was investigated for predictive signatures. The vast majority of mCRC patients receive FOLFOX-based first-line treatment even though neuropathy frequently limits its use beyond four months. MI FOLFOXai™ is a 67 gene molecular signature predictive of efficacy of oxaliplatin-based chemotherapy in patients with metastatic colorectal cancer. The signature was predictive of survival in an independent real-world evidence (RWE) dataset of 412 patients who had received FOLFOX/BV in 1st line and inversely predictive of survival in RWE data from 55 patients who had received 1st line FOLFIRI. Blinded analysis of TRIBE2 samples confirmed that FOLFOXai was predictive of OS in both oxaliplatin-containing arms (FOLFOX HR=0.629, p=0.04 and FOLFOXIRI HR=0.483, p=0.02). In the FOLFOXIRI arm of TRIBE2, the signature predicted a median overall survival difference of 14.1 months between the groups predicted to have increased benefit versus those predicted to have decreased benefit.
Multiple predictors to address additional clinical challenges are in development. Given the enormity of the possibilities and realizing that we cannot be constrained by individual scientists working in a serial fashion, Caris CODEai, which is now available to POA members, allows members the ability explore these interactions in parallel. The future of cancer research is large scale, high quality data combined with large scale, high quality research using artificial intelligence to find non-linear, multivariate patterns – the future is here.