Artificial intelligence (AI) has become an increasingly popular tool for drug companies discovering and designing new therapies. According to analysis by , the AI market in drug discovery is expected to grow from $159.8m in 2018 to $2.9bn by 2025.
Of the almost 180 start-ups involved in AI-assisted drug discovery in 2019, 40% were working on repurposing existing drugs or generating novel drug candidates using AI, machine learning, and automation.
AI-enabled drug design company Valence Discovery, formerly InVivo AI, was founded in 2018. Since its rebrand last month, the company has announced a series of impressive drug discovery and design partnerships, with the aim of making advanced technology accessible to R&D organisations of all sizes.
鈥淭he overarching mission of Valence is really to empower drug discovery scientists with the latest advances in AI-enabled design,鈥� says CEO Daniel Cohen. 鈥淎nd that’s not just faster, cheaper drug discovery, but it’s also about unlocking novel therapeutics space so that we can now address what were previously intractable problems using these AI methods.鈥�
Valence鈥檚 academic origins
Valence has its origins at Canadian AI research institute Mila, where the company鈥檚 founding team focused on developing deep learning tools for drug discovery and design during their PhD studies.
鈥淲hat we’re trying to accomplish is the very rapid and cost-effective design of high-quality drug candidates that are optimised for a broad range of potency, selectivity, safety, DMPK [drug metabolism and pharmacokinetics] parameters that are relevant to whatever particular drug discovery programme we’re working on,鈥� Cohen explains.
US Tariffs are shifting - will you react or anticipate?
Don鈥檛 let policy changes catch you off guard. Stay proactive with real-time data and expert analysis.
By GlobalDataPharma and biotech companies have been using AI tools to make sense of big data for some time, but what makes Valence鈥檚 technology unique is that it鈥檚 centred around 鈥榝ew-shot learning鈥� 鈥� that is, developing a learning model using very little training data 鈥� and finding value in small, noisy datasets.
For targets and indications that have already been extensively researched, there will be large amounts of pre-existing data to use. Companies looking to develop and design novel therapeutic approaches, however, will be working with very limited information. While small data holds the potential for novel therapies in areas of high unmet need, drug discovery teams must be able to effectively work with sparse datasets 鈥� and this is precisely what Valence鈥檚 technology aims to achieve.
鈥淚f we want to move into novel target spaces, or novel chemical areas where we have inherently little pre-existing data, we need entirely new classes of deep learning methods built specifically for low-data environments,鈥� Cohen says. 鈥淎nd that’s what few-shot learning allows us to do for our partners.鈥�
Addressing the limitations of AI-driven drug design
As well as small data鈥檚 inherent trickiness, synthetic accessibility 鈥� how easily chemical compounds can be synthesised 鈥� is another challenge involved in incorporating AI into novel drug discovery and design. Many of today鈥檚 AI systems generally yield low-quality molecules; they are highly reactive, unstable, synthetically infeasible, and therefore difficult to translate into effective treatments.
What sets Valence鈥檚 platform apart, Cohen says, is that it addresses the core limitation of existing AI technologies around synthetic accessibility. 鈥淭he limiting step in these AI-oriented design, make, test cycles is always the make and the test,鈥� he adds. 鈥淚f you can鈥檛 readily synthesise these AI-generated molecules, the value-added AI in a typical discovery programme is going to be limited.鈥�
To circumvent this obstacle, Valence has developed new classes of design technologies that Cohen says enables teams to enforce a high degree of synthetic accessibility and medicinal chemistry quality into AI-generated molecules.
Despite most major biopharma companies now employing AI-driven solutions for drug discovery, effectively integrating this technology into the process remains another major challenge.
鈥淲hen you look at biopharma today, only a tiny fraction of the space is AI-enabled,鈥� Cohen says. 鈥淏uilding high-quality AI capabilities internally is just not a core competency for a lot of discovery-oriented organisations 鈥� the space is evolving really quickly; it’s very challenging to stay on top of the latest methods.
鈥淭he field really needs to move to a point where you have plug-and-play infrastructure that’s been built specifically for drug design, that makes these tools more accessible to drug discovery scientists and to R&D organisations of all sizes, not just the largest pharmas.
鈥淩eally what we’re trying to do at Valence is democratise access to deep learning and drug design.鈥�
Valence launches with a raft of new partners
As Cohen highlights, one strategy for overcoming the biopharma-AI integration struggle is collaboration with AI tech start-ups. In the weeks immediately following Valence Discovery鈥檚 unveiling, the company announced several partnerships with major pharmaceutical companies and research institutes. Despite being made so early in Valence鈥檚 journey, the collaborations were exciting, rather than daunting, for the company.
鈥淲e had many years of peer-reviewed science demonstrating the value of these technologies, we’re headquartered at the largest deep learning research institute in the world, we have some of the world’s leading deep learning scientists, like Professor Yoshua Bengio as close scientific advisors,鈥� Cohen explains. 鈥淎nd we built up this really interdisciplinary team that’s bilingual in computation and also in the life sciences.鈥�
Cohen emphasises that all of Valence鈥檚 deals are structured around the needs of the partner, and that the company is an active collaborator that seeks to 鈥渟hare in the successes of any AI-derived molecules鈥�.
The company鈥檚 first announced collaboration, with pan-Canadian drug discovery and research commercialisation centre IRICoR, Universit茅 de Montr茅al, and the Institute for Research in Immunology and Cancer of the Universit茅 de Montr茅al, seeks to discover novel drug candidates for the treatment of levodopa-induced dyskinesia in Parkinson鈥檚 disease.
The target in Valence鈥檚 collaboration with Repare Therapeutics is similarly specific: precision oncology medicines. Cohen says AI is a natural partner for companies looking to optimise personalised treatments of this kind, allowing them to move through the discovery process as quickly and cost-effectively as possible, and explore chemical spaces they ordinarily wouldn鈥檛 have access to.
鈥淚n Repare鈥檚 case, it’s a really, really nice collaboration because we’re combining the best of both worlds,鈥� he says. 鈥淭hey have this really powerful platform on the biology side for target identification, and we’re combining that with our platform for generative chemistry, really allowing their team to focus on what they do best, which is innovating on the biology.鈥�
Valence鈥檚 most recent partnership, a drug discovery deal with French pharma giant Servier, is far broader. Under the agreement, Servier will leverage Valence鈥檚 technological expertise to generate novel drug candidates for multiple targets, while Valence is set to receive an upfront payment and success-based milestones on any drugs derived from the partnership. While Cohen can鈥檛 go into the specifics of the Servier deal, he says the collaboration involves moving into new chemical spaces to unlock difficult-to-treat targets.
At this relatively early stage of AI鈥檚 development as a drug discovery and design tool, technologies like Valence鈥檚, while immensely promising, are circling around the margins of mainstream drug development 鈥� as Cohen acknowledges, AI currently supports only a small proportion of the pharma sector鈥檚 clinical programmes. But the potential for machine learning to find clinically-relevant links that human minds have missed is clear, and Valence is betting that these technologies will drive a major sea-change in drug development over the next decade.
鈥淲e believe quite strongly that by 2030, the majority of drug candidates entering the clinic will have been designed with meaningful input from AI systems and advancements,鈥� Cohen says. 鈥淲e’re very excited to be playing a role in empowering the shift towards AI-enabled drug design across the entire industry.鈥�
