由维亚生物创新中心投资孵化，致力于将AI技术商业化用于药物研发的生物医药初创公司——Phenomic AI于近日接受采访。Phenomic AI创始人Sam Cooper与Lu Rahman共同讨论了AI在药物发现中的作用。以下为采访全文：LR: Where are we with deep learning and drug discovery and how effective is it for researchers? How well is it received and are barriers to cost an issue?SC: Some promise deep learning will revolutionise drug-discovery overnight; others are over-critical of a powerful new tool in our data analysis arsenal. The reality lies in the middle and is problem-specific.Deep-learning offers a powerful approach for handling complex raw inputs, where abstraction is key to analysis, for example, recognising objects in images, or determining a person’s mood from a speech input. Here, we’ve seen deep-learning achieve human level performance in previously intractable problems. The challenge in drug discovery is identifying key use cases where deep-learning can solve the problem, and the problem is a valuable discovery problem.If you’ve identified one of these key use cases, and demonstrated you’ve a product that solves it – converting the value is not difficult. Similarly, cost is not a barrier – for start-ups, cloud providers have programs letting you get set-up for free; for big companies, compute costs should be a small fraction of the cost of wet-lab experiments and payroll.LR: How can it change the way researchers approach drug discovery in cancer – what benefits and savings does it create for them?SC: A great use case in cancer is in scoring levels of tumour infiltrating lymphocytes (TILs) in a tumour, a key marker of anti-cancer immune activity and thus both an in vivoendpoint, and patient stratification measure. Manual TIL scoring involves a trained histologist analysing a huge number of slides and is expensive and time-consuming. Deep-learning models exist that can accurately quantify TILs across any number of slides for a fraction of the cost of manual analysis.Other use cases in cancer drug-discovery include, more effective patient stratification based on analysis of tumour scans by deep-learning, chemical property prediction to improve molecule design, and analysis of imaging and genetic data from complex in vitroscreens, eg. organoids, which we perform at Phenomic.LR: What are the challenges in discovering and developing new drugs for cancer and how will AI help overcome them?SC: At a high-level, data-analysis in cancer drug discovery is becoming more complex. We’re moving from monotherapies with opportunistic combination trials, to combination ‘native’ pipelines. Targets are shifting from those that are exclusive to cancer cells and their growth, to those that contribute to the state of the wider tumour microenvironment, eg inflamed vs. excluded, invasive vs. non-invasive. Prior failure rates in trials are forcing us to selectively identify specific patient subsets that result in us gaining maximal information on our treatment in the clinic at minimal cost. This results in an environment where we’re unable to process the raw data required to make discovery decisions. Analysts versed in biology and the suite of data-analysis tools at their disposal for processing these data-inputs and providing outputs that can be acted upon will thus become integral to the majority of pipelines.LR: Where do you see the research opportunities in this market?SC: There is opportunity to apply the same strategies that have become the norm in cancer, (eg. patient stratification, combination treatment, population screening) to other major disease areas. This, in part, needs to be driven by concerted efforts from academia to build resources that industry can use, and that are simply too difficult and costly for industry to build. An example of this is The Cancer Genome Atlas (TCGA), which is now an incredible resource employed by companies globally across their discovery pipelines from discovery through to stratification for clinical trials.LR: What long term effects do you see AI having on drug discovery?SC: AI/ML will become ingrained in discovery pipelines in a way that we’ll often forget it’s even there. At Phenomic, we’ve now become so versed in using the probability of a cell being in a specific state (eg. fibroblast vs. myofibroblast probability), we often forget this is not the industry norm. This level of integration, plus continued advances in therapeutic modalities, will be necessary for us to continue to generate new medicines at a productivity level that is acceptable to investors and society as a whole.