Drug Discovery

Viva-Thera, XtalPi Nominate HIV Drug Candidate in Seven Months

Boston – September 22, 2026 -- Viva-Thera has nominated a preclinical candidate targeting the membrane-proximal external region (MPER) of HIV-1 gp41, completing discovery and optimization in seven months through a partnership with AI drug-design firm XtalPi (HKEX: 2228). The biotech, incubated and backed by XtalPi as a strategic shareholder, made its public debut disclosing progress across four HIV modalities: vaccines, long-acting small molecules, gene therapy and cell therapy.

XtalPi's AI screened 10 million compounds to isolate five validated hits

Using generative AI, XtalPi navigated a chemical space of approximately 10 million compounds to identify five hits with strong antiviral activity and favorable drug properties. Continuous optimization of these leads drove the program to preclinical candidate nomination within seven months, according to the companies.

Dualitas Signs Up to $1B Roche Deal for Bispecific Antibody Discovery

South San Francisco, Calif. – September 19, 2026 -- Dualitas Therapeutics has secured a research collaboration and license agreement with Roche worth up to $1 billion, including a $36.5 million upfront payment, to develop bispecific antibody therapeutics for immunology and inflammation diseases.

Dualitas will screen over 300,000 bispecific antibody combinations for Roche

Under the agreement, Dualitas will use its proprietary DualScreen Bispecific Discovery Engine to functionally screen more than 300,000 novel bispecific combinations, in what the company describes as one of the largest-scale bispecific discovery efforts undertaken. Roche will take over all subsequent preclinical development, regulatory, manufacturing and commercial activities for candidates that emerge from the collaboration.

Nona Biosciences Debuts HCAbLM, First AI Model for Human HCAbs

Cambridge – September 18, 2026 -- Nona Biosciences has developed HCAbLM, the first language model trained specifically on fully human heavy-chain-only antibodies (HCAbs), built from a dataset of 31.8 million HCAb sequences drawn from 73 independently immunized transgenic mice. The findings were detailed in a study titled "A foundation model learns the sequence and functional grammar of fully human heavy-chain-only antibodies," published on bioRxiv.

Nona targets the developability bottleneck in AI antibody design

Current AI antibody discovery is limited less by target-binding design than by manufacturability at scale, aggregation risk and stability -- collectively known as developability. HCAbLM was built to learn the underlying sequence grammar that determines whether an HCAb can fold correctly and progress into a viable drug candidate. General-purpose protein models lack this capability because their training data contain almost no HCAb sequences.